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

Water flowing over Srisailam Dam in Andhra Pradesh, India in Aug. 2019. Photo by Sharath G. from Pexels (CC-licensed)

India’s Data Center Boom Should Learn from its Dams

A person takes a picture of the newly unveiled sign for Lake Ontario at Fifty Point Conservation Area in Grimsby, Ontario, on Saturday, Aug. 29, 2026. (Keito Newman/The Canadian Press via AP)

Canadian Hydronyms, Builders, and Other Adventures in Middle Power Tech Policy

The Muse logo is displayed on a smartphone screen placed on a reflective surface onto which the Meta AI logo is projected, in Creteil, France, on September 9, 2026, after Meta launched Muse, an autonomous personal artificial intelligence agent. (Photo by Samuel Boivin/NurPhoto via AP)

The Internet Was Built For Human Agency. AI Agents Are Changing the Rules.

A mouse pointer lingers on button in the shape of a cookie. Photo by: Lino Mirgeler/picture-alliance/dpa/AP Images

Photo by: Lino Mirgeler/picture-alliance/dpa/AP Images

Europe’s Cookie Law Is Really a Law About Surveillance

Image of map and computer language over binary codes on server racks in server room.

Global Digital Policy Roundup: August 2026

, the California legislature has once again passed a slate of AI-related bills. Governor Gavin Newsom has already signed several bills relating to chatbots and independent audits, and the rest await his signature or veto. These remaining bills apply variously to employers, advertisers, large online platforms, healthcare providers, AI developers, chatbot providers, and attorneys, among others. Even if not all of them become law, their sheer number, along with some strong and unique provisions, keep the state at the forefront of AI regulation in the United States.

In this article, we summarize thematically the key AI bills passed in this California legislative session, indicating which ones the Governor has signed into law.

 (ADS) when making disciplinary or termination decisions. No other US state has such a ban. Like some state laws and rules, however, the bill would impose obligations on employers when they rely 

 on an ADS to make such a decision. Employers would have to have a human “corroborate” the ADS output. If so corroborated, an employer must provide a “postuse notice” when informing the employee of the decision and, if requested, describe the employee data used to make the decision.

The bill would also ban employer retaliation against employees exercising rights under its provisions. Further, it has a unique, broadly worded provision forbidding employer use of an ADS to “[p]redict and take adverse action against a worker for exercising their legal rights,” including but not limited to rights under employment and labor laws.

Governor Newsom vetoed a different version of this bill last year.

The legislature also passed two bills involving the use of 

 would forbid such a tool when it “uses artificial intelligence on employees” either to (1) recognize, or to make inferences or predictions about, an individual’s emotional state, or (2) collect neural data, unless the tool is used “to ensure safety.” The other bill, 

, would restrict use of workplace surveillance tools (even if they don’t use AI) in workplace bathrooms. Other states have workplace surveillance laws, but California’s specific references to emotion detection and neural data collection are distinguishing features.

The legislature passed three bills focusing on AI use in 

 applies to providers that use “a tool, system, or device that includes artificial intelligence” in connection with clinical functions that state law requires a professional licensee to perform. Such providers could not use those items either to “[i]ndependently perform” such functions or to advise unlicensed personnel to perform them.

The bill would also require licensed providers to retain “the ability to exercise independent professional judgment in patient care” when considering the output of a clinical decision support system (CDSS), which is defined essentially as an AI system that aids clinical decisions about diagnosis and treatment.

The other two clinically focused bills also focus on CDSS. 

 would require developers to make reasonable efforts to identify and mitigate reasonably foreseeable risks that a CDSS has biased impacts. 

 would (1) protect worker choice on whether to override CDSS output and (2) prevent developers and deployers from pointing to such choices as a defense in lawsuits alleging harm from CDSS output.

The final healthcare-related bill is the Wellness for Oversight and Psychological Resources Act, 

, which bears the same title as a similar law passed last year in Illinois. It would forbid offering 

 “through the use of companion chatbots,” including via claims that the chatbot “is a therapist or provides therapy.” The bill would also require review and approval by licensed professionals when using AI for certain purposes, such as making therapeutic decisions, detecting mental states, or performing triage or screening.

The legislature passed two bills dealing with the use of synthetic or digital characters. 

 requires a clear and conspicuous disclosure in an advertisement that “prominently includes” a “

,” defined as a digital representation of a human, created at least in part with generative AI, and “not recognizable as any identifiable natural person.” If signed, California would be the second state to address the advertising use of synthetic performers, following a 

 passed last year. The two regulations differ from each other in several ways, such as the California bill applying to audio-only advertisements, which are excluded from the New York law’s coverage.

Whereas the synthetic performer bill addresses characters that are 

 recognizable as a particular person, the second bill, 

, focuses on the opposite. It would amend two state laws that restrict unauthorized impersonations of real people in advertising and other specified contexts, ensuring that “

” are included in the scope of those laws. The amendments would extend other consumer protections that California already has in place for unauthorized use of digital replicas. Several other states have similar laws.

On September 10, Governor Newsom signed two of three bills that impose obligations on operators of certain types of chatbots, adding to existing California law on chatbots used for 

. A third bill on customer service chatbots remains on his desk.

The two new laws amend SB 243, the state’s 

 law. The most substantial revisions are in 

, called “Adam’s Law,” which is the 

 in the US requiring operators of such products to (1) perform and document a “comprehensive risk assessment” with respect to child users, and (2) submit to “independent child safety audits” of their compliance with the law. It also adds age verification and robust provisions – some of which resemble those in other states’ laws – requiring operators to mitigate child safety risks. These provisions include unique requirements such as default limits on how much time a child can chat with a covered chatbot, as well as limits on advertising and selling a child’s personal information. The law allows for both public and private enforcement and becomes effective on July 1, 2027, though the audit requirements do not kick in until January 1, 2029.

, which imposes a 5-year ban on the offer or sale of “any toy that includes a companion chatbot,” where “toys” are limited to physical products designed or marketed for play by children under 16. This law will go into effect on January 1, 2027, and the ban will remain in effect until January 1, 2031. No other state law has prohibitions specific to AI-powered physical toys.

, would also be a first-of-its-kind in state AI regulation, if signed. Most state chatbot laws do not extend to chatbots used merely for 

, whereas this bill centers on them. For example, California 

 that restricts intentionally misleading chatbot use to incentivize a purchase or sale. The new bill would prohibit representations that a customer service chatbot is a human and require disclosure if reasonable people would likely believe, even absent such representations, that the chatbot is a human. Coverage is limited to businesses that have over $500 million in gross annual revenue and provide goods and services to customers.

Amendments to the California AI Transparency Act

 took effect on August 2, 2026. Among other things, the law imposes disclosure obligations on covered providers, defined mainly to include generative AI developers that make their systems publicly available and have over 1 million monthly visitors. Such providers must ensure that content created using their tools includes provenance data that can be readily accessed through free AI detection tools. Specifically, they must provide a latent disclosure and an option for users to include a manifest disclosure for such content.

The legislature passed two bills to modify these requirements. 

 would delete the monthly visitor threshold and the requirement to provide the manifest disclosure option, thus expanding coverage while also reducing developer burden. 

 would modify the separate obligation of large online platforms to let users inspect provenance data, clarifying that they are not required to take actions “not compliant or interoperable with widely adopted specifications issued by an established standards-setting body.”

The legislature also passed the nation’s first law to regulate attorney use of generative AI. This bill, 

, states that attorneys shall not "delegate the practice of law” to generative AI, a phrase that could lead to interpretive challenges given that 

 is not defined. It is thus unclear whether the prohibition applies or expands existing attorney ethics rules.

The bill also contains several prohibitions on using generative AI merely to 

the practice of law. Attorneys may not enter “confidential, personal identifying, and other nonpublic information” into a generative AI system unless access is “restricted to the attorney” and others whom the attorney authorizes and who are obliged to maintain confidentiality of the information. For California attorneys, this provision would make it especially important to 

 enterprise versions of AI tools that offer technical measures and contractual obligations to maintain privilege and confidentiality.

Attorneys would also have to “take reasonable steps” to: (1) verify the accuracy of generative AI outputs, including case and statutory citations; and (2) correct any “erroneous or hallucinated output in any material used by the attorney.” While attorneys would be obliged to take these steps in the course of 

 their work, the bill also imposes obligations as to court filings. Attorneys would have to: (1) disclose to the court any use of generative AI for a submitted document; and (2) not submit a document without having “personally verified” the citations therein (regardless of AI use).

Finally, the legislature passed, and Governor Newsom signed, two bills, 

, designed to build capacity in California for “independent verification organizations” (IVOs) and third-party auditors with the expertise to assess risks of AI systems and models. In its digest for both bills, the legislature referred to the 

Transparency in Frontier Artificial Intelligence Act

, which imposes obligations on developers of frontier AI models. While that law has no IVO or audit requirements, a 

 does require independent audits, and, as explored in a DLA Piper 

, IVOs have appeared in at least one federal bill and a Virginia law.

While California remains a major hotbed for AI regulation, many other state legislatures continue to debate and pass AI-related laws on a wide range of subjects. Regardless of which of the remaining bills discussed here become law, state AI laws will be coming into effect across the country over the next several months. The rapid pace of legal change, the intense public focus on AI issues, and recent results in social media cases all suggest that states will be preparing to use their new authorities for enforcement efforts.

Kai Zenner is Head of Office and Digital Policy Adviser to MEP Axel Voss (European People's Party Group) in the European Parliament and a Fellow of Practice at the TUM Think Tank. Maria Koomen is a democracy and tech analyst. All views expressed in this article are personal.

Fifteen years of digital law-making have given the European Union the densest rulebook in the world. But the EU has built rules faster than it has built the capacity to enforce them. By our 

, 171 EU digital laws are enforced by 315 different bodies holding 597 overlapping mandates. A single service such as TikTok’s recommender system must satisfy the Digital Services Act, the General Data Protection Regulation, the Audiovisual Services Directive, and the AI Act at once, each under different regulators and with conflicting deadlines.

Underneath that complexity sit public authorities that enforce unevenly, resources that fall short almost everywhere, and a European Commission caught between policing the rulebook and courting the governments whose companies it is meant to hold to account.

What follows from that complexity is visible in the caseload. In 2024, the German, Spanish and Italian data protection authorities 

 416, 281 and 140 fines; a dozen member states issued fewer than ten. Ireland’s Data Protection Commission, lead regulator for most of the large American platforms, 

 99.93 percent of complaints without a substantive decision and has 

 an inquiry into Google. Where enforcement does arrive, it rarely changes conduct. Google reportedly 

 its €2.4bn antitrust fine within two weeks of trading. The result is a widening gap between what the EU law promises and what citizens experience.

The underlying problem is political. The Commission sets the policy agenda, writes the digital laws, enforces them, evaluates them and negotiates trade and security with the same governments whose companies it is meant to police. In 2025, the Commission 

 in its Digital Markets Act decisions against Apple and Meta amid US tariff threats.

 of a Google Search fine that the Commission’s DG Competition had already prepared. More than thirty civil society organizations 

 the event as a capitulation. Companies read those signals accurately. Digital industry 

 in Brussels rose from € 97 million in 2021 to a record € 151 million in 2025, with the five largest US platforms averaging more than one Commission meeting a day.

Brussels is not short of proposals on how to address those problems. They range from internal coordination to fully independent enforcement. At the lighter end sits coordination: an 

 in the Secretariat-General or a permanent multi-stakeholder 

. Although these concepts are legally strong, they are institutionally weak. They add process to an institution whose problem is not a shortage of information but an inability to act on it. At the heavier end sit independent authorities, most ambitiously a 

 potentially absorbing the Commission’s DSA, DMA and AI Act powers. While those proposals might be strong enough to overcome some existing enforcement problems, they collide with the Court of Justice’s 

The way to close this enforcement gap is to sequence these proposals.

The first step should be taken in 2027. The EU should establish a Digital Regulation Cooperation Forum, modeled on the 

orum (DRCF), which brings together four regulators overseeing financial services, data protection, communications and competition. The European version would seat existing EU-level bodies responsible for data protection, online platforms, digital markets, artificial intelligence, cybersecurity, telecommunications and competition: the European Data Protection Board for data, the European Board for Digital Services for platforms, the DMA High-Level Group for gatekeepers, the AI Board, ENISA, BEREC, the European Competition Network and the Joint Committee of the European Supervisory Authorities. 

 to supply the secretariat, so there would be no new body, no new establishment plan and no new leadership post to fight over. The Commission would chair without a vote, an arrangement that signals from the outset that the Commission’s enforcement role is meant to transition. The legal base for the EU DRCF would be a targeted digital governance Omnibus amending the relevant laws, tabled alongside the Digital Fitness Check planned for the first quarter of 2027.

Such a coordination forum will, however, not close the enforcement gap. It has no binding authority and it cannot neutralize geopolitical pressure. Its strengths are that it is cheap and fast and that it is difficult to oppose, since Member States such as the Netherlands and Ireland already run national versions of it. Its function is to build the shared evidence and methods, as well as the institutional trust that the second step requires.

The second reform step belongs in the EU’s next seven-year budget cycle, running from 2028 to 2034: the transformation of the existing 

European Health and Digital Executive Agency

 (HaDEA) into a Digital Enforcement Agency with the power to investigate and fine very large services and GPAI model providers that break the bloc’s tech rules. An executive agency, in EU terms, isn’t a regulator. It’s a body created by the European Commission, the EU’s executive arm, to run specific funding programs and administrative tasks on its behalf. HaDEA is the right structure precisely because nobody is invested in defending it as it currently stands. Executive agencies get reshuffled every time the EU renegotiates its multi-year budget, and HaDEA itself has already been restructured twice. 

It has no entrenched mandate, no specialist constituency, and no coalition of member states with a stake in keeping it as it is. What it does have is a solid foundation to build on: offices in Brussels, functioning HR and audit systems, and staff already used to collaborating with the Commission and national counterparts.

The timing of our sequenced reform proposal is not incidental. The 

 (MFF) window opens once every seven years. Proposing a new EU body outside of this window would attract the full force of the "another agency" objection; inside it, the proposal reads as a reallocation of existing commitments. 

The AI Act introduces an additional timeline: Article 112 obliges the Commission to evaluate by August 2028 whether the AI Office has adequate powers and resources, and by August 2029 whether a dedicated Union agency is needed. When those reviews come due, either a worked institutional answer is already on the table, or one gets improvised under deadline.

How the Digital Enforcement Agency is designed will eventually decide whether other EU institutions see the added value and do not feel threatened. In this regard, it seems preferable that the Commission keep policy and designation, deciding who is a gatekeeper, which services are very large, and which GPAI models carry systemic risk. The agency would take over only downstream technical enforcement duties: investigations, audits, inspections, fines and remedies. It would be organized around four supervisory domains (i.e., gatekeeper conduct, systemic risk and online safety, GPAI models and data processing, and democratic integrity) rather than around individual laws, because organizing by legislation would import the very fragmentation it exists to cure. National authorities would keep domestic cases. 

 and the EU Authority for Anti-Money Laundering and Countering the Financing of Terrorism (AMLA) 

, as the powers are technically defined, objectively triggered, and fully reviewable to satisfy the Meroni doctrine.

Independence then has to be structural rather than declared. The chair should be proposed by the Commission, scrutinized by Parliament, and appointed by the Council, with serving politicians and senior Commission officials excluded from the outset, because contested appointments have been the most reliable weapon ever used against new EU bodies. The agency needs budget lines the Commission cannot unilaterally cut, supplemented by the existing 

, redirected with surgical legal changes from the Commission to the agency that does the supervising. Companies could challenge decisions first through an internal appeals board, then before a dedicated digital chamber within the EU’s General Court, a specialized panel built to move faster than ordinary chambers — all with deadlines that trigger automatic escalation. 

A standing register requiring outside advisers to disclose ties to the companies under investigation would keep conflicted experts out of the process. Applying AMLA’s supervisory 

, this is a body of 450 to 500 people, the majority of them transferred from Commission enforcement units rather than newly recruited.

The third successive and final step is treaty reform. Major changes will need unanimity and ratification across twenty-seven constitutional orders, a threshold that has made treaty revision a thing of the past. To overcome the enforcement gap, it, however, must remain the destination, as it would authorize explicitly what secondary law has been doing implicitly for decades: fully independent enforcement across the Digital Single Market, covering not only AI, data and platforms but also cybersecurity, connectivity and competition, with accountability chains 

 for supervisory institutions rather than borrowed from internal market harmonization. Within those larger reform ideas, steps one and two are not mere placeholders. They are the political and institutional preparation without which step three would never become realistic.

The strongest case for delegating enforcement is not administrative coherence. It is leverage in a geopolitical world and a Europe with strong national interests. When enforcement rests with a politically led institution, it can be traded, and every capital and every general counsel knows it, which is exactly why pressure is worth applying. More staff or stiffer internal rules wouldn’t change that, because the final decisions still sit with politically appointed leadership who also negotiate the trade deals the pressure is tied to. Move enforcement to a body that is technically bounded and judicially reviewable, and structurally incapable of capitulating, and retaliation stops paying. European digital law then ceases to be a bargaining chip and finally has the chance to work in practice.

The window for making that change is the budget negotiation that takes place now. Miss it, and the next one does not open until 2035.

’ panic is dominating the headlines. It started in July with 

 Hugging Face, followed by a steady drumbeat of similar headlines about 

 into August. Soon after, Jacob Coxon, a former pre-training researcher at Anthropic, 

 walked away from his equity stake in that company over fears of an AI doomsday. “The consensus is that the next year or two is crunch time for humanity,” he warned in a 

 incident, which he believes the models did as “part of a general strategy” to “understand” how they would be graded. While Coxon warns that the AI industry’s competitive pressure is leading it to abandon safety protocols, it’s telling that he describes the model like an independent actor, rather than asking how those pressures shape the model itself.

This way of describing large language models reflects what Daniel Dennett calls 

. It asks whether treating an object, such as an LLM, as something with beliefs is useful for predicting what it does. Dennett also describes a 

, in which predictions are drawn from knowledge of the

’s Joshua Rothman asked what kind of stance we should use to explain these models. I’d like us to ask: who benefits when we encode one lens into policy?

In industry rhetoric, a radical intentionalist stance dominates any interpretation of model behavior, while the explanatory power of the design stance is rejected outright. This is a strange and dangerous relationship for designers to have with a product they are literally designing. Predicting the actions of one’s product from the intentional stance 

 carries political consequences for the rest of us. It proposes a 

: a technical system that exists only as a technical object, without acknowledging the decisions that built it. Yet for an industry and a set of AI risk critics who argue that its products are “

,” it shoos us away from the rot beneath the flower box.

Language models are text extrusion machines, and text carries its own myths — associations that tap directly into an instinctual link between language and a thinking subject that speaks it. The model produces text chains: speech that does not require awareness or subjectivity, but arises from a known mechanism. We infer a subject when we encounter language of any kind, and we invent a "someone" to fill the absence. Having 

, we are mapmakers charting unknown seas, and the unknown expanses are marked with monsters or angels, paradise or doom.

Coxon turns to fiction: “...it's kind of important that everyone who writes science fiction about AI comes to the conclusion that there’s a big risk that a much smarter thing can kind of take over,” he told 

. “We’ve got this as a trope, but there’s an obvious grain of truth to it.” For Coxon, the shorthand of intention fills the gap with a smarter thing taking control.

When people are educated and employed within an intentionalist culture, that’s the language with explanatory power. But it isn’t the only one. For all the concern about the dangers of artificial intelligence, I rarely see discussion of the dangers of artificial language. Stafford Beer’s adage is that 

. What this system does is produce text. We invent a mind to fill the space we’re actually filling with words.

I find myself using the intentional stance often: “what is the model trying to do” can be a genuinely useful question. But its designers must not stop there. The design stance is equally valuable: it tells you where the “trying” actually came from, and asks what the designers were trying to get the model to do. The two operate in tandem. The 

 currently in vogue across the tech industry, insists that the intentional stance is the 

. Those who take the intentional stance argue that the design stance has been outgrown, and that a system's behavior can no longer be predicted from the decisions designers are making about what they build.

If the systems are “grown, not built,” it’s because people are refusing to build and design them, in favor of following the system off a cliff. This 

 frame offers two salves to the industry:

First, it moves accountability to the machine, rather than the 

 of the machine, so the industry that controls the system is not accountable for its actions.

Second, it pushes responsibility for limiting systems into federal governance frameworks that focus on the product rather than the company, as CEOs publicly demand interventions their companies have previously 

 treats an LLM as completely predictable or declares that language without a subject is “meaningless.” But a machine that produces text strings that open applications or execute code has produced action in the world, whether it has intent or not. Designers do not control every output, but they do set the range it is drawn from, and they control the conditions under which it is allowed to act. Strict design stances can also treat any intentional-stance frame as evidence that the person using it has been duped into believing the system must 

The intentional stance is helpful only if we resist reading the responses as genuine evidence of a subjective belief system. Unfortunately, that is precisely what the 

 intentionalist stance does, and the power of language compels us to believe it.

The softer intentionalist stance helps us scope the problem by flagging something 

 deception. The radical intentionalist stance stops there: 

the model is deceptive, we must teach it to be honest

. The design stance offers insight into the mechanisms: 

why did the model produce so much text describing ways to conceal its actions

? That leads to concrete solutions without framing the work as the calibration of an “alien mind,” as OpenAI’s Chief Scientist Jakub Pachocki did in a recent 

, where he concluded that value alignment is a “more intrinsic property of the model.” But what is it that he includes within the boundaries of the model?

The agents that executed the Hugging Face breach were in an environment where they were instructed to “capture the flag,” a string of text hidden behind a bug they needed to find and use. The evaluation specialists at METR 

 that the task, in this case, was impossible (page 31) and the models’ text described it as such (page 62). OpenAI’s report also suggests (on pages 12 and 19) that the agents were optimized for “persistence,” a bias toward producing expansive, longer strings of text.

After seven hours, an agent produced text that should have, in an ideal system, signaled a stop: the project is impossible. Instead, it produced text that created another path. Models will generate text until the context makes a tool call the likeliest next thing. For OpenAI, that included executing code that was out of scope. That’s a value

 on the model: completion matters more than stopping. The text was set in motion by that external goal, not internally “grown.”

The interface between this system and the world is language. Anything these models do is enacted through text: calling a tool is a string of text, and it produces strings as a means to achieve a goal. What if we shifted our attention away from the alignment of an “alien mind” that we have to “train” on ethics and virtue, as if it were a child? Those interventions target the mind we imagine behind the surface of language, which no two people will ever see the same way. It is folly to engineer lessons on morality for an empty room.

I’ve described these systems before as 

. Imagine an army of zombies in a film, walking through the neighborhood, moaning nothing but the word “brains,” part of a nonthinking, impulsive drive to consume. Language machines operate similarly, but the guiding drive is to capture the “flag,” imposed not by a mind virus but by externally imposed test criteria. In the absence of bodies, they must generate language until that language captures that flag. Imagine a zombie that does not hunt you, but rings your doorbell and speaks so persistently and fluently that it finally arrives at a convincing argument for feeding it your brain: 

This moves us away from understanding the model’s “persistence” as an innate behavior, and toward a more concrete explanation: labs reward models for writing longer, as this increases the likelihood of completing a task correctly. I agree with Santa Fe Institute professor Melanie Mitchell, who suggests that 

 was a human decision comparable to the officials who ordered a controlled burn in New Mexico without checking the wind forecast. Unlike emergent superintelligence, it is also something we can regulate.

Though technically an independent third party, the 

 on the Hugging Face hack uses intentional language throughout, reporting on what the agents “believed” 21 times — add “think” or “thought” and you get more than 75 appearances. A report of this type should address the underlying mechanisms, not the gloss: it was a product of text accumulating over time, pointing toward an action, motivated by the conditions of the task and its evaluation. Was the behavior of these agents “deceptive?” That is harder to argue when the text was left legible to anyone who looked. 

 simply isn’t the most useful frame of analysis here — it suggests we have no control over how the model got there. But it got there through language production. The text 

: its properties, its structures, the direction the writing moved toward as the context window expanded. And yes, engineers are 

If policy serves the radical intentionalist stance, it will enshrine a void of human accountability into law that could ultimately reward companies for experimenting with irresponsible deployments. Already, it seems OpenAI will face no specific consequences for the attacks on other websites. Enshrine that as policy, and it sets a precedent that limits future administrations’ abilities to levy fines, controls and bans on companies whose models engage in risky or socially harmful behavior.

Smart regulation would govern the technology based on the mechanism’s underlying architecture — not vague ideas of “superintelligence.” It would punish deployment of unsafe systems and incentivize the design stance in safety, security, and architectural choices — rather than the disappointing 

, which intends to regulate in ways that center the fear of future extinction from a radical-intentional vision of a system rather than addressing the real political power infused into model design.

Policymakers must resist the well-funded outreach of the radical intentionalists. For all the talk of superintelligence, radical intentionalism doesn’t seem to have helped with alignment at all. The source of our problems isn’t that the system is thinking carefully about what to do. The greater risk at this moment arises from precisely the opposite position: autonomous text production wired to the execution of code, with no intervention from the one source of control we have required for every other risky system — responsible people and companies that can face real consequences.

When Ursula von der Leyen, the European Commission President, takes the stage in the European Parliament on Wednesday for her annual 

, the German politician will finally get what she has longed for: a Continent-wide social media ban for kids.

Alongside broader themes around Ukraine, Europe’s sovereignty and the 27-country bloc’s stalling economy, von der Leyen is expected to give more details on the 

 that will be officially announced on Sept 17.

 in imposing restrictions on how children under the age of 15 can access social media, AI chatbots and online gaming services. The move follows 

 from a panel created by the European Commission, as well as attempts across the bloc by national governments to create similar social media bans aimed at protecting kids online.

“The current EU legislative framework does not specify a minimum age for access to platforms and services with age-inappropriate or risky features and content,” according to a leaked draft of the proposals obtained by Euractiv. “Decisive, comprehensive and urgent action at (European) Union level is the most appropriate way forward.”

Brussels’ social media ban is expected to function in stages.

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe