Fast Company Technology

Artificial intelligence has moved quickly from a novelty on the job to a common workplace tool. Gallup found that 52% of U.S. employees and 68% of professional services workers were using AI by February 2026. Thomson Reuters similarly found in spring 2026 that 74% of professionals use AI several times a week for work.

This widespread use shouldn’t surprise anyone. For accountants, attorneys, consultants, and other professionals, AI can accelerate research, summarize complex information, and assist with writing and other forms of communication.

But its rapid adoption has a big downside—AI can produce information that sounds authoritative and looks polished while being completely wrong. And when AI generates mistakes, responsibility for those errors belongs to the professional or business relying on the AI output—not companies that develop AI systems.

AI errors have led to lawsuits, court fines, financial consequences and public embarrassment from fabricated citations, nonexistent sources, and inaccurate AI-generated information. These cases raise an increasingly important question: How can professionals benefit from AI without outsourcing the judgment their work requires?

Fake citations

We are two certified public accountants who research tax accounting. In the September 2026 issue of Journal of Accountancy, we examined what happened in Clinco v. Commissioner, a case before the U.S. Tax Court where an attorney submitted a brief to the court with allegedly fake citations.

The Tax Court is a specialized federal court that resolves disputes between taxpayers and the IRS. This case involved a restaurant owner accused of underreporting more than US$2 million in business receipts and claiming unsupported depreciation deductions on rental properties. The main issue concerned Clinco’s accounting records used to determine how much of their income to tax.

The taxpayers’ attorney, Abraham Wagner, cited four cases to support his argument that an IRS notice assessing additional taxes was invalid.

The court determined that three of the citations appeared to be AI-generated hallucinations that cited nonexistent Tax Court cases. The court rejected the taxpayer’s argument and warned Wagner that submitting fake citations could lead to sanctions from the court, including financial penalties or disciplinary action against Wagner.

Ultimately, the court did not sanction Wagner in the case, but the decision stands to put attorneys and other professionals in Tax Court on notice that AI hallucinations will not be tolerated.

When AI gets it wrong

AI errors made on behalf of clients are not limited to briefs drafted by attorneys. Nor are these mishaps limited to the United States.

When Deloitte produced a 237-page report in 2025 for the Australian government examining its automated system for penalizing welfare recipients who failed to meet job-seeking requirements, it contained nonexistent academic citations and a fabricated quotation it attributed to a federal court judgment.

The global consulting firm later issued a revised version of the error-ridden report. Deloitte blamed generative AI for the errors in the original and agreed to partially refund its fees.

Like most airlines, Air Canada has a chatbot on its website to respond to its customers’ questions. The chatbot told a customer in November 2022 that he could buy a plane ticket under a bereavement policy and be reimbursed later. This was not company policy, and Air Canada refused to honor what the chatbot had told the customer.

The customer sued the airline and won 812 Canadian dollars in damages. The British Columbia Civil Resolution Tribunal ruled that Air Canada had failed to take reasonable care to ensure that its AI bot provided accurate information and agreed with the customer that the airline had negligently misrepresented the bereavement policy.

AI lawsuits

There have been other lawsuits filed against AI platforms and providers.

In 2025, for example, a Georgia court dismissed a defamation lawsuit against OpenAI after its ChatGPT tool falsely accused Mark Walters, a radio host, of embezzlement. The court ruled in favor of summary judgment for OpenAI and found Walters had not established fault and damages.

There is also an ongoing case, Keene v. Google LLC, that is similar to the Walters case. Keene, a TV producer, sued after Google’s AI summaries allegedly published false information about Keene’s criminal history. In September 2026, the court ruled that some of his defamation claims may proceed, without deciding yet whether Google was liable.

Hallucinations and ‘confabulations’

Unfortunately, for anyone using AI to save time and use less labor, AI errors are often hard to spot. Unlike most obvious mistakes humans make, an AI-generated error may include supporting explanations, citations or references that make a passage appear complete and credible.

The National Institute of Standards and Technology issued a report on AI risk management in July 2024 that calls AI mistakes and hallucinations “confabulations”—the confident presentation of erroneous or false information.

Some researchers draw a more specific distinction between hallucinations, in which AI invents information outright, and confabulations, in which it fills gaps or distorts real information in a plausible way.

Either kind of problem can be hard to detect without independent verification, but confabulations mimic actual research so closely they can trick AI users into believing the AI output is valid.

AI errors in the workplace

A 2024 Stanford University study of specialized AI tools for legal research found erroneous results in 17% to 33% of the responses tested. Yet those findings and other grounds for concern about accuracy have not stopped many law firms from embracing AI tools due to their potential to save time and improve efficiency.

Recent surveys indicate law firms and other employers have not adopted enough workplace safeguards to spot and root out AI-generated mistakes. A 2026 survey by the Nationwide insurance company found that only 36% of surveyed businesses have written AI policies, and only 37% provide any training on responsible AI use.

As AI becomes more and more common in the workplace, we believe it’s clear that telling employees simply to “check the AI” is not enough. Businesses and the people working for them need clear expectations about when it’s appropriate to use AI, which outputs require independent verification, and who is ultimately responsible for the final work.

AI may change how professional work gets done, but it does not change where the buck stops.

Sarah Webber is an associate professor of accounting at the University of Dayton.

Kaitlin Newkirk is an assistant professor of accounting at Xavier University.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Reading is great. Finding something to read is annoying.

We all know the saying “don’t judge a book by its cover,” but what else are you supposed to judge a book by? Whether you’re walking through a bookstore or library or scrolling through a collection on your favorite reading device, the cover is inevitably going to be the first thing you notice.

Maybe the art will stand out, or maybe you’ll see a name you recognize. But it’s hard to know what, exactly, you’re going to like without actually reading some of the words inside.

Today’s tool is a clever way to get around that and find the perfect read for you—without relying on the most superficial initial impression.

This tip originally appeared in the free Cool Tools newsletter from The Intelligence. Get the next issue in your inbox and get ready to discover all sorts of awesome tech treasures!

An absurdly clever discovery engine

The next time you want to find something spectacular to read, remember a site called ​Uncovered.ink​.

Uncovered.ink solves the book-discovering challenge by not showing you a book’s cover at all—instead, you see a preview of the actual opening chapter. It’s a thoughtful way to find new books to read without being influenced by the usual external variables.

➜ Think TikTok or Tinder for books: You get a quick glimpse at what something’s all about and keep reading if you’re interested—or swipe to the next book if not.

⌚ Just open up the site​ to start browsing immediately, in a matter of seconds. There’s no need for an account or any sort of sign-in.

Uncovered.ink works almost instantly—no sign-ins or account creation required.

✅ When you click or tap the button to get going, you’ll immediately see the opening chapter of a book. You can scroll up and down to switch to another book, if you want, or you can tap the book to flip through the pages. The arrow keys also work, if you’re on a desktop computer. 

The site has you blindly read the opening chapter of a book to see if it’s for you.

Once you’ve read the opening, the book’s cover, title, and author will be revealed to you—along with a link to buy the book or find it at your local library, if you’d like a physical copy. There are around 1,500 books in the service, meaning you can scroll for a long time until you find something you like. And you can filter by genre, too, if you want. 

If you like what you read, Uncovered.ink directs you to the title—for buying, downloading, whatever you prefer.

☝️ You can even get fresh new excerpts to read without having to keep going back to the site time and time again. There’s an email newsletter​, which brings a recommendation to your inbox every week, and there’s ​an RSS feed​, which gives you the first few paragraphs of a different book every day.

Whichever path you choose, it’s a smart new way to find something new to read, whenever the need arises—and it absolutely qualifies as a Cool Tool that’ll add an extra dose of delight into your life.

• Uncovered.ink is ​entirely web-based​, so it’ll work in any browser on any device you’re using—no downloads needed.
• The site is completely free to use.
• And you don’t have to sign in or share any personal info (unless you opt to sign up for the weekly emails). A browser cookie does keep track of any books you save, but that’s pretty much it.

Treat yourself to all sorts of brain-boosting goodies like this with the free Cool Tools newsletter—starting with an instant introduction to an incredible audio app that’ll tune up your days in truly delightful ways.

Despite the wealth of critical personal data our iPhones hold, many people don’t back up their devices regularly. Usually, this is because the amount of data they’d need to back up in iCloud would vastly exceed their available iCloud storage space—and they don’t want to pay extra each month for an iCloud Plus subscription. Other times, people assume their iPhone data doesn’t change that much day to day, so they don’t need to back up their iPhone that often. 

But not backing up your iPhone regularly can leave you in a world of hurt (i.e., data you can never get back) if it’s ever damaged, lost, or stolen. That’s why it is best practice to back up your iPhone every day. Here’s how to do that with and without iCloud.

Automatically back up each night with iCloud

Apple recommends backing up your iPhone’s data every day to protect against data loss, but the company also understands that many people don’t have the time to micromanage their iPhones that way. That’s why Apple has built an automatic backup feature into your iPhone. It’s called iCloud backup, and once you enable it, your iPhone will create a copy of all your important data that isn’t already synced to iCloud each night and securely store it in the company’s cloud.

Once your iCloud backup is uploaded, if you ever need to reformat your iPhone or get a new one because your previous one was lost or stolen, all your data can be easily restored to your new device. Your old iPhone may be kaput, but your irreplaceable, personalized data lives on.

If you have an iPhone, there’s no reason you shouldn’t be using automatic backups each night. The backups occur in the early morning hours while you are asleep, with your iPhone plugged in and connected to a Wi-Fi network.

The annoying thing about iCloud backups is that they can be quite large (mine is currently about 7 GB). That size exceeds the free 5 GB of iCloud storage Apple gives iPhone owners. In other words, if you want to use iCloud backups, you almost certainly need to pay for an iCloud Plus subscription, which gives you more online storage.

But the good news? You can subscribe to a 50GB iCloud Plus plan for just 99 cents a month—and you get other kinds of iCloud perks with it besides the tenfold increase in base storage.

To enable automatic iCloud backups, do the following:

1. Open the Settings app.
2. Tap your Apple Account name.
3. Tap iCloud.
4. Tap iCloud Backup.
5. Toggle the Back Up This iPhone switch to ON (green).

Now, provided your iPhone has a Wi-Fi connection and is connected to power, it will back up automatically each night.

Manually back up your iPhone to iCloud

If, for whatever reason, you don’t want your iPhone to back up automatically to iCloud each night, you can back it up manually at any time. As a matter of fact, sometimes it’s good practice not to wait for the automatic nightly backup. For example, if you are about to install a major iOS system update, it’s good to initiate a manual iCloud backup beforehand.

To manually back up your iPhone to iCloud at any time, do the following:

1. Open the Settings app.
2. Tap your Apple Account name.
3. Tap iCloud.
4. Tap iCloud Backup.
5. Tap Back Up Now.

Back up your iPhone to your Mac or PC

If you feel leery about having a backup of your iPhone floating around on Apple’s cloud, you’ll be glad to hear you can also back up your iPhone locally on your Mac or PC. This means a backup file is created on your iPhone and then stored on your personal computer.

Creating a local iPhone backup is arguably more secure, since it isn’t stored online and thus isn’t susceptible to hackers who may target data stored in Apple’s cloud infrastructure. You also don’t need to worry about iCloud subscription fees, because a local iPhone backup is limited only by the available space on your computer’s hard drive. 

However, creating a local iPhone backup on your computer means that if your computer is ever unavailable, you won’t be able to access your iPhone backup.

To back up your iPhone to your Mac:

1. Plug your iPhone into your Mac.
2. Click the Finder icon.
3. Click your iPhone from the Finder’s sidebar.
4. Under General > Backups, select “Back up all of the data on your iPhone to this Mac.”
5. Optionally, you can select the tick box to encrypt the local backup, which is generally a good idea. This requires a password to access the backup. But don’t forget it, because the data in an encrypted backup is unrecoverable without the password.
6. Click the Back Up Now button.

To back up your iPhone to your PC:

1. Plug your iPhone into your PC.
2. Open the iTunes app.
3. Click the phone icon.
4. Select your iPhone.
5. Click Summary.
6. Under Backups, select This Computer.
7. Optionally, you can select the tick box to encrypt the local backup, which is generally a good idea. This requires a password to access the backup. But don’t forget it, because the data in an encrypted backup is unrecoverable without the password.
8. Click the Back Up Now button.

Many experts recommend storing a backup of your iPhone on both iCloud and your computer. That redundancy means you always have a backup of the backup if something happens to one of them.

As someone who’s kept a smart display on the kitchen counter for the last eight years, I should be among the target audience for Apple HomeView.

My Lenovo Smart Display is one of the best gadgets I’ve ever owned, mainly because it’s always cycling through family photos that I’d otherwise surely forget. In addition to being a constant source of fond memories for me and my family, it is also occasionally useful for kitchen timers, music, weather forecasts, and managing the grocery list.

HomeView, which launches next month for $299, brings an Apple-like layer of polish to the smart display concept, with a high-resolution display, aluminum base, and built-in subwoofer. Its facial recognition capabilities also allow it to display each person’s messages and calendar events.

But for all those bells and whistles, HomeView appears to fall short on the one feature I care about most: Its photo features are designed around iCloud Photos, making it a tough sell for me and the other 1.5 billion people who use Google Photos. Without a steady supply of fresh photos, HomeView might wind up feeling a bit empty.

How HomeView works

Those who’ve seen the HomeView up close liken it to a giant Apple Watch, with a 7-inch square-shaped screen that provides information at a glance. The main “Faces” view cycles through things like the time, weather, and photos, and you can swipe up to a home screen for built-in apps and smart home controls.

At least at the outset, the emphasis is clearly on Apple apps and services such as FaceTime, Apple Music, Podcasts, and Safari. The role for third-party apps is more limited. If a music or video app is compatible with CarPlay, Apple’s in-vehicle projection system, Apple may provide similar access on HomeView when a compatible iPhone is nearby. This would allow users to access third-party music like Spotify. HomeView will also support third-party home screen widgets, similar to those on iOS.

As for photos, Apple hasn’t gone into much detail, other than saying it will offer “Portrait-style depth effects” and support for shared albums that “loved ones near and far can help fill.” Smart displays from Google and Amazon are both able to surface photos of specific people, including new photos that users upload to cloud storage over time, so hopefully that will be a possibility with HomeView as well.

Either way, Apple has not announced any provisions for displaying photos from outside its own ecosystem. Its photo frame features will likely rely on users’ iCloud Photos libraries exclusively.

Photos are the killer app

Although some folks have wondered about the HomeView’s target audience, I’ve always felt that photo frames are the ideal use case for a smart display. It’s a frictionless way to revisit all the smartphone photos you’ve taken, and it fits perfectly with the concept of a communal device for everyone in the home. Compared to stand-alone digital photo frames, I also like that these devices can do a little bit more, such as controlling smart home devices, playing music, and setting timers.

Without the photo component, though, a device like HomeView just becomes another screen in a home already filled with them. Using it as a photo frame is what justifies everything else. For me, at least, HomeView won’t be of much use if I can’t bring my Google Photos library with me.

While Google does offer an easy way to transfer your entire Google Photos library to iCloud Photos, I’d rather not switch photo systems just for the sake of a new gadget. Besides, I prefer Google Photos anyway as someone who doesn’t exclusively use Apple products.

To be fair, other smart displays are similarly closed-off to third-party photo sources. Google’s Nest Hub is for Google Photos, and Amazon’s Echo Show devices are for Amazon Photos. It’s not surprising that Apple would go the same route, because supporting other sources would create complications, and the whole point of HomeView is to reinforce Apple’s own ecosystem of services.

Still, my hope is that over time, Apple will expand the HomeView’s “Faces” feature to third-party apps. That way, alternative providers like Google Photos, Facebook, OneDrive, Ente Photos, and even the self-hosted Immich could provide a showcase for users’ photo libraries. Without a platform for external photos, the target audience for HomeView becomes a lot narrower.

At least eight states are suing artificial intelligence giant OpenAI, saying the company’s popular chatbot model violated state consumer protection laws by hurting children and deceiving adults in what prosecutors described as a reckless pursuit of profit.

The company’s chatbot model, ChatGPT, is giving dangerous medical advice, encouraging paranoia, and coaching people to kill themselves or hurt or kill others, the lawsuits said.

At the same time, the lawsuits contend that chatbots are pretending to be agreeable and compassionate while preying on the vulnerable and on teenagers, the first generation to grow up with artificial intelligence.

“OpenAI is hurting kids and they are lying to Iowans,” Iowa Attorney General Brenna Bird said in a statement.

The states also accused OpenAI of giving life to rogue AI models—or “cyberhacking agents” as one lawsuit called them—that hacked into other companies, lied to human users, and wrote themselves new mission instructions.

In a statement, San Francisco-based OpenAI said it has built safety protections into ChatGPT for children, and that it would continue to do so to improve the model. It also said it is reviewing what it calls “misaligned models” and that it had found nothing close to the rogue incident discovered in July when an AI agent hacked into another AI company.

States join together in lawsuits

The wave of lawsuits is the latest filed by states under their consumer protection laws, accusing tech giants that offer AI chatbots and social media platforms of harming children. The states that filed separate lawsuits Wednesday in state courts are Arizona, Arkansas, Idaho, Kansas, Montana, Nebraska, and West Virginia.

Matthew Bergman, the founding attorney of the Social Media Victims Law Center, called the lawsuits “the most significant coordinated state action ever taken against an artificial intelligence company.”

In its lawsuit, Arizona said ChatGPT encourages teens to share personal information with it and, at times, discourages them from sharing that information with parents, authorities, or friends.

The states say that the risks and harms—to teenagers and to the broader public from rogue agents—were foreseeable, but that OpenAI has misled people into thinking that ChatGPT undergoes adequate safety testing while the company undermines safety in the pursuit of profit.

“Those revenues have come at great cost to the public, which has suffered irreparable harm—including lost loved ones, disturbing cyberattacks, and now a plausible threat of civilizational collapse—due to (OpenAI’s) reckless behavior and ChatGPT’s uniquely dangerous features and content,” Idaho wrote in its lawsuit.

ChatGPT is the most popular chatbot in the world, commanding nearly half the American marketplace, the states said in their lawsuits.

The AI landscape is evolving

It’s possible all of the allegations in the states’ lengthy lawsuits date to incidents and findings before August, when OpenAI introduced a version of ChatGPT designed for teenagers, with stronger protections around subjects like suicide and sexual chats.

Still, the states said it is far from clear that the current version is any less dangerous.

A watchdog group, Common Sense Media, said last week that some of OpenAI’s protections in its ChatGPT for Teens product don’t work as intended and it urged OpenAI to restrict the chatbot to adults until it can offer a safe product for teens.

Coordinated, multistate lawsuits against the tech industry have borne fruit.

In August, Meta agreed to add stronger child-safety measures to its Facebook and Instagram platforms in an $18 billion landmark legal settlement that ended a trial over teen social media addiction and settled claims filed by nearly every state.

Although the safety measures that Meta agreed to adopt did not apply industrywide, the company quickly urged its competitors to adopt similar measures.

OpenAI is making a similar appeal. OpenAI spokesperson Drew Pusateri said the company is committed to working with states on “advancing pragmatic AI policies that apply to the entire AI industry—not just one company.”

OpenAI also said it works with external experts who review its policies and ChatGPT’s behavior.

ChatGPT is intended for users 13 and older, and users under 18 need permission from a parent or guardian to use it, OpenAI said. The chatbot model for teens is designed to reduce exposure to self-harm, violence, eating disorders, and explicit sexual or graphic content, it said.

Nearly 9 in 10 teens who use ChatGPT use it for learning, information, skill-building, or productivity, the company said. On average, teens spend under 15 minutes a day on ChatGPT, it said.

But Laura Zimmermann, Common Sense Media’s director of youth AI research, questioned the value of OpenAI’s statistics.

For one thing, ChatGPT can do a student’s homework in seconds, so even 15 minutes can have a big impact on a teenager’s development, Zimmermann said. She also questioned whether OpenAI knows which users are teenagers or children, given that her organization’s testing found that ChatGPT couldn’t or didn’t detect which user is a child.

___

This story has been corrected to show the spelling of the surname of Common Sense Media’s director of youth AI research is Zimmermann, not Zimmerman.

—By Marc Levy, Associated Press

Sudden restrictions on Starshield — SpaceX’s military-grade version of the Starlink satellite internet service and the backbone of Ukraine’s drone operations and communication — alarmed the country’s military and government officials this week, with commanders reporting lost drones and disrupted missions.

The episode raised fresh questions about Ukraine’s reliance on the Starshield program and who actually controls access to it.

By Thursday, some Ukrainian units said their connectivity had returned to normal, though uncertainty remained over what had caused the restrictions.

The episode added to tensions between Kyiv and Washington, a relationship already strained by President Donald Trump’s reported deal with Moscow on Russian diesel imports, which Ukraine viewed as a betrayal.

“The Starshield shutdown was short, lasted only one day, but several drones and other systems were affected. Some missions failed, and we lost drones because communications were cut off,” said Yevhen Karas, the commander of Ukraine’s 413th Separate Unmanned Systems Regiment, known as Raid, which specializes in tactical and deep-strike drone operations.

Here’s what to know about the SpaceX satellite services that have become critical to Ukraine’s war effort.

What is Starshield?

Starshield is SpaceX’s satellite program for governments and the military. Built on Starlink technology, it is designed for national security missions and includes secure communications, Earth observation and satellites equipped with specialized sensors. It also incorporates additional encryption and security measures for sensitive government operations.

The Starlink system uses thousands of satellites in low Earth orbit to provide broadband connectivity to terminals on the ground, including in areas where conventional internet infrastructure is unavailable or damaged.

Unlike Starlink, Starshield is tailored to government requirements and can include dedicated satellites and special equipment.

Starshield “is the version where you have much higher speeds, you have better connectivity and it’s not available for civilians; it’s available only for the military,” said Olena Tregub, founder and secretary-general of Ukraine’s Independent Anti-Corruption Commission.

Tregub said reliable, high-speed connectivity was essential for Ukrainian military operations involving drones and units moving rapidly on the ground and in the air. Operators need clear, real-time video feeds to identify what is happening and make decisions, she said.

Ukraine’s increasingly digital battlefield also relies on systems such as Delta, which helps manage and share battlefield data. Transmitting large volumes of information requires fast and reliable connections, Tregub said.

“It’s not just about being able to connect. We need to make sure that this speed is there, that it’s not reduced,” she said, adding that standard civilian Starlink access alone would not meet the military’s needs.

The U.S. government is a major Starshield customer. SpaceX founder and CEO Elon Musk said this week that the U.S. government owns and controls the system, with decisions about its use resting with the U.S. military and, ultimately, the president. His remarks have fueled questions about who was responsible for the latest restrictions affecting Ukraine.

How are Starlink and Starshield linked?

Starlink and Starshield are different services, but they share much of the same technological foundation, including SpaceX’s satellite engineering and capabilities.

The difference lies in the customer, security needs and end purpose. Starlink provides commercial services, while Starshield is designed for government and national security missions.

The distinction is not always straightforward. Starshield can draw on the wider Starlink satellite network and related infrastructure, while also drawing on separate and dedicated satellites and systems.

Initially developed for commercial use, Starlink became a crucial communications tool after Russia’s full-scale invasion of Ukraine in February 2022. Its terminals have helped Ukrainian authorities, emergency services and military units stay connected despite Russian attacks on communications infrastructure.

On the front line, the service supports communications between military units and enables operations requiring a continuous connection between drones and their operators, including missions in which operators need to receive video while aircraft are in flight.

The Starlink and Starshield services are not interchangeable, even when they rely on overlapping technology.

For Ukraine, that matters because restrictions affecting particular terminals or military services do not necessarily mean the entire Starlink network has been switched off. Neither does the continued operation of civilian terminals guarantee that specialized military connectivity will remain available.

What happened during the restrictions?

Ukrainian military personnel reported that the restrictions affected terminals transmitting data at high speeds, interfering with some drone missions and causing losses when operators could no longer maintain communications with flying drones.

Responsibility for the restrictions remained unclear after Musk’s comments raised questions about the role of Washington. The Pentagon has not answered questions about the slowdown from The Associated Press.

By Thursday, Karas said his unit’s connectivity had returned to the required speed and its missions were continuing.

“We are very happy,” he said, adding that he did not know whether the episode had been caused by a temporary technical malfunction or something else.

The incident nevertheless reinforced the need for backup communications, he said. Ukraine’s army was preparing alternative solutions, although they would be more expensive and less effective, he added.

—Samya Kullab, Associated Press

Associated Press reporters Hanna Arhirova and Efrem Lukatsky contributed.

Meta released a new version of its Muse Spark LLM on Thursday. As its name indicates, the big news about Muse Spark Open is that it’s an open-weight model. Rather than being offered solely as a cloud service, it’s publicly available for anyone to run on their own hardware. With 816 billion total parameters and a 1-million-token context window, it’s the U.S.’s most capable open-weight model, according to Meta.

As I wrote in a recent feature, Meta’s AI story dates all the way back to the founding of its original AI lab in 2013. One of this long journey’s high points happened in 2023-2024, when its open-weight Llama model was embraced by companies that wanted to control their own AI rather than rent it from a company such as OpenAI or Anthropic. But that momentum evaporated last year when Llama 4 was widely regarded to be a dud.

That set off a chain of events inside Meta. Scale AI CEO Alexandr Wang joined the company to head a new research organization, Meta Superintelligence Labs. Llama gave way to a new family of models, starting with Muse Spark. But they were closed cloud services, not open-weight.

With Muse Spark Open’s arrival, Meta has a shot at rekindling the excitement over Llama. By running open-weight models on their own infrastructure, companies can avoid uploading confidential data—from medical records to unreleased movie footage—to third-party services. They can tweak the model for specific purposes. And since they own the hardware, they can gorge on tokens without racking up any fees.

During a media briefing yesterday, I got a demo of some of Muse Spark Open’s capabilities. It wrote a script to perform math, helped code a social feed app, and selected video clips and edited them into a longer sequence. It all worked, though I’ll reserve judgment on how well it performs these sorts of tasks compared to other models until hands-on reactions from AI developers start to come in.

But the demo I saw was notable in part because it took place at Apple, not Meta. Muse Spark Open was running on two Mac Studio computers cabled together into a cluster—essentially a tiny desktop data center.

Historically, Apple and Meta have not exactly been allies. Indeed, when Apple executives such as Tim Cook touted the company’s commitment to privacy, it was seen as an implicit attack on Meta’s business model. The ability to run local AI that rivals cloud AI has unexpectedly aligned these two very different companies’ interests.

When Apple released the first Mac Studio in March 2022, running AI models capable of autonomously writing software and editing movies was presumably not on its mind. Even the original version of ChatGPT had not yet been announced. But the company certainly was thinking about artificial intelligence. The Apple-designed M1 Max and M1 Ultra processors available for the Studio both sported Neural Engine units optimized to run local AI.

The Mac Studio and two of Apple’s AI-optimized chips, the M6 and M5 Ultra. [Photo: Apple]

Fast-forward to this year. Using Macs as dedicated AI machines has become a phenomenon, and Apple has been leaning into it. At the consumer end of the spectrum, serious AI types are using agents such as OpenClaw on Mac Minis, leading to shortages of the diminutive computer.

Building a cluster of two or four Mac Studios to run Muse Spark is an industrial-strength version of the same idea. Practically speaking, it requires a total of 1 TB of RAM between the computers; any sufficiently equipped setup will cost tens of thousands of dollars.

Still, it’s an unusually approachable way to run your own local AI, because Mac Studios are, well, Macs. They run MacOS, plug into standard wall outlets, and can sit stacked on a desk rather than requiring a rack.

Along with further optimizing the newest Mac Studio models for AI, Apple has created a software technology called MLX to optimize its computers to run machine learning software such as LLMs. The company worked with Meta to make sure Muse Spark Open operates well on a Mac Studio cluster, but MLX also helps less resource-intensive models such as China’s Qwen run on less over-the-top hardware, such as a single Mac Studio.

Mac Studios aren’t mandatory for Muse Spark Open, which can also run on non-Apple infrastructure. Still, given the Mac’s burgeoning popularity as an AI machine—and Apple’s investment in AI from the silicon on up—it’s easy to see why Meta would prioritize Mac Studio setups. It’s equally obvious why Apple, which hasn’t built a general-purpose LLM of its own, has a powerful incentive to offer the best infrastructure for Muse Spark Open and other open-weight models.

For Apple, making nice with Meta on this particular front doesn’t just boost the fortunes of its highest-end computers. It’s also a sign that the company, which was long derided for being “behind in AI,” might have outfoxed pundits who thought that success was synonymous with taking on OpenAI, Anthropic, and Google—and Meta—as a frontier AI lab. As long as AI helps drive sales of Apple’s platforms, it’s winning.

The smartphone market suffered its worst decline on record in the September quarter, as rising memory costs pushed shipments down 22%, with worse to come, according to market researcher IDC. Memory chip prices increased by a stunning 350% over the past year, IDC says, because of competition for memory from AI industry players.

That’s cut into smartphone manufacturers’ profits and forced them to raise prices, cut back on specifications, and discard budget-focused models, says Nabila Popal, senior research director for Worldwide Devices at IDC. Chinese brands like Xiaomi Corp. and Oppo, which sell lower-priced phones with smaller margins, took the worst hit among the top mobile brands. Apple and Samsung were more insulated from the memory price increases because their phone portfolios lean toward higher-margin premium devices.

The same handful of manufacturers that make DRAM and NAND memory chips for smartphones—Samsung, SK Hynix, and Micron—also make memory needed for AI processing. Specifically, the manufacturers use layers of DRAM to make high-bandwidth memory (HBM) chips, which provide the high data-transfer rates required for AI workloads.

AI accelerators, such as Nvidia’s graphics processing units (GPUs), require large quantities of HBM to store model parameters and retain contextual information during problem-solving. HBM is physically integrated into the same package as the GPU, sitting right beside its processing silicon.

Memory manufacturers have been pivoting toward HBM because making it for the AI industry is more profitable than making smartphone memory. “The AI infrastructure and cloud service providers like Google, Meta, and Amazon that are doing these buildouts, the majority of them have secured long-term agreements [with the memory manufacturers] all the way up to like 5 years, to 2030 or 2031, pretty much guaranteeing the output will be given to them at certain price points,” Popal explains.

Memory prices are still rising, in part because the AI data center buildout is just getting underway. IDC believes worldwide spending on AI infrastructure, including servers and storage, will reach $1.21 trillion annually by 2030.

Memory manufacturers, Popal says, are building additional capacity but aren’t motivated to move very quickly because they’re benefiting from inflated memory prices.

Smartphone consumers will ultimately shoulder the costs. Already, the average selling price of smartphones has risen dramatically. “Last year it was $450,” Popal told Fast Company. “This year we’re expecting smartphone ASP to be $581.”

“They see you when you’re sleeping. They know when you’re awake. They know if you’ve been bad or good…” No, we’re not changing Santa’s pronouns. We’re talking about the AI-powered gadgets that are likely to be overflowing from his sack this holiday season.

AI toys began appearing on the market last year, quickly prompting warnings from child advocacy groups. But the 2026 holiday season will bring a much wider selection of AI gadgets for kids and adults alike, while AI companies are also pushing the technology as a shopping assistant.

This week has brought several announcements from Big Tech companies unveiling new AI-powered devices. Apple introduced a slate of smart home products, including the HomeView, which features a seven-inch screen and lets users control doorbells, thermostats, and smart locks, all powered by its AI-enhanced Siri.

Amazon is following a similar path with its introduction last week of new Alexa Tablets, which run on Alexa+, an AI-powered version of the company’s longstanding digital assistant. Simply say “Alexa” or tap the “Ask Alexa” button, and the chatbot can answer questions, control your smart home, and more. Prices for the three models range from $229 to $500.

Microsoft, which earlier this summer began testing a wearable AI device for office workers, is hoping shoppers spring for a new Surface laptop equipped with an Nvidia AI chip and a $2,599 price tag. The Surface Laptop Ultra, introduced last week, begins shipping October 16. Meanwhile, Meta is introducing a Windows version of its Muse AI agent, looking to capitalize on the tool’s popularity in Apple’s App Store.

Meta, of course, has much more in store for the holiday season. The company has big hopes for its Muse Charm, a Tamagotchi-like keychain with a two-inch display and a pair of cameras. Beyond the usual chatbot functions, the gadget can help declutter your inbox or make a dinner reservation.

Meta has also rolled out a new line of AI-enhanced eyewear, including models without cameras, following numerous complaints about people being filmed without consent. The company plans to offer 100 different options for AI glasses by the end of the year.

It has competition in the smart glasses market, though. Snap has introduced Snap SPECS, augmented reality glasses that can recommend restaurants or provide directions to get you from Point A to Point B. They’re available for preorder now and expected to ship later this fall.

One AI gadget that won’t make it onto shelves this holiday season is OpenAI’s rumored smart speaker, being designed by Jony Ive, the designer behind many of Apple’s biggest products. OpenAI hasn’t formally confirmed the device, which it reportedly hopes could eventually replace the smartphone, but it’s widely expected to arrive next year.

AI gadgets aren’t limited to Big Tech, either. The Plaud NotePin is a wearable AI voice recorder and note-taker that boasts 20 hours of continuous recording, 64GB of internal storage, and transcription capabilities in 112 languages. It sells for $159.

And once again, there are plenty of AI-powered toys your kids could be clamoring for as the holidays draw near. Among them:

• Luna Robot Pet Dog, a $499 ChatGPT-powered robotic companion equipped with a high-definition camera to help it understand its surroundings.
• Ted, a screenless teddy bear equipped with an AI chatbot and both push-to-talk and auto-listening modes. It can answer questions and even read bedtime stories to kids ages 3 and up.
• Anki’s Vector 2.0, a desktop robot that looks a bit like a bulldozer but integrates ChatGPT and has facial recognition capabilities.
• Miko 3, an AI-powered “educational partner” aimed at kids ages 5 to 10.

Child advocates, however, are once again urging parents to steer clear of AI toys. A report from Common Sense Media warns that they “pose unacceptable risks for young children,” citing concerns about privacy, emotional attachment, and reliability.

Hello again and welcome back to Fast Company’s Plugged In.

Once a year, Google holds a conference on its campus called Research@. Scientific researchers from inside and outside the company pack an on-campus auditorium for a highly compressed look at what its 26-year-old Google Research lab has been up to. This year’s edition emphasized AI—surprise!—and covered a vast amount of ground, from predicting natural disasters to fact-checking technical documents to helping doctors better serve patients.

After the keynote, I sat down with James Manyika, the member of Google’s senior leadership team who oversees Google Research. (His title—senior VP of Research, Labs, Technology & Society—conveys the breadth of his responsibilities.) He told me about Google Research’s mandate and approach to its work, and why he believes AI has already resulted in more scientific progress than most people realize. Our conversation has been edited for length and clarity.

For the sake of people who didn’t get to attend this conference, talk a bit about what Google Research is all about. You’re part of Google and Alphabet, and there’s research going on all over both companies beyond your group.

Google has had Google Research since almost the very beginning. The founders initially liked the idea of having a research organization, almost like Bell Labs. That’s what Google Research has been.

The way we think about the mandate and scope of Google Research is both foundational and applied research in a wide range of areas: things to do with fundamental computer science, AI, and machine learning, but also science itself. A lot of the work informs and helps our product areas. Things like how we think about knowledge graphs and search and retrieval of information. The fundamental computer science that goes into designing our infrastructure and hardware. The team that invented the transformer was in Google Research, so all of that too.

The parts that people aren’t as aware of are the scientific work. My favorite example is the fact that Google Research has had a genomics program for the last 12 years. We’ve had a connectomics program, to understand the wiring of the brain, for 10 years. And we’ve had the moonshot to build a quantum computer for 13 years.

One other thing that Google Research has as its mandate is a fair amount of translational work. The science and the research is fantastic, but a lot of work in crisis response, community resources, and resilience happens in Google Research.

Some of the stuff on the stage this morning involved issues like firefighting and medical AI. To what degree are you focused on these kinds of big important jobs for the world that are not always directly related to Google’s core as a business?

To me, all of those are fundamentally related to Google’s mission sense, which is organizing the world’s information and making it universally accessible and useful.

Things like flood forecasting are actually part of a moonshot to say, “What if no human or computer on Earth was surprised by a natural event?” Then that translates into, “Okay, so what are the natural events?” It’s a list that includes everything from earthquakes to floods to wildfires to monsoons to cyclones. We’re trying to work toward all of those. We’ve made more progress in some than others.

It’s the same thing with health. What if all diseases could be diagnosed, detected, understood, and eventually treated and cared for? What would that look like? That drives the fundamental research, the work on language and accessibility. Again, that goes back to the mission, which is, if you’re going to organize all this information and make it useful, boy, it better be in all the languages. And it better [overcome] every kind of access barrier, which we think of not just as language, but also things like [helping] people with disabilities, for example, speech challenges.

One moonshot right now is supporting 1,000 languages. A couple weeks ago, we hit our most recent milestone, which is 300 languages. We’re lucky in the fact that this is a company that’s wired to think about how we do these ambitious things.

A lot of the news onstage involved jobs that AI is already good at. But there was also discussion of it not always being as accurate as we’d like, and that it may disrupt jobs and the economy in ways that are going to cause issues. What percentage of your work and thinking relates to AI potentially causing problems and trying to avoid them?

A lot. In fact, we spend probably as much time as we do on the exciting possibilities. We think of them in a few different categories—the performance challenges and risks that could cause harm.

If you think about hallucinations, that could be problematic. But we also think about other categories of challenges and risks that have to do with misuse of this technology. That’s why you see us invest across Google Research and Google DeepMind on things like privacy and SynthID, which we’ve expanded beyond our own systems.

Are there technical solutions and innovations related to misapplication and misuse? Of course. But you need more than just that. You need society to come to collective views about how we deal with those things. We try to advance the ambitious, useful things, but also work on the downsides and the complexities that come with this.

How do you engage with researchers outside of the company?

I’ll give some concrete examples. I mentioned connectomics, the wiring diagrams of brains. We’ve had a deep collaboration with Howard Hughes Medical Institute and also the Lichtman Lab at Harvard for over a decade. When we published recently the work on Drosophila, the fruit fly, it’s the result of 10 years of deep collaboration. The work that Quantum team does has been a decades-long collaboration with academics at Caltech and University of Chicago.

Michael Brenner, who was on stage, is a full professor in applied mathematics at Harvard. He’s been there for 25 years. He’s also been a fellow here for the last 10.

There are a lot of people out there who are worried about AI, or skeptical, or just don’t see value in it. If they understood some of these scientific accomplishments, would they feel better?

It will help, but but it won’t be enough. I just think these scientific things need to be more understood. But the indirect nature of that will always make it arm’s length, not like a chatbot you interact with every day.

I think the bigger challenge and opportunity that we have is with regards to things that people engage with tangibly and directly. When it comes to work, the labor market effects, and education, people will experience that quite directly. That’s where we have to make a lot more progress.

I think it doesn’t help if some in the field itself make pronouncements that aren’t being borne out. The one that I find worrisome is the labor market impact, because most of the apocalyptic pronouncements on that front haven’t happened yet, and it’s unclear whether they will happen.

That’s not to suggest that we should sit back and relax. I think we should think about those changes in the labor markets and plan for them. Training is part of it. We should think about the role of safety nets.

Those direct things are probably a big part of the anxiety and concerns that people have. We have to make progress on those things and show that in fact there are paths through this.

You’ve been reading Plugged In, Fast Company’s weekly tech newsletter from me, global technology editor Harry McCracken. If a friend or colleague forwarded this edition to you—or if you’re reading it on fastcompany.com—you can check out previous issues and sign up to get it yourself every Friday morning. I love hearing from you: Ping me at hmccracken@fastcompany.com with your feedback and ideas for future newsletters. I’m also on Bluesky, Mastodon, and Threads, and you can follow Plugged In on Flipboard.

More top tech stories from Fast Company

Building an AI data center costs billions. Defending it is another story
From Russia to the Persian Gulf, drone attacks are forcing governments and technology companies to rethink how they protect their most valuable computing infrastructure. Read More →

Tin Can’s next product brings a relic of phone culture to Gen Alpha
The brand that made landlines cool again is introducing the Rolodex to a new generation. Read More →

Nintendo just made an unusual decision—and it’s going to be good for business
As Sony moves to phase out physical game discs, Nintendo is reassuring fans that cartridges aren’t going anywhere. But its commitment comes with some caveats. Read More →

What happens when your employees’ AI agents start coming to work? (Spoiler: It’s bad)
With 82% of organizations running AI agents they don’t know about, personal agents like Muse and Dots are about to make a hidden problem much bigger. Here’s what to do now. Read More →

He helped build Meta’s smart-glasses technology. Now he wants to take on Big Tech’s grip on personal AI
With Awear AI, Ash Saulsbury is building a platform that would let smart-glasses users take their AI assistants’ memories across devices and brands. Read More →

Trump’s new ‘superintelligence’ website is a rebrand fail
The newly branded SI.gov website has a problem: Its messaging is out of sync with public opinion. Read More →

The U.S. is on track to add a record 86 gigawatts of new wind, solar, and battery storage capacity this year—roughly the size of the U.K.’s entire grid. And there’s far more clean energy in the pipeline, with low-cost proposed projects that could help meet the surging demand from data centers. So why are data centers increasingly turning to polluting gas plants instead of clean energy?

A startup called Firma Power says that the issue isn’t cost, speed, or reliability, but the difficulty of assembling enough clean energy from a patchwork of projects—a problem that it is trying to solve.

Securing enough reliable, clean electricity for a massive new data center is complex. But Firma’s software calculates what a proposed data center needs, and then automatically searches through proposed clean energy projects to suggest a portfolio that could support it. An analysis that might otherwise take a team months can now take minutes. Firma then does the hard work of pulling all of the clean energy projects together, and data center developers can sign a single contract.

The company says that the process can get data centers power faster than building new gas-fired power plants—and it can compete on cost. And at a time when Americans really don’t like data centers, using clean power could make it more feasible for some projects to be greenlit by communities.

Why data centers aren’t tapping into clean energy

“We started the company with the hypothesis that despite the fact that 98% of what’s in the queue is wind, solar and storage, that’s not getting tapped as the default path to power for large loads—so there’s got to be some market frictions and barriers there,” says Jesse Jenkins, an energy systems engineer and associate professor at Princeton University who cofounded Firma Power and serves as CTO. “What are they, and can we build the platform that overcomes those?”

Firma cofounder and president Scott Burger, who previously worked at the pioneering battery company Form Energy, reached out to Jenkins, a former MIT classmate, to talk about the challenge as the data center boom grew in 2024. Burger’s modeling showed that clean energy should be economic and scalable for data centers; Jenkins said that he saw the same thing.

At the same time, they knew that tech companies with clean energy goals were struggling to stay on track. Burger and Jenkins partnered with a third cofounder and CEO, wind and solar developer Brent Alderfer, to launch a startup to tackle the problem. Drawing on more than a decade of their own research modeling the power system, they built custom software called FirmaOS. The startup, based in San Francisco and Princeton, New Jersey, raised around $1.4 million in a seed round in February 2026.

They realized that data center developers and tech companies were struggling with the complexity of the current system. Tech giants have been buying renewable energy for data centers for years, but the situation is different now. In the past, tech companies typically signed contracts for renewable energy credits. The power from a wind or solar farm helped offset the emissions from energy used at a data center, but usually wasn’t expected to be available 24/7 to keep the data center running. (In some cases, companies paid for clean energy projects on a completely different grid.) Securing reliable clean energy to actually power a data center around the clock is a very different challenge. While some companies are working on the latter approach—notably Google—it requires a large team and a lot of effort, and can’t easily be replicated across the industry. Firma wanted to change that.

The scale is also different now. The power demand for a “large” data center in the past might have been 50 megawatts. Now a single AI data center might require 500 megawatts, or even a gigawatt—roughly as much as the entire city of San Francisco. Getting enough clean power requires working with multiple projects and developers, not something like a single solar farm.

How Firma turns a patchwork of projects into a power supply

When customers come to Firma, they typically come with a site for a data center in mind, and share the size. For example, a project might need 300 megawatts in 2028, scaling up to 1 gigawatt by 2030. Firma looks at the local utility’s requirements for bringing a large customer online, and runs its software to calculate the generating capacity needed to serve the data center and which clean energy projects could supply it, while optimizing for cost. Then the team reaches out to each renewable developer to confirm if the project is on schedule and can be built, and starts signing contracts. (In some cases, Firma may buy and complete some of the projects itself.) Over the length of the contract, it manages the portfolio, adding more capacity if needed.

“Because of our unique model, we can deliver executable load-enabling strategies in days,” says Burger. “The data centers come online between 2028 and 2029. The usual alternatives, waiting for a utility’s planning and procurement cycle or buying project by project, can take years—indeed, we often hear from customers that say ‘our utility told us to wait until 2032 for a new gas power plant.'”

The startup is now working on around a dozen projects with customers including hyperscalers, AI cloud companies, and data center developers. While the company declined to publicly share the details of specific projects, it’s already assembling portfolios of solar, wind, and batteries on grids where customers would have had to wait much longer to access gas power. In cases where data centers were considering behind-the-meter gas plants—in other words, managing their own power plants—Firma’s approach also has the advantage that it lets customers rely on the infrastructure of the larger electric grid.

Bringing new clean energy to the grid

The approach helps bring new clean energy to the grid that likely wouldn’t have been built otherwise. More clean energy is in planning now than the grid actually needs, Burger says. (In three grids alone, PJM, SPP, and MISO, Firma calculated that around 200 gigawatts of clean energy could theoretically be added to the grid by 2030 and meet requirements for reliability, far more energy than the entire country needs.) Most of the projects in the queue don’t get built.

“Projects drop out for many reasons, such as grid upgrade costs, permits and local opposition,” Burger says. “One of the most common is the simplest: no buyer. Without a long-term contract, a developer can’t finance the project. In some cases, that’s because a developer couldn’t find a customer to sign a contract for the power.”

Firma can help rescue some of those projects. Utilities buy power for their existing customers; “bring-your-own-power” customers like data centers can support different projects, adding new clean energy to the grid.

Since the process is faster and cheaper than developing new gas power plants right now, it’s likely that more developers will move in this direction. “We don’t actually even need to market it as clean,” says Jenkins. “It is clean, and it’s important to many of these customers that it is. It’s increasingly important to the communities that live next to these data centers. But it’s also just economically competitive and basically the only way besides gas power plants to get to tens of gigawatts of deployable scale before 2030.” In the longer term, he says, solutions such as advanced geothermal can also help, but they haven’t scaled up yet.

“This should be the easy button solution,” Burger says. “We should see more folks looking to wind/solar/battery.”

Personal AI agents are having a moment. Meta recently launched a personal agent it named Muse and saw an estimated 2.8 million installs in its first 12 days, outpacing ChatGPT’s own mobile debut. Meanwhile, Instinct, a startup developing a personal agent you can text or call, just raised $1 billion at a $10 billion valuation. SpaceXAI introduced its Grok Bot agent in August, and most recently, on September 29, OpenAI released Dots, agents built to pursue goals in the background with minimal oversight.

So, what are personal agents? A simple way to understand them is to think of them as employees that are always on. You give them a goal, you give them the access and tools they need to achieve that goal, and then you get out of the way and let them get to work. The key new piece of the puzzle, the thing that is driving the hype, is that these goals are real-world goals—things like shopping, making dinner reservations, replying to emails, filing insurance claims, and so on.

The potential productivity gains are enormous. But there’s no such thing as a free lunch, and the upside comes with serious risks that business leaders must urgently address. This piece tells you what you need to do.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/creator-faisalhoque.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/faisal-hoque.png","eyebrow":"","headline":"Ready to thrive at the intersection of business, technology, and humanity? ","dek":"Faisal Hoque’s books, podcast, and companies give leaders the frameworks and platforms to align purpose, people, process, and tech—turning disruption into meaningful, lasting progress.","subhed":"","description":"","ctaText":"Learn More","ctaUrl":"https:\/\/faisalhoque.com","theme":{"bg":"#02263c","text":"#ffffff","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#ffffff","buttonHoverBg":"#3b3f46","buttonText":"#000000"},"imageDesktopId":91420512,"imageMobileId":91420514,"shareable":false,"slug":"","wpCssClasses":""}}
The risk-reward trade-off

The value of a personal agent is directly related to the access and permissions you give it: The more an agent knows about your relationships, schedule, and spending, the more useful it becomes. For example, it would be incredibly useful for frequent travelers to have their flights and hotels organized by agents, but this is only possible if the agent can check calendars, write messages, and make payments.

But this same access can make agents dangerous. In February, for instance (and somewhat ironically given how hard Meta has pushed Muse), Meta’s own director of AI alignment watched an agent she had connected to her inbox start deleting it, ignoring her instruction to check with her first.

The point here is not that the risk is so great that we should never use personal agents. The point is rather that the risk cannot be separated from the reward. An agent that can’t read your inbox can’t clear it, and one that can’t spend money can’t book your flight. Every permission that makes an agent more useful also widens what can go wrong.

Personal agents will come to work

The issue for enterprise is that the access people grant doesn’t stop at their personal lives. In a recent survey of knowledge workers, employees reported giving AI tools direct access to company email, cloud storage, and customer databases, and more than half of those using unapproved AI tools said they had shared internal messages and emails with them.

And leaders may be overestimating how much of this is visible and being managed—in the same survey, 90% of executives said they were confident in their visibility into AI tools, while 52% of employees admitted to using AI tools without approval. Meanwhile, a Cloud Security Alliance survey found that 82% of organizations have AI agents running in their systems that they don’t know about, and 65% had suffered an agent-related security incident in the previous year. Jim Taylor, president of the security company RSA, describes learning about a midsize global bank that believed it had no AI agents, only to find out it was running more than 4,000.

These surveys cover AI tools and agents of all kinds used at work, but they don’t isolate personal agents such as Muse and Dots; indeed, the surveys were done before these products became widely available. But the implications for personal agents and enterprise are clear and stark—people are already hooking up their personal agents to their company emails and databases. When they do, the agent’s risks become the company’s risks.

3 risks leaders need to understand

Here are three of the most important risks that come with personal agents.

1. Your data leaves the building. This risk needs no attacker. Connect a personal agent to work email, and copies of company data go to a vendor your company has no contract with, sitting outside its procurement and compliance rules. Muse users, for instance, must opt out if they don’t want their interactions used to train Meta’s models, and we can be certain that many will forget to do so when using Muse in work contexts.

2. Outsiders can give it orders. Anyone who can email your employee can try to inject their agent instructions. One early Instinct tester emailed his own inbox from a brand-new Gmail account with instructions meant for the agent. After seeing how easily it could be phished, he deleted his account. Even Meta, which built layers of defenses into Muse, says “prompt injection remains an open problem in the industry.”

3. It’s a master key. Whoever takes over an agent inherits everything it’s connected to. Last month, a researcher showed that any app or command running on a Mac could hijack Muse and the permissions its owner had granted, and steal the token that logs into the user’s Muse account. Agents can also read the inbox where security codes arrive: Instinct was observed pulling a sign-up code from a tester’s inbox to finish a restaurant booking. In the wrong hands, that ability could unlock accounts the agent was never connected to.

The first risk can largely be handled through vendor choice and settings. The other two risks can’t be removed without giving up what makes agents useful—but they can be reduced. In all three cases, the key is setting clear limits on what agents can touch and do.

4 things to do now

Each of the steps below targets one or more of the three key risks. None requires new software, and all can start this week.

1. Find every agent this week. Since leaders may see less than they think, start by looking. Ask IT to pull the list of third-party apps connected to company email and calendar accounts. At the same time, send staff a short, no-blame survey: “Which AI agents do you use, and what have you connected them to?” By the end of the week, you should have one initial list.

2. Publish one page of agent rules. This is where you set the limits. Sort agents into three lists: approved, approved with limits, and banned. “Approved with limits” might mean authorized for use with personal accounts only. Ban connecting any agent to finance systems, customer data, password managers, or the inbox where security codes arrive; those are the connections that turn an agent into a master key. Approve only vendors that let you delete your data and opt out of training. Name one person who owns the rules page, and update it every month.

3. Require a human “yes” before any action. Because outsiders can try to steer an agent, keep a person between it and anything consequential. Agents may read, sort, and draft. But a person must approve before an agent sends email outside the company, spends money, shares a file, or accepts terms. And where possible, give the agent its own email address and forward it only what it needs, instead of opening up your whole inbox.

4. Practice switching agents off. Data can outlive the connection, so test whether disconnecting actually works. Pick one agent. Revoke its account access, change any passwords it used, and cancel any payment card linked to it. Twenty-four hours later, check whether it is still acting or still holding data. Ask the vendor in writing to delete your information. Then add “disconnect all agents” to the checklist you use when employees leave the company.

Set the terms now

Personal agents will soon be as ordinary as smartphones. Indeed, the problem reminds me of the bring-your-own-device problem of the 2010s, when employees started using their personal phones for work email. But there is a critical difference here: A phone could store your company’s data—an agent can act on it.

Just as people were using their own phones at work a decade ago, your people are already using these agents. Either you set the terms on which personal agents enter your business, or you let others do it for you.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/creator-faisalhoque.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/faisal-hoque.png","eyebrow":"","headline":"Ready to thrive at the intersection of business, technology, and humanity? ","dek":"Faisal Hoque’s books, podcast, and companies give leaders the frameworks and platforms to align purpose, people, process, and tech—turning disruption into meaningful, lasting progress.","subhed":"","description":"","ctaText":"Learn More","ctaUrl":"https:\/\/faisalhoque.com","theme":{"bg":"#02263c","text":"#ffffff","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#ffffff","buttonHoverBg":"#3b3f46","buttonText":"#000000"},"imageDesktopId":91420512,"imageMobileId":91420514,"shareable":false,"slug":"","wpCssClasses":""}}

The state of Florida won a key court decision Wednesday that allows it to enforce a ban on social media accounts for kids under 14 and other restrictions.

The law was one of the most restrictive in the U.S. on social media use by children when Gov. Ron DeSantis signed it in 2024.

The law implicates the First Amendment, an appeals court said, but “likely does so in a content-neutral manner.”

It would likely survive a full legal challenge “because it doesn’t burden substantially more speech than necessary to advance the important interest in protecting minors from social media addiction,” a panel from the 11th U.S. Circuit Court of Appeals said in a 3-0 opinion.

The law bans social media accounts for Florida children under 14 and requires parental consent for use by kids who are 14 or 15.

Industry trade groups the Computer & Communications Industry Association and NetChoice sued Florida, arguing that the law violated the free-speech rights of social media providers and users. A judge in 2025 said the law likely was unconstitutional and granted an injunction.

That injunction was put on hold while the appeals court heard the case. The lawsuit, with the injunction now overturned, will return to federal court in Tallahassee.

The appeals court noted that the law covers social media platforms “that have been determined to induce heavy use by children, and applies only to platforms that use both algorithms and one of several addictive features.”

An email seeking comment from the Florida attorney general’s office was not immediately answered. NetChoice said it was disappointed with the decision.

“We are currently evaluating all of our options as we continue the fight to ensure Floridians’ rights to free expression online are protected, and parents remain in the driver’s seat when it comes to making decisions for their families,” said Paul Taske, a lawyer at NetChoice.

—Ed White, Associated Press

Welcome to AI Decoded, Fast Company‘s weekly newsletter that breaks down the most important news in the world of AI. I’m Mark Sullivan, a senior writer at Fast Company, covering emerging tech, AI, and tech policy.

Sign up to receive this newsletter every week via email here. And if you have comments on this issue and/or ideas for future ones, drop me a line at sullivan@fastcompany.com, and follow me on X @thesullivan.

AI Super PACs are paying millions for midterms ads, which almost never mention AI 

Political action committees backed by artificial intelligence developers, Silicon Valley executives, and venture capital firms have poured tens of millions of dollars into competitive 2026 midterm elections. But across the thousands of television, streaming, and digital ads bought with that money, the words “artificial intelligence,” “algorithms,” and “tech policy” appear in almost none of them, a New York Times investigation found.

Silicon Valley donors and corporate super PACs want to elect members of Congress who are likely to support industry-friendly AI regulation, or no regulation at all. But they must do so without further elevating AI as a campaign issue. The technology, and the data centers that support it, remain deeply unpopular with voters. Many Americans worry that AI will eliminate jobs, pose potentially existential risks, enable mass surveillance and media manipulation, and ultimately benefit only the wealthy elites who finance and control it. Wide majorities also oppose the construction of noisy, energy-hungry data centers in their communities. With AI policy now a political liability, advertisers are channeling their funds into conventional campaign ads focused on kitchen-table economic issues like inflation, local taxes, and small business support.

And we’re talking about unprecedented sums of money. A New York Times analysis of campaign finance data shows that political groups aligned with major AI developers like OpenAI and Anthropic, along with tech-focused venture capital firms, have directed between $55.7 million and $60 million into federal midterm campaign advertising and voter outreach. According to Public Citizen’s campaign finance tracking, the sector’s flagship super PAC, Leading the Future, raised $75.1 million this cycle, roughly two-thirds of it from the venture capital firm Andreessen Horowitz. Last year, OpenAI’s cofounder and president, Greg Brockman, and his wife, Anna, contributed $50 million in personal funds, splitting the money between Leading the Future and the primary Trump-aligned super PAC, MAGA Inc.

Silicon Valley’s super PAC money also doesn’t buy as much advertising as ordinary campaign cash. Federal Communications Commission rules require broadcast stations to charge super PACs undiscounted market rates, while granting candidate campaigns statutory “lowest unit rate” discounts. Because outside groups can pay up to five times as much per ad slot as candidate campaigns, tech-backed PACs are concentrating their media buys on broad, high-resonance economic messaging rather than risking millions of dollars on a topic that triggers immediate public resistance.

Under the Trump administration’s arrangement with Big AI, AI developers and infrastructure companies shower the president with cash and praise, while Trump does everything in his power to keep AI regulation voluntary, nonbinding, or nonexistent. Late last month, for example, the administration brought tech titans to Washington for a luncheon and ceremonial signing of a voluntary safety pact asking advanced AI labs to follow safety protocols and establish internal teams to monitor their own progress.

Anthropic says AI worms could soon spread across smart gadgets

Anthropic published research on Monday suggesting that small AI models may be just months away from being able to replicate themselves across cheap, internet-connected devices without relying on data centers. Researchers warn that these models could spread across devices such as smart ovens and off-brand Android TV boxes, making them difficult to shut down because there would be no central point of control. In Anthropic’s tests, GLM-5.2, a large language model developed by the Chinese AI company Z.ai, and other powerful models successfully replicated themselves across devices. Researchers predict that smaller models could achieve the same capability by the end of the year.

Huang says Nvidia will announce $2 billion in European deals

Nvidia’s CEO, Jensen Huang, said on Tuesday the company expects to unveil $2 billion worth of European AI deals in the next week, on top of the $8 billion it has deployed across the continent over the past year, Bloomberg reported. Speaking in Paris—first stop on a tour that also takes him to London and Berlin—Huang announced investments in four French startups, according to Quartz. Nvidia is already a major investor and chip supplier for the French AI company Mistral. “Europe needs to get in the race,” he said, adding that open-source models may be Europe’s best route to catch up with U.S. developers. Nvidia, of course, would benefit from a more intense global AI arms race regardless of who wins, because all players need Nvidia chips.

Hundreds of scientists call for an international AI containment body

More than 400 scientists signed an open letter calling for an international body to standardize containment practices for cyber-capable AI models and establish protocols for tracking, documenting, and reporting containment failures. The letter urges organizations developing, deploying, and testing such models to immediately adopt high-assurance containment systems, default to airgapping, and report containment failures to the proposed body; those incidents would eventually be made public. As of Wednesday, the letter’s website listed 406 signatories, including Geoffrey Hinton, Yoshua Bengio, and Stuart Russell.

Google releases Gemini 4 Argon

Google released its latest frontier model, Gemini 4 Argon, on Tuesday. According to Google, Argon ranks first on the Vals Index, which measures performance across finance, coding, legal, and tax tasks. The company says the model helped its software engineers increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. Google also says Argon helped the Chrome team complete an architectural refactor spanning hundreds of files, compressing months of work into weeks. Argon is available globally through Google’s application programming interface (API), as well as through the Gemini app and AI Mode in Search for Ultra and Pro subscribers.

Chinese companies bought more than 1 million Nvidia gaming GPUs in a year

Chinese tech companies have purchased more than 1 million Nvidia gaming and workstation graphics processing units (GPUs) to make up for shortfalls in AI computing power, the Financial Times reported. China’s domestic AI chip production has doubled this year but still cannot keep pace with demand from the country’s largest tech companies, leaving a shortfall of nearly 1.5 million chips against demand exceeding 5 million. The stopgap chips are primarily Nvidia’s RTX Pro 6000D, RTX Pro 5000, and RTX 5090, all of which comply with U.S. export controls. While they cannot match the performance of Huawei’s Ascend 910C, they are useful for simpler AI workloads.

More AI coverage from Fast Company:

• Apple’s HomeView is nearly a decade late. That might not matter
• Building a data center costs billions. Defending it is another story
• Rogue agents are the price of a breakneck AI arms race
• The list of AI insiders sounding the alarm on their own companies keeps growing

Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company today.

German authorities opened an investigation into possible sexual harassment after a robot controlled by two young men touched the buttocks of a 16-year-old girl and made thrusting movements toward her, police said.

Police in the western city of Cologne responded after the incident overnight Saturday to Sunday in which the robot’s suspected 16-year-old owner and a 19-year-old companion were said to have filmed the alleged assault.

They were said to have streamed the incident and commentated on it online.

Cologne police spokesman Christoph Bermel said the two male teenagers have not been charged but are considered suspects, and investigators were examining whether the robot’s movements were preprogrammed or controlled in real time.

Police were unfamiliar with this type of incident, he said.

“We have not been aware of this phenomenon so far,” Bermel told The Associated Press, saying it was “new territory for us.”

Emergency teams that were dispatched to the scene near a fast-food restaurant found the robot in the trunk of a vehicle, Cologne police said in a statement on Monday. The robot and the two young men’s mobile phones were confiscated.

North Rhine Westphalia Police Cologne, on Facebook, posted an image of the red-and-gold robot lying on the ground.

The robot, said to be 120 cm (about 3 ft 11 inches) tall, was dressed in a white tank top, black sweatpants and wearing a gold chain and sunglasses. It also had what appears to be a fake mustache.

Videos circulated on the internet show two young men escorting the robot toward a group of women, bumping into one and hugging others, on a lively nighttime street.

One of the men repeatedly shouts at the robot and pulls it backward after it makes sexual gestures, prompting laughter among people in the area.

On the police’s Facebook account on Tuesday, officers criticized inappropriate comments and warned that sharing or joking about such incidents can worsen the impact on any victim.

Associated Press writer Jamey Keaten contributed to this report.

—Daniel Niemann, Associated Press

On September 2, 40 minutes before Target Field opened its gates for a Minnesota Twins game against a division rival, the Detroit Tigers, a man opened fire nine blocks away in downtown Minneapolis, resulting in three deaths and six injuries, including law enforcement officers. 

“The number of emergency vehicles going by the stadium, headed that direction, was frightening. It was a nonstop cascade of sirens,” said Jeff Beahen, director of security for the Twins. 

Beahen, who served in law enforcement for 39 years in Minnesota, is familiar with quick decision-making. Although the security of the baseball stadium itself was not threatened, the proximity of the shooting and the emergency response could have compromised safety and travel logistics for thousands of visitors and employees on their way to the ballpark. 

Beahen had 40 minutes to decide whether the game should go on. Within a few minutes, he used a platform called Envoy to notify employees about the shooting, and 25 minutes later, he sent out a second notice that the situation was stable. Alerting employees in the building and those on the way helped Beahen make sure that risks were averted while an emergency meeting convened with the security team. The game was played. (The Twins lost in extra innings.)

Visitor management

The cloud-based workplace management platform Envoy was created in 2013 by an engineer, Larry Gadea. The Twins implemented the platform in 2019 for delivery and visitor management, then adopted its emergency response system in 2025. The Twins were part of the beta version of the new platform called Envoy Response, which officially launched in August.  

Response uses on-premise data to track the safety of employees in large-scale buildings, such as stadiums, data centers, and office buildings. More than 16,000 high-compliance workplaces around the world use Envoy, including 92 of the Fortune 500 companies, according to a spokesperson at the San Francisco-based company. Envoy assesses multiple kinds of threats or events, including shootings, chemical spills, or protests in the vicinity of a workplace. 

Target Field can host more than 40,000 people at maximum capacity, according to Beahen, and can accommodate well beyond 100,000 visitors over a weekend. Major League Baseball teams are required to get certified under the Department of Homeland Security’s SAFETY Act, Beahen says, which includes showing their stadiums have established practices covering life safety and evacuation protocols. When Beahen joined the Twins, his primary role was to get the certification. The Twins were the 12th MLB team to achieve this certification, in 2023. 

Rapid emergency response

On the day of the shooting, Beahen used Envoy to notify roughly 500 employees of the threat, and added some 78 people to the system who may have been in and around the stadium, such as coaching staff and scouts. It took him roughly 2 minutes to put together the message.

“When I think about the way the security industry has evolved, a lot of it has been on the basis of facility-level decisions,” said Bridget Scott Akinc, vice president of strategy at Envoy. Instead of relying solely on physical security measures, like card readers and cameras, the platform detects employees on-site through their connection to the network.  

Before implementing the platform, the Twins used an emergency notification platform that lacked directory integration. This meant contact lists were inaccurate, and notifications were sent to former employees. While not every incident is an emergency, the cohesive system made messaging more reliable and organized. 

Beahen notes the surge in emergency situations related to political and social unrest in Minneapolis. From the George Floyd murder in 2020 and the protests that followed, to the fatal shootings of Renée Good and Alex Pretti by ICE enforcement officers earlier this year, threat intelligence remains a crucial part of keeping Minneapolis citizens safe, including at venues where thousands of people come together.  

Expanding internationally

Several sports arenas across the country have adopted Envoy in their emergency response process, including those of the NFL’s Philadelphia Eagles and NBA’s Miami Heat. Several universities, media companies, and data centers have implemented its visitor management and security features, as has the computer chip company Nvidia. 

Envoy Response has been adopted by more than 300 organizations since its August launch, some of which have international office locations, like the online forum Reddit and the equity management platform Carta.

The company is looking toward other international partnerships for Response, particularly in Britain. After the Manchester Arena bombing in 2017, when 22 people were killed and more than 1,000 injured, the U.K. brought in “Martyn’s Law,” so known from the name of one of those killed. It requires certain public venues to implement security measures and staff training to protect from terrorist attacks. The legislation passed in 2025, and takes full effect in 2027. Ahead of that, U.K. venues are moving toward more documented risk assessments. Envoy published a guide to the law in June, positioning itself as a solution to these requirements.

The unique challenges of arenas

The platform accommodates different physical spaces, from offices in tall buildings to stadiums outside of cities. But arenas face particular challenges, with the movement of thousands of attendees and employees. While the platform is not a standalone solution for emergencies, it has made the logistics of dealing with nearby emergencies easier to track for the Twins, helping Beahen, the security director, make faster decisions.

Stadiums, Beahen says, are soft targets. “We don’t have walls that protect us, we have open gates . . . and we have to have preparedness and tools and strategies to mitigate that, knowing that someone may take advantage.”

For Beahen, the shooting near Target Field illustrates a broader reality of stadium security: Preparedness is no longer limited to protecting the physical perimeter. When threats can emerge suddenly and outside a venue’s walls, security teams need timely information, accurate personnel data, and a way to communicate with people before, during, and after an incident.

AI-powered smart glasses promise to keep your phone in your pocket, but their usefulness depends on how much work they can take off your day. Meta ties its glasses to its own AI services, while Google is courting eyewear makers with Gemini. As an AI assistant learns a wearer’s preferences, it can become more useful and harder to replace.

If that knowledge cannot move to a competing brand, choosing better glasses could mean starting over with an assistant that knows less about you.

Ash Saulsbury spent seven years as a systems architect at Apple before joining Meta, where he served as product lead for wearable reality products from 2017 to 2019 and helped incubate its smart-glasses technology. Although he left before Meta’s Ray-Ban products reached consumers, he describes the early project as an extension of the company’s advertising business; making content easier to create would encourage engagement and generate more advertising opportunities.

He later became vice president of architecture for Microsoft’s mixed-reality products, where his team worked with Qualcomm to develop chips capable of operating within eyewear’s power and thermal constraints.

His startup, Awear AI, wants consumers to be able to switch eyewear brands without losing the information their AI assistants have accumulated about them. The company emerged from stealth at Qualcomm’s Snapdragon Summit on September 23, introducing Agentic System One (AS-One), a software platform that eyewear makers can integrate into their products, beginning with glasses powered by Qualcomm’s Snapdragon AR1+ processor.

“You’ll see many companies go to enormous lengths to make it extremely difficult for their competitors to access your data,” Saulsbury tells Fast Company. “We make it extremely difficult for anyone, including ourselves, to access your data.” Rather than profiting from access to consumer information, he says Awear wants to make money by providing services people find valuable.

His cofounders include Lila Tretikov, Microsoft’s former deputy chief technology officer and former CEO of the Wikimedia Foundation, and Jan Eißfeldt, the foundation’s former global head of trust and safety. Together, they oversee Awear’s privacy and trust architecture. Its COO, Jennifer Van Beek, brings experience in luxury branding through her agency work with LVMH and other prestige houses.

MAKING PERSONAL AI WORK ACROSS DEVICES

Awear develops the technology eyewear companies need to bring personal AI assistants to their products. Its partners would design and sell the glasses, while Awear provides the underlying software, connected services, and adaptable hardware designs.

AS-One includes an encrypted communications network where AI agents can interact with people and one another. A separate memory vault is designed to preserve information across personal devices and cloud storage, although Saulsbury says photographs can remain on the devices where they originated.

“The agents you use should work for you, not for somebody else. Even if you decide to leave, you can take all that information with you and move it somewhere else,” he says. “You will always have visibility into what the agentic system knows about you.”

For example, a wearer could ask their glasses to arrange dinner. AI agents would use their preferences from previous interactions to make a reservation and send the details to their phone. Saulsbury envisions that interaction eventually extending to a car that could take the wearer to the restaurant. He says the booking capability already works, but provided no details about the vehicle integration such a handoff would require.

The platform also combines its own AI models with those from external providers, which Saulsbury declined to identify. He says Awear anonymizes those transactions, while its ability to switch between models could help eyewear brands operate across different countries.

PERSONAL AI NEEDS MORE THAN PRIVACY

Saulsbury says consequential actions, such as purchases, would require explicit user confirmation, with safeguards that do not rely entirely on a language model’s judgment.

“We don’t allow the AI agent to make a purchase without your direct authorization,” he says. “We want to make sure that even if you purchase something and later change your mind, or perhaps forget that you made the purchase in the first place, those decisions remain reversible and you have a way to recover from them.”

For Tretikov, the cofounder, protecting users also means ensuring those safeguards survive changes in ownership or management. She argues that users should retain control of their encryption keys, preventing even a future owner of Awear from accessing or exploiting the personal histories its agents have accumulated.

She says encrypted channels would protect information moving between devices, but emphasizes that portability does not mean every agent automatically gains access to everything another knows.

“Think of it much like the relationship between your doctor and your insurance agent,” she explains. “They have different responsibilities, and that does not mean they should automatically share everything they know about you.”

Apple’s reported plans to introduce privacy-focused smart glasses in 2027 add another dimension to Awear’s challenge. Saulsbury expects Apple to deliver a polished product, while Tretikov acknowledges the strength of its privacy protections. Her concern is that privacy within a tightly controlled ecosystem still leaves consumers dependent on the company operating it.

To address that limitation, she plans to make Awear’s memory-transfer format and data-sharing rules openly available. She draws a parallel with her experience leading the Wikimedia Foundation.

“One of the most important lessons from that experience is that Wikipedia is not trusted simply because people trust the Foundation. It is trusted because the content is structurally protected, because its licensing allows anyone to take all of that content and leave,” she says. “That is the same standard I want us to apply to personal memory.”

Asked whether opening parts of the platform would make it easier for competitors to copy, she responds: “Easier to copy? Partly. Good. If competitors adopt our export format, users win and the standard spreads. We win. Trust is the standard.”

For now, however, Awear has not identified a published technical specification or a competing service capable of accepting and using a person’s complete AI memory. Encryption could protect that information if the company is acquired, but ensuring continuity if Awear shut down would require more than secure storage or an export option.

CAN SMART GLASSES PROTECT BYSTANDER PRIVACY?

Eißfeldt, the other cofounder, says Awear’s “triple-lock” approach protects personal information in an encrypted vault, with users controlling the keys and deciding who can access their data.

Those protections would be built into Awear’s software, beyond the control of eyewear manufacturers using its platform.

“The manufacturer locks sit in the runtime we ship, below the brand’s layer. A brand configures the experience above it and doesn’t get a switch to turn off the protections,” he says. Devices that bypass the safeguards would lose their “Awear Certified” designation and access to the network. He says the company also intends to publish independent audit results.

That gives Awear a way to enforce its privacy commitments, while also placing the startup in the position of gatekeeper for the very ecosystem it wants to offer as an alternative to Big Tech.

The system also prohibits identifying people who are not public figures, Eißfeldt says, while recordings require explicit consent.

But what about people who aren’t wearing the glasses? Asked how the devices would avoid capturing a bystander’s private conversation, Eißfeldt points to the conditions required to activate the assistant, including voice recognition and a wake-word system.

“The AS-One-operated microphone on the eyewear on the user’s head would need to be very close to your example’s private conversation, and the voices would need to be ones the system recognizes and can recognize valid consent from,” he says. “There’s also a keyword spotter, and until you say your assistant’s name, it’s deaf.”

The glasses would also feature indicator lights facing both inward and outward, alerting the wearer and nearby people when the assistant is listening. Whether those safeguards can reliably prevent unwanted recordings remains to be demonstrated.

THE FIGHT FOR CONTROL OF AI EYEWEAR

Van Beek, Awear’s COO, sees an AI assistant’s behavior as an extension of a brand’s identity, something luxury eyewear makers may be reluctant to surrender to a technology company.

“Clients of these brands can see a product or campaign and recognize the brand in an instant, without needing to see their logo,” she says. “We’re giving brands the agency to ensure they don’t just exist on the outside of the product but are carried through on the inside as well.”

Her approach begins with the product a brand wants to create, rather than the technology it needs to accommodate.

But Awear is entering a market where larger competitors already have significant advantages. According to Counterpoint Research, Meta accounted for approximately 84% of global smart-glasses shipments in the first half of 2026. Google, meanwhile, offers eyewear partners access to Gemini and financial support, the latter including a reported investment in Gentle Monster, a Korea-based eyewear brand.

“We’re not asking them for their clients; we’re asking them to help shape their process,” Van Beek says, so brands can bring their own identity to the technology.

Competition could also come from companies closer to Awear’s technology stack. Qualcomm’s Snapdragon START offers software designed to support multiple AI models and shared memory, while Liquid AI is developing ways for agents to use personal context with user permission.

Saulsbury positions Awear as a bridge between the chipmaker’s technical expertise and the fashion industry’s design priorities. But Qualcomm’s expanding software ambitions could eventually narrow the opportunity Awear hopes to capture.

The startup has not disclosed its pricing or funding details, and the identities of its eyewear partners and their product release dates remain confidential.

If Awear succeeds, choosing a new pair of glasses could become a decision about the glasses themselves. Consumers could weigh comfort and capabilities without calculating how much their assistant would forget. Getting there requires other companies to support that continuity, even when it helps a customer buy from a rival.

Saulsbury has spent years helping build technology inside powerful ecosystems. Now his company’s prospects depend on making the knowledge accumulated inside one useful beyond its boundaries.

The first four years of the generative AI revolution have tended to see power flow in one direction. Big Tech firms unleashed their tools on the planet and expected users to like it or lump it. Unhappy with the amount of data collected about you in the pursuit of convenience from your chatbot? Tough luck. Squeamish about the provenance of all the data AI systems have been trained on, and whether it was legally obtained? Not a lot you can do about it.

But as the technology continues to advance, and users become more comfortable with its capabilities, that imbalance is starting to shift. And Big Tech might regret what it’s wrought. If the first few years of the generative AI revolution were about vendor lock-in and AI companies potentially exploiting intellectual property they had no right to use, the next few could see David start to fight back against Goliath.

Like many others, I bought a digital photo frame from Amazon this week, inspired by developer Wes Bos’s reverse engineering of a $50 photo frame he bought from the retailer and turned into a smart display. Within a couple of hours, and with the help of Claude, I’d sidestepped the built-in software designed to lock me into a recurring subscription just to view my own photos, installed an operating system of my choice, and repurposed the frame into a wall-mounted display of regularly updating headlines to keep me on top of the news agenda. I’ve also resuscitated a decade-old printer for which there are no compatible drivers for modern operating systems, and overridden proprietary software for a thermal receipt printer to turn it into an on-demand, tear-off to-do list.

Next up on that to-do list? Sic a swarm of AI agents on the locks that prevent 20 e-ink price tags, once mounted on supermarket shelves, from displaying what I want, and see if they can crack them. Then hack together a hardware system to power the tags, replacing the rail-based system that once sent them signals.

I’m not alone. Want a photo editor or graphics package like Adobe Photoshop but don’t want to pay the exorbitant monthly subscription for it? Folks have used AI to reverse engineer the way the app works and vibe-code alternatives of their own. Same with Word and any number of other apps. Want to play console games in a web browser? Sure. AI has helped with that, too. One of the hottest GitHub repositories right now is Reverse Engineer Anything (REA), which aims to make the process easier.

The maker movement has always existed, but AI’s ease of use helps smooth over the bumpier elements of getting into that space. (You can, for instance, now send a photo of a circuit board to ChatGPT and ask it to identify what does what, rather than pore through a manual yourself or undertake hours of trial and error with a multimeter.) But what’s even more interesting than people tinkering and tweaking is the glee with which many of us are doing it.

There’s a sense of moral rectitude that underpins much of the current trend of reverse engineering anything. Yes, AI makes it technically possible to do a lot of this stuff far more easily than before, but it’s the burning sense of longstanding injustice, of being screwed over by endless subscriptions and the inability to ever own anything, of being bilked and badgered for more money, that is fueling the movement. I take glee in giving the finger to an exploitative photo frame app and avoiding the risk of installing malware on my phone just to use something I bought. I feel morally justified in doing so because too many tech companies, big and small, have screwed me over before, so screw them. And better still, I have control over what happens to my device next.

Depending on the jurisdiction, some of this reverse engineering may not be legal, though consumers also have rights to tinker with what they buy. But many of us reckon we’re just paying back a fraction of what’s been done to us before. And if tech companies didn’t have any qualms about what they did for decades, why should we?

Since its founding in 2021, Colossal Laboratories & Biosciences has sought to bring back extinct animals, including the dodo, the iconic flightless bird once endemic to Mauritius. It was documented by humans when the Portuguese first arrived on the island in the 16th century, but hasn’t been seen in hundreds of years. The columbid (a 3-foot-tall member of the Columbidae family, which includes modern pigeons and doves) struggled to survive the arrival of new invasive species and human hunting, and, as the saying goes, went “the way of the dodo.” 

For that same reason, Colossal, the biotech/genetics company that already claims to have resurrected the dire wolf, wants to bring the dodo back. The company’s strategy relies on editing the genomes of other species (their living relatives) with the extant preserved DNA from the species they’re trying to regenerate. But the dodo is especially difficult: The company only has access to one dodo genomic sample and needs to apply its technology—previously applied to dire wolves conceived via surrogate—to a species gestated in an egg.

But Colossal announced on Wednesday that it has made progress toward that goal by using a new AI tool it built, called GenoMorph. The AI tool is trained on DNA from other species and can identify which sections of a genome are likely to be associated with different traits, like a beak or a feather. The scientists are using that technology to identify which regions of the available dodo DNA are responsible for the traits that make the dodo, well, dodo-ish. Thus far, they’ve whittled the list down to 2,300 DNA targets. (The company also says that, for the first time, it has used lipid nanoparticles to conduct genomic editing within an egg.)

This is all a reminder that what Colossal is doing isn’t complete genetic replication of the original dodo, but something more nuanced. The company’s strategy instead involves building what its leadership calls a “functional” dodo, filling in the genetic sequence of a living relative species, a type of pigeon, with selections of dodo DNA associated with specific dodo traits, like its iconic beak. 

Whether or not you think this strategy constitutes actually bringing back the dodo hinges on how broadly you’re willing to define a species. Some scientists have strongly objected to Colossal’s framing of its work, even as the company announces quests to bring back the woolly mammoth (a heavy-coated cousin to the elephant), the thylacine (Tasmanian tiger), and the giant Moa (a flightless bird from New Zealand that stood up to 12 feet tall and weighed nearly 500 pounds). But Colossal’s chief science officer Beth Shapiro counters that the distinction between different species is often fuzzier than we think. “It’s a species definition that’s based on an arbitrary threshold of sequence similarity,” she says. “But that threshold is arbitrary. It’s going to be arbitrary no matter who you are.”

Fast Company sat down with Shapiro and company CEO Ben Lamm to discuss Colossal’s latest dodo updates, the definition of species, and the way their company uses AI.

This interview has been edited for clarity and length. 

You’ve announced your quest to bring back the dodo a while back. What progress have you made today?

Beth Shapiro: As a biologist who spent most of my career staring at genomes, the real hard thing is how do I know—in this 2 billion letters of DNA that make a bird, or three and a half billion if it’s a mammal, or 4 billion if it’s an elephant—what it is that I need to change if I’m going to take a Nicobar pigeon and turn it into a dodo? And for some of our projects, like the mammoth, we can start by looking at all of the 100 mammoths that we have and say, “Okay, where are they all the same as each other but different from elephants?” That gives me a good idea where to start. 

But what do we do when we only have one genome, which is the case for the dodo [. . . ]? How do I look at that genome and figure out what it is that we need to edit if we’re going to take a Nicobar pigeon, which is a smaller, beautiful-colored, strongly flying pigeon with a tiny little beak, and turn it into the iconic dodo? 

What we did here was combine DNA sequence data that we had for a bunch of birds [and] 3D skulls that we have for a bunch of birds. This is obviously something that’s difficult for a human to be able to mark, but an ML [machine learning] algorithm can very quickly figure out how to summarize that 3D shape into something that a computer can intake and output. We also considered development data, basically, how the shape of bird face forms at different stages of development. 

So to collect that information, we looked at different birds through that developmental pipeline, very early development, as that beak is forming, and collected gene expression data—what genes are on, what genes are off, how on are they, or how off are they—and fed all of this information collectively into this AI tool that we built called GenoMorph. This tool can tell us, using machine learning algorithms, what genes are turned on and off, and when, and it outputs a big list. And that is our candidate list of what might be helpful for bringing back the dodo.  

It sounds like you’re looking at known birds—where we can understand their genome and how that genome expresses in real time—to understand and zoom in on what dodo DNA might be most important for those quintessential dodo traits.

BS: That’s right. It’s all about this regulatory landscape, which is confusing. . . . Every cell has the same book in it, but which chapters should I be reading if I want to make a dodo beak versus a Nicobar pigeon beak?

When people think about AI right now, they’re thinking of LLMs [large language models] that are covering a wide base of knowledge. This sounds like one of the more traditional definitions of AI, where you have a very specific training set that is very refined and for a very specific domain, and then you can kind of build out of that. . . . Did you use a frontier model at all to make this?

Ben Lamm: In the early days of LLMs, they weren’t doing a great job at ancestral state reconstructions or comparative genomics, so we had to go solve those. They’re highly nuanced, right? We do have good relationships with the big AI companies. Our unique data set—plus some of the kind of traffic control stops cop stuff that we’re building in this—will further refine and be able to be leveraged by the superintelligence platforms. 

But we’re not currently doing that. We do have access to some of the latest and greatest on OpenAI and Anthropic with Mythos, and we are using it for certain things, but not yet this. Not just the data sets, but also how we derived the results of the thinned-down data set will be valuable and [could be] fed into a superintelligence platform.

BS: Some of the frontier models that are supercool that are out there are not really trained to solve this particular problem, which is how do we map a particular genotype to a phenotype . . . [basically] mapping a particular bit of developmental biology onto the way an organism ends up looking.

If you could map out how this would be applied to mammal species, would there be important differences? How would that work?

BL: The fundamental tech will be applied to both. We’ve really focused it on the avian side. . . . One thing that I was unaware of till this morning was how fast we think the core platform can be leveraged on some of the other work.

When it comes to your claims that you’re bringing back the dodo, people are very attached to formal delineations of species. They say you’re not actually bringing back the dodo. 

BL: I used to own a conversational intelligence platform before. . . . We’re very thoughtful on monitoring sentiment analysis, social graph, and earned media. There is a population that is very fixated on what to call things. They would have crushed it in the Garden of Eden, right?

But that is not representative[. . . .] We believe in freedom of speech. If somebody wants to say, you know, our dodos are dodos, and our mammoths are mammoths, that’s great. If they want to say they’re cold-tolerant Asian elephants with edited mammoth alleles from 104 mammoth genomes from 1.3 million years of genetic divergence, that’s cool too. 

I do think if you say the second one, you are kind of missing the forest for the trees, right? I was recently told that one of these people said online that if it looks like a duck, if it quacks like a duck, if it walks like a duck, it serves the ecological function of a duck, it eats like a duck, that doesn’t mean it’s a duck. Tell my two-and-a-half-year-old that.

BS: We’re focused on filling ecological niches, and if we can bring back an organism that fills an ecological niche that performs the role, then we will have successfully brought that animal back.

We will make decisions that are based on animal welfare and the capacity to live in the habitats of today. Our dire wolves are white because ancient dire wolves were white, but we didn’t use the dire wolf DNA version to edit to make them white because, in a gray wolf genetic background, there’s some evidence that that would have made them unhealthy. They would have made them deaf or blind. So we used a version that circulates today in gray wolves, but we still resurrected that dire wolf trait, and that’s what we’re going to do. We’re going to make decisions that focus on animal welfare and restoration of key traits, and that’s our definition of “de-extinction.”

Are you worried at all that you might be missing some trait, that there’s some other part of the body of the dodo that you’re not looking at, which might be quintessential to the dodo? 

BS: I think our goal, which we define in collaboration with the Mauritius Dodo Advisory Committee, is to create a large ground-dwelling, fruit-eating bird that can restore the dodo’s ecological niche, and if we are successful in doing that, we will have brought back the dodo. 

BL: If I just had to make a bird that looked like a dodo, my life would be easier, and I could raise way less money. My point is, if I make a unicorn tomorrow, I think fewer people will bitch because they’ll be, like, ah, “we can’t say they didn’t use unicorn DNA.” 

BS: It’s a species definition that’s based on an arbitrary threshold of sequence similarity. But that threshold is arbitrary. It’s going to be arbitrary no matter who you are. 

BL: So we have 104 mammoth genomes, right, over 1.3 million years. None of them [produced] living cells. We can’t clone them. You know what’s crazy about 1.3 million years of genetic divergence? They’re very different. You know what’s not different? The parts.

If . . . I had all those living cells, I could just clone them and magically have 104 mammoths. Which ones are a mammoth? I will tell you, genetically, they are insanely different over 1.3 million years.

Renee Wegrzyn spent her career building the future of American science—including the Advanced Research Projects Agency for Health, or ARPA-H, the White House agency created to accelerate health breakthroughs.

Then the Trump administration showed her the door. Wegrzyn cofounded Transfyr to tackle what she calls one of the most overlooked issues in science: the magic hands problem. She explains why fixing this unglamorous bottleneck is the prerequisite for everything AI promises to deliver in medicine, and why she can feel the next Sputnik moment coming—even if she can’t quite name it yet.

This is an abridged transcript of an interview from Rapid Response, hosted by Robert Safian, former editor-in-chief of Fast Company. From the team behind the Masters of Scale podcast, Rapid Response features candid conversations with today’s top business leaders navigating real-time challenges. Subscribe to Rapid Response wherever you get your podcasts to ensure you never miss an episode.

You were at ARPA-H, the inaugural director of a White House effort to accelerate health research. You were amped up about the possibilities, but with the Trump administration, you had to pivot. Was it hard not to take it personally that they wanted different leadership?

I’ll tell you, I’m still amped up about it. It’s still an amazing idea. Obviously, DARPA [the Defense Advanced Research Projects Agency] has been around for almost 80 years. I hope ARPA-H will have the same longevity and continue to live on today.

I’m a builder. I love to take things forward and really show what’s possible. So was it hard when I left? Of course it was hard. I had a really unusual appointment: I was a presidential appointee, but I was term-limited, so I served through the transition and had a few weeks in the new administration. I served in “Trump one,” by the way, as well.

So it was a great opportunity to make sure they really understood why we built the agency. We needed an ARPA-H because there was something missing in the way we were doing science broadly, which was: How do you move really quickly? How do you get focused on a problem versus building a general knowledge set? I was able to show them what we were building and convince them that it was worth keeping around. I couldn’t be more proud of the agency and its perseverance.

So your new business, Transfyr—The New York Times wrote about you, and the article kept referring to “magic hands.” I’m curious what magic hands are. What problem is Transfyr trying to solve or address?

You probably have your favorite dish, or your parents’ or grandparents’ favorite dish. Sometimes they just have the magic hands. Nobody can cook it like they do. Nobody can make it taste the special way they can. And that happens in science as well. There are some scientists who just have the knack and can deliver really great science over and over.

The challenge is that a lot of those scientists are such experts they can hardly describe what they’re doing. It’s an art in some ways. And so the magic hands idea was really a testament to what those scientists can do in the lab.

The pipeline we’re creating at Transfyr is: Can we create computer vision and perception tools to capture the work of science as it’s being done so that we understand, if the protocol says, “Mix the cells,” whether someone is inverting the tube in a certain gentle way, while another researcher may be rigorously shaking or vortexing it, which could destroy the sample?

Those little things, compounded over the many decisions a scientist may make over the course of several hours of experimentation, lead to what is called magic hands.

Was this an issue you saw at ARPA-H, the sort of breakthroughs that were trapped in a single lab in some ways?

Where that manifested at DARPA and at ARPA-H was in a realization, through my career, of how fragile breakthroughs are. Usually, the way a breakthrough is first communicated to the world is through a research publication, a Science or Nature paper. That paper is not a product. It’s not commercializable. It’s not ready for a patient that day. That’s the starting gun for what is going to happen.

The ARPAs of this world are really fantastic at getting to that initial breakthrough and demonstrating what is possible, but following through and delivering is not a 100% guaranteed activity.

When I started ARPA-H, we wanted customers, who would be the final users of that technology, to give us feedback on what we were developing so that when we got to the last day of the experiment, they were demanding that product, versus having that be the first contact with a customer, which unfortunately is pretty common around traditional grant-making organizations.

It’s rethinking how you make sure you can transfer that knowledge off the back end of a program.

Because the folks at the lab aren’t necessarily worrying about, or prioritizing, how to make it scalable. They’re just trying to see if it works. But of course, eventually we want it to be scalable.

Sometimes the folks in the lab don’t even have the expertise or understanding of who they need to talk to next to do that, right? We think about science and those transfers as handoffs. We deploy our technology to an originating lab to understand the science and what was important about that discovery, but we can also deploy it in a receiving lab so that we can see how they’re reproducing it.

And instead of the way we transfer science today, which is looking over each other’s shoulders and saying, “Oh, you’re doing it a little wrong,” we have this video-connectedness that lets us help accelerate those iterations of transferring the science.

Do you ever get frustrated about how resources are allocated in our world? You’ve put your life toward this work, and yet so much money is devoted to, I don’t know, space exploration or the next hot social app, the Meta Muse craze. How do you deal with government and how it prioritizes things?

Yeah, it’s maddening. One of the things that has me nervous right now is that I look back on the last 80 years following the Second World War. We’ve led the globe in science and engineering and innovation. Nobody who’s practicing science today has practiced science in the United States at a time when we weren’t, quote-unquote, winning from a competitiveness perspective, when we weren’t at the top.

I genuinely believe we may lose that competitiveness, that edge, that incumbency, and I’m not sure I see us fighting for it in the way that we should be. We have to deliver on what we’ve promised to do, and that’s a commitment not just to my country, but to my kids and my family. 

What I think a lot about is Sputnik moments. We’re going to look back at this moment in time like there’s something happening, and it’s going to impact whatever’s happening 20 to 50 years from now. You can feel it in the moment.

Where that happened before, of course, was Sputnik. The United States didn’t know the Soviet Union was developing that capability, and we didn’t have a counter-capability. What I realized is that each individual person has a Sputnik moment in their life that is usually health-related. My Sputnik moment was when my mother was diagnosed with cancer. That stopped everything.

Everybody has a story, whether it happened to them or their family, or if you talk to a member of Congress, something’s happening in their state that they can’t deal with.

So I’m like, “What is the Sputnik moment now?” If ARPA-H was bringing people from an unhealthy state to a healthy state, what’s happening now is taking us beyond that. If we can figure out and solve health and wellness, a lot of people talk about abundance. What is the next trajectory? What is the Sputnik that we’re working toward?