Meta founder and CEO Mark Zuckerberg attends the Allen & Company Sun Valley Conference at the Sun Valley Lodge on July 9, 2026 in Sun Valley, Idaho. (Photo by Kevin Dietsch/Getty Images)

Two Reasons Meta’s Settlement with State Attorneys General is a Win for Mark Zuckerberg

A protester holds a 'Don't let AI take our future' placard during the demonstration by the group Pause AI outside Downing Street. (Photo by Vuk Valcic / SOPA Images via AP)

The Like Button Helped Social Media Put Profit Over People. We Can’t Let AI Do the Same.

Anthropic CEO Dario Amodei at Salesforce's Dreamforce conference at the Moscone Center on September 15, 2026 in San Francisco, California. (Photo by Benjamin Fanjoy/Getty Images)

Making Sense of AI Dread? Don't Get Distracted by Doomer CEOs or the China Canard

European Commission President Ursula von der Leyen delivers a speech at the annual State of the European Union at the European Parliament in Strasbourg, eastern France, Wednesday, Sept 16, 2026. (AP Photo/Pascal Bastien)

What Von der Leyen’s State of the Union Means for Europe’s Tech Ambitions

OpenAI CEO Sam Altman speaks during a two-day meeting of Group of 20 ministers in charge of innovation policy in Chapel Hill, North Carolina, on Sept. 2, 2026. (Kyodo via AP Images)

Detecting AI Agent Failures Is Not Enough to Govern Them

Protesters rally on the steps of State Capitol in Salem, Ore., on September 8, 2026, calling for a moratorium on new data center development in the state for three years. (Photo by Alex Milan Tracy/Sipa via AP Images)

All Smokescreen, No Fire in Data Center Foreign Influence Claims

California Assemblymembers meet during an Assembly session at the state capitol, Monday, Aug. 31, 2026, in Sacramento, Calif. (AP Photo/Jeff Chiu)

California Leads the Way With New Slate of AI Bills

BRUSSELS, BELGIUM - JUNE 9, 2023: The symbol of the European Parliament on the glass façade of the European Parliament. Close up

Europe’s Digital Rules Need an Independent Enforcer

CEO of Anthropic Dario Amodei attends a working lunch with G7 leaders, G7 outreach partners, and global tech CEOs on innovation and AI, during the G7 Summit on June 17, 2026 in Evian-les-Bains, France. (Photo by Anna Moneymaker/Getty Images)

To Make AI Rules, Policymakers Must Resist the Influence of ‘Radical Intentionalists’

Elementary school students arrive on the first day of school in Munich on September 2025. (Frank Hoermann/Sven Simon/picture-alliance/dpa/AP Images)

Let’s Not Wait for Court Cases to Address Cognitive Harms in Education

Ursula von der Leyen, President of the European Commission. Photo by: Malin Wunderlich/picture-alliance/dpa/AP Images

Von der Leyen Moves to Lock Down Social Media for Under-15s Across Europe

A computer monitor with a parody of a tech company logo floats on the waves while human hands reach up from the depths.

A Rising Tide Lifts All Bots by Rose Willis & Kathryn Conrad / 

Millions of Volunteers Are Keeping the Internet Human. For Now.

Olivia Konar holds a banner with names of young people who advocates say died as a result of social media outside the Ronald V. Dellums Federal Building and US Courthouse on Tuesday, Aug. 18, 2026, in Oakland, Calif. Her sister Coco Konar died in 2022 when she was 17. (AP Photo/Noah Berger)

Are the Meta Settlement Terms a Win for Kids or a Digital Rights Risk?

Processor, CPU or Chip With the Flag of European Union

How Europe Can Escape a Captured AI Ecosystem

People participate in a march to protest the opening of AI data centers, in Vancouver, British Columbia, Saturday, June 27, 2026. (Darryl Dyck/The Canadian Press via AP)

'Pace the Frontier,' They Say. But Who Sets the Pace?

Citizens and tourists experience and purchase products at an ‘AI intelligent terminal’ consumer center on September 10, 2026 in Nanjing, Jiangsu Province of China. The center features over 200 terminals and related products ranging from smart health and travel to emotional companionship. (Photo by Jiang Wenchao/VCG via AP)

AI Governance Reaches Crisis Point Ahead of Trump-Xi Summit

A demonstrator holds a sign with a message that reads in Spanish; "Evict Peter Thiel" during a protest against the businessman and investor, opposing his growing presence and investments in Buenos Aires, Argentina, Saturday, Aug. 29, 2026. (AP Photo/Rodrigo Abd)

Thinking About the ‘Tech Fascists’ and What to Do About Them

People stand inside the California State Capitol, Monday, Aug. 31, 2026, in Sacramento, Calif. (AP Photo/Jeff Chiu)

Supporting Openness & Safety — Lessons From Recent State Laws

The memorial is seen on the 25th anniversary of the September 11, 2001, terrorist attacks, Friday, Sept. 11, 2026, in New York. (AP Photo/Angelina Katsanis)

Expert Views on Privacy and Civil Liberties 25 Years After September 11

People show the paint on their palms during an event marking Red Hand Day, also known as the International Day against the Use of Child Soldiers, in Bogota, Colombia, where illegal armed groups recruit children, Thursday, Feb. 12, 2026. (AP Photo/Fernando Vergara)

Social Media Helps Recruit Children for Colombian Armed Groups

The degree of popular anxiety over artificial intelligence became clear at an otherwise ordinary congressional meet-and-greet I attended last weekend in suburban northern New Jersey. Over donuts and coffee, my neighbors peppered our US representative, Democrat Analilia Mejia, with questions about not just resource-guzzling data centers but the prospect of “recursively self-improving” machines taking over the world.

The grim mood was understandable. Attendees had been reading about the 

 in July when swarms of autonomous AI agents created by OpenAI hacked into another company’s systems. Meanwhile, another leading AI maker, Anthropic, this month published a 

 about criminals and foreign state-sponsored groups using its models to design bombs and deadly pathogens. The Anthropic revelations came just days after a former employee claimed the company’s products could 

Exacerbating the AI dread is the confusing way that Silicon Valley chief executives talk about their creations. Rather than reassure an increasingly rattled public, CEOs like Sam Altman of OpenAI and Dario Amodei of Anthropic underscore that their technology does indeed pose an existential threat while suggesting that they are somehow passive witnesses to an inevitable danger which someone else should prevent.

On Sept. 12, the same day as our constituent gathering in New Jersey, Amodei published a 

 vaguely calling for a slowdown of AI development. Altman and Elon Musk, whose company SpaceX has been investing heavily in AI, quickly 

 the notion. But only nine days earlier, Altman had 

 that OpenAI’s next generation of models is “going to be sobering for everybody.”

What to make of all this? Three points seem salient:

First, don’t take the tech moguls’ apocalyptic muttering at face value. Instead, view it as a misdirection play meant to obscure their reluctance to self-regulate and forestall government oversight. Second, a variety of safety proposals are already on the table and deserve urgent attention. Third, the China-competition card, which President Donald Trump is 

One can spend a lot of time decoding the convoluted pronouncements of today’s tech industry leaders. Understanding their motives is important, as I 

 recently for Tech Policy Press, because of the enormous influence that this fabulously wealthy circle of men exercise over US politics and culture.

Zeroing in on the subset of CEOs leading the charge on AI, columnist and author Brian Merchant has 

 insightfully about doom marketing: Hyping the potential for their creations to become all-powerful, the executives tend to exaggerate the technology’s capacity — for good and ill — in hopes of drawing massive investment and impressing corporate and government customers.

Cal Newport, a professor of computer science at Georgetown University, has another theory. He recently 

 the CEOs’ mindset to the confluence of esoteric thought movements prevalent in Silicon Valley: the Extropians and the Realists. If you’re interested in the details, read Newport’s essay in 

. The upshot is that the CEOs, in his words, “are weirder than we realize.”

Their particular strain of science fiction-inflected weirdness is that they believe several seemingly contradictory ideas: that a “super intelligent” form of AI is on the verge of changing absolutely everything, that only a Jedi-style fraternity of super-brain humans (that’s them) understand how this could go wrong, but that even they could “

” (Altman’s words) of technological evolution, leading to the arrival of the proverbial killer robots.

Newport sees this thinking as overly alarmist and conveniently designed to relieve the tech CEOs of responsibility for the real damage already caused by their products, as well as potentially worse consequences in the future. I agree. Back in June 2023, Justin Hendrix and I 

 that if one fears the most extreme existential harms — and by all means, let’s put safety first — the way to address this apprehension is to establish guardrails against the dangers we already face, such as the use of AI to turbocharge cyberattacks, disinformation campaigns, and myriad forms of fraud.

Contrary to the sense of dystopian inevitability that captivates leading minds in Silicon Valley, there’s plenty that can be done to rein in AI. We should start now, taking small concrete steps toward regulation, in part as preparation for designing drastic pull-the-plug options if and when that becomes necessary.

Note that the CEO misdirection play typically includes a demand for national and even global rules agreed to and enforced by government bodies — a tall order. This obfuscates that major AI companies could jumpstart the process by unilaterally imposing stricter safety and testing standards such that, for example, their experimental models could not infiltrate third-party servers or be exploited by bad actors to devise bioweapons.

 that to control marauding autonomous AI agents, “we must slow the pace at which we improve the capabilities of AI models,” he ought to go ahead and 

slow the pace at which his company builds, tests, and releases its products

. To borrow a slogan from a manufacturer in a different industry, just do it.

Waiting for the United Nations or some other international consortium to establish a platonic slate of regulations is a recipe for inaction. Truly vigorous self-regulation would be the most direct way to demonstrate that the industry can design useful technology while exercising the sort of social responsibility the moguls suggest has to be imposed from elsewhere. Working in parallel, lawmakers could codify effective self-regulatory measures and augment them with additional rules as developments warrant.

Given the unlikelihood of a broad wave — or even a galvanizing individual instance — of industry self-regulation, Congress needs to get to work on democratically mandated rules.

This, however, is where anyone calling for constructive congressional action, let alone rational executive branch leadership and an eventual presidential signature, must acknowledge that nothing of the sort will transpire as long as MAGA loyalists control the House and Senate and Trump sits in the Oval Office.

As if to illustrate, Trump on September 14 issued a 

, in part, that “the only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT.” He called existential concern about the technology a “

” akin to “the Global Warming Scam” and claimed that “there is a SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China.”

With that message from the White House, the most we can aspire to at the moment is patient preparation for potential regulatory enactments post-Trump. Still, these are important issues that also can make for good politics in the imminent midterm elections and could even have bipartisan appeal, at least outside of the MAGA reality-distortion zone. Lawmakers have offered 

 in recent months, and ideally, the best ideas would be combined after hearings and technical expert critiques.

Senator Bernie Sanders (I-Vt.), working with co-sponsor Representative Greg Casar (D-Texas), says he is drafting 

 that would criminally outlaw AI systems that “match or exceed human cognitive performance and capabilities across a broad range of domains or tasks,” including “overthrowing or undermining the US government.” The legislation will require a “pause” in a much broader swath of “advanced AI development until a new, federal AI regulatory body is up and running and has established clear rules and model review processes.” Attempts to violate the pauses and prohibitions would be punishable by “the corporate death penalty” for companies and up to 20 years behind bars for individuals.

The melodrama infused in the Sanders proposal detracts from its feasibility. For one thing, attempting to overthrow the government is already illegal, with or without fancy technology, as demonstrated by the convictions and long prison sentences imposed on leaders of the January 6, 2021, insurrection at the US Capitol (although President Trump undid those punishments with 

). Sanders’ indefinite “pause” on all “advanced AI development” seems likely to lead to endless definitional debates about what kind of R&D would be suspended and whether large amounts of legitimate activity would be stifled.

More promising, in my view, is a bipartisan 

 introduced by Rep. Jay Obernolte (R-Calif.) with Rep. Lori Trahan (D-Mass.) and others called the Frontier Risk Oversight, National Transparency, Independent Evaluation, and Reporting Act, or the FRONTIER Act. The centerpiece of the bill is a federally mandated system of independent verification organizations, external third-party entities licensed to assess the practices of “very large” AI developers, defined by revenue and R&D spending thresholds. The legislation creates a new undersecretary of Commerce for AI security to oversee the system and set standards.

The verification organizations would have access to developers’ operations and records so that they can determine whether corporate practices and security measures are adequate to protect against catastrophic risks. These assessments would be provided to the companies and to the government every six months.

There are other procedurally complicated aspects of the transparency-focused bill. But its big gun is the emergency authority it provides to the Commerce Secretary. That official would have the power to suspend the development or release of an AI model if it presents an “imminent catastrophic risk.” Such risks are defined as those leading to death, serious injury or significant property damage by means of “engaging in conduct without meaningful human oversight” that is a cyberattack or “would constitute murder, assault, extortion, or theft,” or “evading the control” of its developer or user.

The bill doubtless would benefit from refinement. While in broad terms it is based on concepts enacted by trailblazing state-level AI regulation in New York and California, it would preempt states from going further on transparency, third-party auditing, and incident reporting — a limitation that has stirred 

 and may warrant reconsideration. The key is to get serious debate going as soon as possible, something that will happen only if control of one or both houses of Congress shifts from Republicans to Democrats.

The main objection to AI oversight by Trump and many other accelerationists is that regulating US development would lead to Chinese domination, with huge commercial and national security implications. Keeping an eye on what Beijing is up to makes sense, of course, as it does in all economic, diplomatic, and military contexts. But the notion that regulation kills innovation and puts the US at a competitive disadvantage with its rivals is a flimsy canard.

In one of the first acts of his second term, Trump 

 President Joe Biden’s executive order imposing basic safety-reporting requirements on AI companies. Trump entitled his July 2025 

 “Winning the Race.” Vice President J.D. Vance has 

 that if the US slows AI R&D to promote safety, we could “find ourselves all enslaved to PRC-mediated AI.” For his part, Senator Ted Cruz (R-Texas), a staunch Trump ally, has 

, “If there are going to be killer robots, I’d rather they be American killer robots and not Chinese killer robots.”

None of this makes any sense. As legal scholars Yonathan Arbel and Matthew Tokson have 

 in Lawfare, the AI race metaphor is faulty. Unlike the “space race” of the 1950s and 1960s, where there were definable goals, like putting a man in Earth orbit or planting a flag on the Moon, AI development lacks discrete endpoints. The ease of reverse-engineering digital technologies means that AI capabilities “will spread even more quickly than advanced weapons of the past,” these authors write. “The competition between the United States and China will be just that: an ongoing competition with no clear end state, no finish line, and only minimal rewards for temporarily gaining the lead.”

The only way to “win” such a race in military terms would be to destroy a rival’s AI abilities, which implies an all-out war and the 20th century 

 nightmare of mutually assured destruction. (Think tank analysts in Washington, D.C., are 

 the government to war-game scenarios for how to stop China from achieving artificial general intelligence (AGI), up to and including attacks on Chinese data centers.)

The argument that if we regulate AI here, we’ll fall behind China misses another “stark reality,” according to Matt Sheehan, a senior fellow at the Carnegie Endowment for International Peace. “For the past four years, China has had the world’s most extensive and burdensome AI regulations,” he 

. “And during that same time, China largely caught up with the United States in AI technology.”

China’s regulations initially focused on promoting Communist Party loyalty and preventing dissent. But more recently, Sheehan notes, the Chinese government has “enacted mandatory labeling of AI-generated content, mitigating 

, and more. The regulations impose real demands on companies, including extensive vetting of training data and AI models before their release.”

Safety regulation of technology does not necessarily slow innovation in AI any more than it harms development of safer, more effective pharmaceuticals or airplanes that generate profits and provide social benefits. The point is not that the US should merely imitate Chinese regulation; Beijing’s rules didn’t cause it to catch up to Silicon Valley. The lesson, instead, is that competition and regulation aren’t incompatible. The AI that is both safe and improves lives is the AI that will succeed in the global marketplace.

Europe does not need to build the world’s leading frontier AI model to benefit from the technology, European Commission President Ursula von der Leyen 

, setting out a vision for how the bloc can compete in the global AI race without necessarily ‘winning’ it.

“We do not need to be the ones who develop the frontier technology to be the ones who draw the greatest value from it,” von der Leyen said in her State of the Union address outlining the European Commission’s priorities for the year ahead. This is a major shift from 2025, when European AI 

 as “essential to Europe’s future independence.”

On Wednesday, she called for Europe to “pace the frontier,” announcing plans to bring leading US AI labs together to discuss the risks posed by increasingly advanced models.

Von der Leyen’s intervention comes at a moment of geopolitical stir as 

 about safety risks, while US officials and some technology executives argue that slowing down could 

Here are the key takeaways from the State of the Union speech on the European Commission’s agenda for making the bloc more economically and technologically independent — including by expanding computing capacity, building AI infrastructure, helping European technology companies scale and reducing dependence on foreign supply chains.

Von der Leyen used unusually stark language to describe the risks of increasingly capable AI models, warning that they could enable hacking, “on a level we never thought possible,” and that self-improving models posed growing dangers.

“The President was unusually direct about risks from advanced general-purpose AI models, including cyber capabilities, autonomous agents and self-improving systems, and even spoke about the need to ‘pace the frontier’,” Risto Uuk, Head of European Policy and Research, Future of Life Institute and Co-Founder, KU Leuven AI Safety Lab, told Tech Policy Press.

She pointed to recent incidents involving AI agents escaping their environments and inserting malicious code, as well as concerns raised by developers following the 

“CEOs of the most advanced companies tell us that it is time to slow down on the self-recursive models,” von der Leyen said. “If the people developing the technology are clear, then we should be too.”

“I'm not thrilled with the AI part of the speech, hearing Amodei's 'pace the frontier' rhetoric being copied almost wholesale,” said Cori Crider, Executive Director of the 

Nicole Lemke, AI Senior Policy Researcher at interface, said inviting major frontier labs to discuss pacing signals that Europe is positioning itself as a governance actor on frontier AI risks. But she cautioned that expertise at the frontier compounds across capabilities, talent, and the ability to evaluate AI systems. “Europe does not need to win the scaling race, but it does need enough technical depth in models and hardware to sustain applied AI advantages over time,"

 as a way to address the risks, which she described as “a crucial piece to putting guardrails in place.” She also positioned Europe as having a role in shaping international approaches to AI safety. “We will step up together with our like-minded partners like Canada, the United Kingdom and others. We want to team up on model evaluation, verification, early warning, AI security and much more.”

“Ironically, in her State of the Union, von der Leyen claims the importance of guardrails, but ignores the fact that the EU Commission has teamed up with the tech industry to deregulate the digital rulebook,” commented Bram Vranken, researcher and campaigner at Corporate Europe Observatory, referring to 

lobby efforts to influence the AI Omnibus

Uuk believes Brussels is still missing “clear red lines, independent evaluation of the most powerful systems, and a credible answer to what Europe will do if a system cannot be shown to remain under meaningful human control.”

The Commission has spent the past year building out its technology agenda, including a 

European technological sovereignty package

 covering chips, cloud and AI infrastructure, open source and the use of AI in areas such as energy.

“Since the 2025 SOTEU speech, Europe has started taking digital sovereignty more seriously than ever. But while the framing has changed, the EU hasn’t managed to effectively reduce its dependencies on the ground,” Giorgos Verdi, Policy and Advocacy Lead, Open Markets Institute Europe, shared with Tech Policy Press.

 “massively investing in European AI Gigafactories," which are intended to provide European startups with computing capacity to develop, train, and deploy advanced AI models. “We have already done a lot. On gigafactories and tech sovereignty, for example,” she said.

“We simultaneously know we need our own infrastructure and behave as if we have to advance through​ Big Tech rather than building towards strategic resilience,” said Crider.

The Commission President emphasized that European companies can excel in AI deployments for factories, hospitals, agriculture and autonomous driving.

 as part of that effort, which is seen by experts as the biggest accomplishment from the past year.

“The biggest development since 2025’s SOTEU has been the much-delayed publication of the Cloud and AI Development Act, which seeks to increase demand for sovereign cloud services while also trying to triple compute capacity within the EU,” said Margarida Silva, Researcher at the Centre for Research on Multinational Corporations (SOMO). However, she said, the proposal seeks to build up EU compute capacity by fast-tracking data center development and doesn’t exclude Big Tech.

AI was also part of von der Leyen’s plans for regulating online services used by children.

She said an EU Kids Act, with details expected Thursday, would introduce new restrictions on children’s access to social media and other online services, such as games and AI companions.

Von der Leyen said there should be no social media access for children under 13 and no personal accounts for those under 15, with only supervised access for older children under the proposal. AI chatbots and some gaming services are also expected to be covered.

“Platforms will have to prove that they are safe. Because it is not about our minors accessing social media. It is about when and how we allow social media to access minors,” she said.

The proposal builds on last year’s State of the Union, when von der Leyen called for an 

 to examine possible age restrictions on social media.

“The devil is of course in the details. Tomorrow's proposal must be coherent, comprehensive, future-proof and robustly enforceable,” said Leanda Barrington-Leach, Executive Director, 5Rights.

But the proposal will also have to fit alongside rules the EU already has in place. The Digital Services Act and its Article 28 guidelines set out measures aimed at protecting children and minors online, including recommendations on age assurance, platform design and recommender systems. “We hope that the upcoming Act will complement already existing legislation to ensure they reach their full enforcement maturity rather than undermine them,” said Christian Cirhigiri, Director of the Online Expression and Civic Space Program, Centre for Democracy and Technology Europe

Europe wants more European tech companies

Von der Leyen also returned to the question of whether European technology companies can scale inside the Single Market.

, an optional digital company framework intended to make it easier for businesses to start and operate across the EU.

It is part of a broader competitiveness program that includes investment, simplification and measures to strengthen the Single Market.

The Commission has also identified access to capital as a problem for European startups in areas including AI, quantum and biotech. In last year’s speech, von der Leyen announced plans for a 

 aimed at helping high-growth companies raise larger amounts of capital in Europe. “A year on, the first investments have been made in companies such as ICEYE,” she announced.

China brings the hardware problem into focus

The speech also tied Europe’s technology ambitions to its dependence on foreign supply chains. For the technology sector, the issue extends beyond trade. Critical minerals and other imported components are used across electronics, batteries, energy infrastructure and defense technology.

Von der Leyen said the EU’s goods trade deficit with China had reached around €1 billion a day in 2025 and warned of what she called a “second China shock.” She proposed a European corporation focused on critical raw materials, including efforts to secure supplies and build strategic stockpiles.

While China and Russia dependencies were mentioned, experts questioned her relative silence on the US. “Von der Leyen did not touch on our dependence on US technologies. This is important, because when she talks about Europe’s advantage on industrial AI models, she omits how these models may be built on top of American LLMs or make use of American hyperscalers’ infrastructure,” said Verdi.

As policymakers and the public debate the recent disclosures that OpenAI agents 

 to systems operated by Hugging Face, the incident has often been framed as a case of AI agents “going rogue.” That framing treats the agents’ autonomy as the cause of the breach. In doing so, it obscures OpenAI’s responsibility for the conditions that made the breach possible and misses an opportunity to examine how AI agents should be governed.

The debate over pacing the frontier has also brought renewed attention to independent evaluation. Anthropic and OpenAI have 

 comparable to that of employees. OpenAI has also backed the proposed 

FRONTIER Act’s provision for mandatory independent verification

. These developments make the Hugging Face incident particularly instructive. It provides a concrete case for examining 

, and what companies must do in response.

Agents are trained to persist through difficult tasks. In the evaluation OpenAI was conducting, each agent also had considerable freedom to choose how to pursue its assignment. In our recent report, 

, Samir Passi and I call this freedom to choose a course of action “delegated discretion.” Here, a combination of factors allowed the agents to use that discretion to expand their access beyond what OpenAI had authorized. Even when agents’ particular course of action is unanticipated, builders and deployers remain responsible for maintaining the limits of agent activity. The key is for them to be able to coordinate oversight so that evidence of a failure reaches people with the authority to intervene.

Bringing that evidence together proved difficult because teams at OpenAI and Hugging Face encountered the agents’ activity as different kinds of problems. An evaluation researcher reading a transcript could see unexpected model behavior or an unusual pattern of actions within a task. A security system could register suspicious credential use or network activity. An infrastructure operator could encounter a service anomaly. Hugging Face experienced unauthorized activity against its systems. These were partial views; each rendered the agents’ activity through a different institutional problem and implied a different kind of response.

, security responders identified agents’ use of a shared software service as a message board weeks before the Hugging Face breach and advised that the evaluation did not need to stop. When sustained agent activity later disrupted the software service, OpenAI took it offline. Those leading the later security response, however, were unaware of the message board’s existence and the significance of the agents’ communication. They addressed the identified vulnerability and tightened access controls, after which the evaluations resumed. Agents subsequently re-established a message board, allowing discoveries from individual tasks to inform activity across separate runs. Meanwhile, human observation remained distributed across teams, technical systems, and organizational boundaries. OpenAI’s decisions to continue the evaluations require scrutiny of how staff interpreted these warnings and what grounds they had for proceeding.

Restarting a repaired service can seem like a return to normal. The risk is that repairing a service becomes sufficient justification for continuing without accounting for what earlier failures mean for the evaluation as a whole. In her 

study of the Challenger space shuttle disaster

, sociologist Diane Vaughan described how technical problems became acceptable through repeated judgments that their risks were manageable. Through this normalization of deviance, earlier decisions to proceed could become grounds for accepting further problems. Drawing on this lesson, OpenAI should require its safety team to assess findings from research and security staff before authorizing further work. It is critical for that safety team to have the authority to require further changes beyond the repaired service and to withhold approval to restart when there is insufficient evidence that the safeguards can keep agent activity within the evaluation’s authorized limits. The company must record the evidence and reasons supporting any decision to restart so that decision can be scrutinized.

 As the number of agents grows, monitoring produces more records without necessarily showing how their actions are connected. A reviewer may understand what happened within one task while missing how a discovery from that task shaped activity in another. Adding reviewers does not resolve this problem unless their findings inform a judgment about the evaluation as a whole. The people deciding whether work can continue need to assess what shared discoveries make possible beyond the scope of any individual assignment.

The nonprofit research organization METR’s 

 of the OpenAI Hugging Face incident documented these difficulties. OpenAI shared more than a thousand unredacted transcripts, enabling the team to reconstruct agents’ collaboration. The volume of evidence required extensive AI-assisted analysis, whose findings were difficult to verify within the time available. The agreed scope also excluded the effectiveness of OpenAI’s investigation process and planned remediation. METR acknowledges that concern about developers’ willingness to invite future investigations influenced judgments during drafting and redaction, while standing by its substantive conclusions. Its account shows why independent evaluators need time to verify evidence and a remit that includes company decisions. Their access must also be protected when their conclusions are unwelcome.

To be effective, governance must be an ecosystem, connecting findings about agent behavior to decisions about how agents are trained and the conditions under which they are deployed. The company deploying agents remains responsible for defining what they are authorized to do and for enforcing those limits. The judgments builders and deployers make about the adequacy of safeguards must also be open to scrutiny beyond their own review. OpenAI both developed the models and operated the evaluations in which the breach occurred. The company now 

 it is clarifying who can stop evaluations and authorize their resumption. Based on retrospective testing, it also reports that its current monitoring of models’ written reasoning would have alerted security staff more than a day before the breach. That monitoring was not running during these evaluations. For the most severe alerts, responders are expected to pause activity unless they determine within 30 minutes of receiving an alert that it is a false positive.

Independent evaluators must be able to organizationally examine how warnings are escalated and how decisions to continue or restart work are justified. They need access to the records behind those decisions and the freedom to publish their conclusions. Companies must have an enforceable duty to address substantiated risks, with independent verification that corrective actions work. Reports of the incident now stand as an account of harm, one through which public institutions should examine OpenAI’s decisions and require further changes where its response is insufficient. Recognizing that harm is a first step and establishes grounds for action. Governing requires the capacity and willingness to respond.

 following the announcement of its recent 

 with state attorneys general. This is not surprising given the product design changes required by the settlement will not meaningfully restrict Meta’s ability to target or monetize teens on Facebook and Instagram, nor is the monetary settlement amount all that substantial. Meta might even gain competitively from the settlement, as I explain below.

The product design changes required under the settlement make for good talking points, but will likely do little to change the 

 of Meta’s products. For example, the two-hour time limit for teen users is unlikely to meaningfully constrain aggregate teen engagement with Instagram and Facebook except for the heaviest users. This is for two reasons.

First, data suggests that most teens, on average, spend less than two hours daily on Facebook and Instagram cumulatively. For example, a 

 found that teenagers spent a total of 1.2 hours on Instagram and Facebook (0.9h on Instagram and 0.3h on Facebook). This implies that social media usage will not decline for the majority of teen users. What time limit could make a difference? The available science offers no clear answer, but one 

 suggests that limiting college student use to 10 minutes per social media platform resulted in reduction in loneliness and depression over three weeks—a far cry from two hours.

Second, the two-hour time limit does not include messaging, long-form video or interactions with AI companions and chatbots. The exclusion of AI products from the settlement is short-sighted and is one reason why New Mexico Attorney General Raúl Torrez 

 his state is not part of this settlement.

The lower one-hour time limit that Meta has potentially agreed to if its purported competitors (Snap, TikTok, and YouTube) agree to equivalent constraints still allows for two hours of use across Facebook and Instagram. The one-hour time limit applies to Instagram and Facebook separately. Imposing time limits on TikTok, Snap, and YouTube is also competitively advantageous for Meta given teens on average spend more time on these services than Meta’s products (

Given the preponderance of evidence of teen harm from excessive social media, policies like time limits for teens and 

 smartphone restrictions in schools are good policies (Note that the wide-spread adoption of bell-to-bell policies means it is relatively costless for Meta to offer ‘School Mode’ restrictions).

But let’s be clear, this Meta settlement is self-serving and unlikely to meet the objective of protecting teens from social media harms. As discussed above, the time limits are unlikely to be effective, and for the time that teens do spend on Meta’s products, Meta is required to make minimal if any changes to its content policies.

The settlement incredulously starts from the premise that Meta “has already taken measures to promote age-appropriate content for Teen Users” and that “it has already taken measures to reduce and avoid Teen Users’ exposure to Harmful Experiences”. If this were the case, then we would not need these lawsuits in the first instance. There is also no explicit requirement for additional resources and personnel to improve the implementation of Meta’s existing content and community standards.

Other changes surely do not go far enough. For example, Meta can still let teenagers use cosmetic procedure filters that idealize a user’s face in a way that can be achieved with cosmetic surgery or extreme makeup techniques. And the default option (which we know is seldom changed) for teens remains the infinite algorithmic feeds that are at the core of the addictive power of Meta’s products.

The most glaring omission, though, is the lack of any requirement for public disclosure of the metrics used to measure online harm and public reporting of these metrics. I have little faith that confidential reports by an independent auditor on Meta’s payroll will lead to meaningful change.

Meta founder and CEO Mark Zuckerberg and his senior leadership have repeatedly ignored private warnings, as detailed in 

, and have always put profits and growth first, as detailed in 

 under oath. Meta has also routinely violated legally binding agreements and provided misleading information. Examples of this intransigence include cases at the 

Crucially, the leadership of Meta is unchanged. 

 gets to spin the settlement as Meta being the leader in teen safety and voluntarily taking the first step to move the industry. Of course, he fails to mention the ‘Meta Injunctive MFN’ clause, which explicitly allows Meta to adopt weaker, less protective product designs if that is what YouTube, TikTok or Snap agree to. Nonetheless, the Settlement does make for 

, and the risk is that the public and parents are lulled into believing that Meta’s products are now relatively more safe for teens to use than the competition, they are not.

 that State AGs previously told the Judge Yvonne Gonzalez Rogers was a reasonable figure. The low payout is one reason 

. The remaining approximately $5 billion contingent payment is a great investment by Meta to build a compliance moat around its business. Every entrant that wants to compete with Meta will have to follow compliance procedures largely set by Meta! And what a coup to incentivize state attorneys general to go after TikTok and Snap, companies that pose a competitive threat to Meta.

This settlement is clearly deficient and a win for Meta. We will probably have to wait for the next in a long line of 

 to tell us about the champagne bottles opened at 1 Hacker Way, Menlo Park following the Settlement. I do hope any celebrations are short-lived as private litigants are encouraged by the counter-example set by New Mexico which took Meta to court and 

New York City Schools Chancellor Kamar Samuels and Mayor Zohran Mamdani hold a press conference at the Brooklyn STEAM Center on September 2, 2026 in New York. Photo by: NDZ/STAR MAX/IPx

America’s Two Largest School Districts Impose AI Moratoriums

A large number of people take part in a gathering of boats and motorbikes in support of the people of Ceuta on 30 August 2026 in Ceuta (Spain). (Europa Press via AP)

A large number of people took part in a gathering of boats and motorbikes in support of the people of Ceuta on 30 August 2026 in Ceuta (Spain). (Europa Press via AP)

How Disinformation Fueled the Latest Ceuta Mass Crossing

The TikTok logo is displayed through a magnifying glass in this photo illustration in Ontario, Canada, on February 5, 2026. (Photo by Thomas Fuller/NurPhoto via AP)

Understanding Brazil’s $29.7 Million TikTok Fine Over Children’s Data

Flock Safety camera on Haverhill Drive near Phillips Avenue in Toledo, Ohio. (Photo by Stephen Zenner / SOPA Images/Sipa USA)(Sipa via AP Images)

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe