In this photo illustration, the icon for the AI app "Claude" by Anthropic is seen through running water as it is displayed on a digital screen on September 16, 2026 in London, England. (Photo by Leon Neal/Getty Images)



Anthropic Needs to Disclose Its Tokenomics Before Its IPO

Anthropic CEO Dario Amodei looks on as United States President Donald Trump holds a press briefing with AI executives following a meeting on artificial intelligence, outside the West Wing of the White House on September 29, 2026 in Washington, DC. (Photo by Kevin Dietsch/Getty Images)

The SEC Should Block Anthropic’s IPO to Protect the Public

Director of National Intelligence and recently named head of the Super Intelligence Task Force Jay Clayton attends a roundtable discussion hosted by President Donald Trump on cybercrime in the East Room of the White House on October 21, 2026 in Washington, DC. (Photo by Alex Wong/Getty Images)

October 2026 US Tech Policy Roundup

A photo of a group of men in suits and ties gathered outdoors, speaking and smiling with trees in the background. Anthropic's CEO Dario Amodei is in the foreground speaking with big hand gestures and eyes closed. 

President Trump, left, speaks with Meta CEO Mark Zuckerberg as Dario Amodei, CEO and co-founder of Anthropic addresses reporters outside the White House, Sept. 29, 2026, in Washington DC. (AP Photo/Alex Brandon)

How AI Companies Want to Be Governed, and What That Means for AI Governance

Canada’s Prime Minister Mark Carney shakes hands with European Commission President Ursula von der Leyen following their closing news conference during the EU-Canada summit in Montreal on Friday, Oct. 30, 2026. (Christinne Muschi/The Canadian Press via AP)

Europe and Canada Plot a Digital Alliance as US Ties Fray

Michael Kratsios, Assistant to the President and Director of the White House Office of Science and Technology Policy attends Security Council meeting on Artificial intelligence and international security at UN Headquarters in New York, NY on September 23, 2026. (Photo by Lev Radin/Sipa USA via AP Images)

What’s Really Behind the Science Slowdown the White House Wants to Fix?

Photograph of a child staring at his smartphone in the dark with the light from the phone illuminating his face. 

Why Parental Consent Policies for Social Media Do Not Help Families

A photo from behind of a man in a checkered shirt pressing his hand against a wall with a hand-drawn black and white image of many raised hands with a bright red staircase shaped like an upside down heart painted in the center.  

A man fixes a billboard illustration of hands raised in protest before the start of a student led protest to demand the resignation of India’s Chief Election Commissioner in Fort Kochi, Kerala, on Oct. 18, 2026. (Photo by Sivaram Venkitasubramanian/NurPhoto)

Meta Streamlines State Censorship in India with Automated Takedowns

Citizens visit a venue for early voting for the midterm election in Virginia on October 30, 2026. (The Yomiuri Shimbun via AP Images)

There is No Universal Election Integrity Playbook for Platforms

How the EU AI Office Can See the Choices Behind Frontier Labs’ Safety Claims

An underwater photo taken looking up to a large circular school of fish. The image is distorted by digital artifacts.

Distorted Fish School II by Lone Thomasky & Bits&Bäume / 

When, How, and Why to Abandon the Internet

A collage illustration of a person holding a smartphone in front of their face, with their head replaced by a burst of social media icons (likes, hearts, comments, notifications, location pins and shares), stock charts and torn paper in blue, orange and teal, set against a cream background.

Mapping the Political Influencer Industry and the Problems it Causes for Democracy

Electrocutions, Intimidation and Wage Theft Haunt Europe’s Data Center Workers

A pedestrian passes a downtown mural honoring World War II veterans in California’s 22nd Congressional District on October 22, 2026 in Bakersfield, California, a competitive southern Central Valley district with a majority Latino population. (Photo by Mario Tama/Getty Images)

What Latino Voters Can Teach Washington About AI Governance

A digital collage combining a historical painting of a miner with a large grid of small online images. At the center, a muscular miner works underground with a heavy drill inside a dark tunnel. To the left and along the bottom, hundreds of tiny square images form a dense mosaic, suggesting scraped internet content. A bright beam of light connects the image grid to the miner’s drill. On the right, a block of colorful digital noise contrasts with the detailed painting, reinforcing the link between physical mining and data extraction.

Primitive Accumulation by Daniela Zampieri / 

What Training Dataset Transparency Mandates Hide

Demonstrators wear masks depicting Meta CEO Mark Zuckerberg, Brazilian presidential candidate Flavio Bolsonaro and US President Donald Trump during a protest outside Meta's headquarters in Sao Paulo, Brazil, on October 19, 2026. The protest was called over allegations that restrictions on Instagram affected the campaign of Brazilian President Luiz Inacio Lula da Silva ahead of the presidential runoff election. Meta attributed the reported issue to a technical failure and denied political bias. (Photo by Allison Sales/NurPhoto via AP)

Brazil Shows Platform Illiberalism is Big Tech’s New Normal

Grok app icon is seen next to the Meta AI and Claude app icons on a smartphone home screen in this illustration photo taken on October 27, 2026. (Photo by Matteo Della Torre/NurPhoto via AP)

The Solution to AI’s Alignment Problem is Precise Guardrails on How We Use It

A stock market chart with lines in different colors on a dark screen with lines trending downwards to the right. 

Chart on a dark screen by Maxim Hopman on 

Will Prediction Markets Affect Democracy? Take a Bet

Canada's Prime Minister Mark Carney, front left, speaks with European Commission President Ursula von der Leyen at the European Parliament in Strasbourg, eastern France, Sept 17, 2026. (AP Photo/Pascal Bastien, File)

Canada and Europe Need a Way to See the AI They Cannot Access

A surreal collage depicting a dishevelled pile of items arranged like a traditional vanitas painting. To the right is a vintage computer monitor displaying the OpenAI logo. Objects representing various instances of AI “slop” litter the scene, such as an idol of Jesus with a shrimp replacing its head, and an anthropomorphised cappuccino cup with cartoonish eyes and lips.

Slop Vanitas II by Georgina Leeuwendal / 

Creating a New Internet Bargain to Preserve Fair Use and Protect the Commons

Matt Scherer is co-author of the recent report 

Brace for Impact: Protecting people and the economy when the AI bubble bursts

Things are seldom what they seem when it comes to the finances of the AI industry. The AI ecosystem’s 

 make it hard for investors, regulators, or workers saving for retirement to get an accurate financial picture of AI companies.

That is particularly true of Anthropic, which filed the 

 for its initial public offering (IPO) last week ahead of a planned November stock market debut.

The company’s topline financial figures look quite impressive for a startup, with revenues that appear to have grown quickly over the course of 2026 and a fairly healthy balance sheet — on the surface, that is.

Dig a little deeper, however, and a far bleaker picture emerges. Footnotes to the company’s most recent 

 in spending commitments, most of them stretching into the 2030s. The company’s financial results for 2026 have not been subjected to formal auditing, which makes them less-than-reliable, but the third-quarter results suggest that the company’s revenue growth 

Anthropic will have to reverse that trend and sustain extraordinary revenue growth to meet its vast obligations. But the financial pressures that are squeezing Anthropic — and, indeed, the AI industry as a whole — will make it exceptionally hard for the leading model developer to do so. If the company goes public, workers, retail investors, and retirees will increasingly be left holding the bag when Anthropic’s bills come due.

The financial mountain that Anthropic must climb is unprecedented for a company making its stock market debut. While Anthropic has an impressive $96 billion in cash reserves, its 

 reveals at least $582 billion in non-cancelable spending commitments to just five companies (Amazon, AMD, Broadcom, Google, and Microsoft) to gain access to the vast computational power required to train and run Anthropic’s generative AI models.

Even though they do not appear on Anthropic’s balance sheet, those obligations are effectively debts 

: they are payments that Anthropic is contractually required to make in the future, regardless of whether the company can afford to do so.

Moreover, it is not clear how much of the remaining $300 billion or so of the $895 billion in contractual commitments will ultimately land on Anthropic’s balance sheet. With the company continuing to negotiate 

further massive infrastructure and compute contracts

, Anthropic may well end the year with a trillion dollars or more in contractual commitments.

Generating the revenues necessary to meet those obligations would be, to say the least, quite difficult for any company. But the margins of AI model developers are being squeezed by several forces that will make that task extraordinarily challenging for Anthropic.

First, it isn’t clear that there is nearly enough demand for generative AI products and services to justify Anthropic’s spending. Multiple 

 notable for their bullishness on AI, suggest LLMs are 

otherwise generating positive returns on investment

. Worryingly for enterprise-focused Anthropic, one study found that the number of Americans using AI at work 

 between the summers of 2025 (46%) and 2026 (33%).

Anthropic’s spending commitments suggest that the company believes it will be able to convince businesses to pay for ever-larger and more expensive models. But there are growing signs that most businesses are not interested in 

for more powerful and expensive models. When Anthropic released its much-hyped Fable 5 frontier model this summer, its commercial reception was 

Even if a future Anthropic model captivates businesses, the company would run into a second obstacle: the AI industry’s frenzied competition. Anthropic is part of a crowded Silicon Valley ecosystem where numerous well-heeled corporations (including Google, Meta, Microsoft, OpenAI, and SpaceX, along with Anthropic) offer generative AI models that are largely indistinguishable in their quality and capabilities. As a result, AI companies can quickly offer competing products whenever a rival identifies a use case that shows even a hint of commercial potential.

In such an environment, companies can rarely afford to raise prices because doing so will prompt customers simply to switch to a rival. The rush to grab customers ahead of their IPOs led Anthropic and OpenAI to launch a 

 this summer that remains ongoing. Competition from Chinese companies offering open-weights models puts 

 on prices, as does the increasing availability of models that can run locally on phones and laptops — no expensive data centers required.

That leads to the third problem: the extraordinarily high costs of the chips and data centers needed to train and run AI models, which put powerful and continuous upward pressure on Anthropic’s costs. Between the chips, servers, networking equipment, electricity and cooling infrastructure, construction costs, and other expenses, a 1 gigawatt (GW) data center costs an estimated 

, and the prices for some components are soaring.

That is for 1 GW. Anthropic has committed to 15 GW of data center capacity in the coming years and the company 

 it will ultimately need 20-25 GW, which explains the company’s half-trillion-plus dollars in spending commitments.

And that is just the cost of the physical infrastructure. Once a data center is up and running, training AI models entails using tens of thousands of chips running in parallel that require massive amounts of energy to power, run hot, and need to be kept cool. Moreover, each time someone prompts a model to answer a question, create an image, or perform a task, additional computational power, energy, and cooling are needed to process the query and generate a response or complete the task.

Finally, there are external factors that are delaying data center construction and piling on additional costs. The fierce and still-rising backlash to 

 across the country has scuttled a number of data center projects and delayed many others. Every delay limits the computational capacity Anthropic is counting on to generate the revenue it will desperately need, and rising costs due to the Iran War and supply-chain challenges could drive Anthropic’s operational costs still higher.

Anthropic will further pad its cash reserves if its IPO is successful, but the brutal economics of its business are not likely to improve anytime soon. Even if demand picks up significantly, intense competition will apply unrelenting downward pressure on prices even as high costs and resource constraints squeeze it from the opposite direction. That is a deadly mix for Anthropic in light of its slowing revenue growth and massive off-balance-sheet debts, a vast and immovable fixed cost.

If Anthropic were to go public now, it would pile additional risk onto a stock market that is already 

 in a handful of tightly interconnected mega-cap companies from the tech sector. That has already created 

a risk of a historically damaging stock market crash

; never before has such a huge amount of wealth depended on the performance of a small number of companies whose fortunes are inextricably linked. Adding another huge, unprofitable tech giant to public exchanges would exacerbate both of those factors and create a higher risk of contagion if (or, rather, when) Anthropic reaches a financial precipice.

It also would mean that an ever-larger share of the fallout from a potentially devastating AI crash would fall on ordinary people. As a private company, Anthropic’s shares are overwhelmingly owned by corporate insiders, tech giants, and venture capitalists. Once the company goes public and shares start getting funneled into 401(k)s, retirement accounts, and index funds, it will be workers, retirees, and retail investors that are increasingly left holding the bag when the forces squeezing Anthropic’s margins start to bite.

The Securities and Exchange Commission (SEC) has the power to stop that from happening. The agency can delay or stop companies from going public if it believes a company’s registration statement is “

on its face incomplete or inaccurate in any material respect

.” That legislative language might as well have been written with Anthropic in mind, considering that the company’s prospectus includes only unaudited financial results for the period where it claims the greatest revenue growth, relegates its most significant future obligations to the footnotes of its registration statement, and has continued to announce new 11-digit financial commitments as its IPO approaches.

 Anthropic’s stock market debut until the company provides two consecutive quarters of audited financial statements, with all material information about the company’s spending commitments disclosed. Failing to do so could have dire consequences for workers, investors, and ultimately the financial system and economy.

Much of the debate about election integrity on social media focuses on harmful claims: what misinformation is spreading, where it’s appearing, how far it travelled, and whether platforms are doing enough to stop it. This approach understandably focuses attention on the content itself and the potential impact on democratic processes.

But the same misleading claim can behave very differently depending on where it appears. A recommendation-driven video feed distributes information differently from a social graph-based network, a private messaging service, or a livestreaming platform. Each gives its users different tools for creating, amplifying, contextualizing, monetizing, and responding to political information.

Election integrity is not only a content policy problem; it’s also a product design problem. Before asking what election-related content a platform should prohibit, election integrity teams should examine the architecture of the product itself. This includes how information moves, how users interact, where authority and credibility lie, and which affordances can be exploited.

A framework for understanding platform affordances

When assessing election integrity risks, it’s important to distinguish between a platform affordance and the harm that may result from its use. An affordance is a capability, opportunity, or constraint created by product design. Five dimensions are particularly relevant to election integrity: distribution, or how information reaches users; content; the formats a product enables; interaction, how users respond to and reshape content; identity, how credibility and accountability are established; and monetization, how attention can become financial support.

These features aren’t harmful by themselves. They shape the incentives, opportunities, and pathways through which harms can emerge. The same affordance can support legitimate political participation in one context while creating an avenue for manipulation in another. Election integrity teams should ask not just what content their policies prohibit, but how political information is created, distributed, and acted upon through the product. Those affordances don’t operate independently. Different combinations produce different platform architectures, concentrating risk, authority, and opportunities for intervention.Various platform architectures create different pathways for political information to spread and therefore require different approaches to election integrity. Three broad models illustrate how architecture redistributes election integrity risk, although many platforms combine elements of each. Recommendation systems can distribute content well beyond established follower networks, making the speed and scale of algorithmic amplification particularly important to detection and response. Social graph systems place greater weight on established accounts and relationships, making networks, influence, and trusted intermediaries more important. Open or federated systems distribute some governance functions across services or moderation layers, creating additional intervention points but also greater challenges for consistency and coordination.

None of these architectures is inherently more or less safe. They redistribute risk, authority, and responsibility. Election integrity programs need to be designed around the architecture they govern rather than assuming that interventions effective on one type of platform will transfer cleanly to another.

Affordances also interact with the political and institutional environments in which they sit. A recommendation system can help voters encounter accurate information about how to vote while also amplifying misleading procedural claims. Where election rules vary across jurisdictions, information that’s accurate in one place may be misleading in another. Local threats matter, too. Features that support participation or transparency may create different risks where candidates and election workers face elevated harassment or violence. Platform architecture shapes risk, but electoral rules, institutional trust, political norms, and local threats manifest it.

The 2024 US presidential election demonstrates how the same election integrity problem can be distributed through different platform architectures. On October 31, 2024, a video appearing to show a voting machine in Kentucky switching a vote from Donald Trump to Kamala Harris circulated on TikTok, where it reached 

. Election officials investigated and found no evidence of vote switching. The device printed a paper ballot that voters could review before casting their vote, and the voter ultimately submitted a ballot reflecting her intended selections after all. Officials described the episode as isolated and found no other reported problems with the county’s ballot marking devices.

As the video moved across platforms, different affordances shaped its reach and interpretation. On X, influential accounts supplied their own framing. Podcaster and pundit 

 reposted the clip and asked whether it showed a broken machine or election fraud, drawing millions of views. Versions also circulated on 

 in English and Spanish, including posts presenting the incident as evidence of broader election interference. By November 1, 2024, the allegation had reached Meta’s microblogging service, 

; with versions later appearing in Portuguese, Greek, German, Vietnamese, and French. The video remained the same, but recommendation systems, influential accounts, and social networks created distinct pathways through which users encountered and interpreted it.

The 2026 US election cycle also shows how new content production affordances can complicate election integrity. In Texas, Republican Senate candidate Ken Paxton released an 

 depicting Democratic opponent James Talarico reading and writing in a fictional “King James (Talarico) Bible.” The ad used an AI-created likeness of Talarico in a montage combining authentic remarks with missing context and words he had not used. A small disclaimer identified some video content as AI-generated, but the format demonstrates how campaigns can use generative tools to depict opponents saying or doing things that never occurred. The challenge is not only whether synthetic content is disclosed. Realistic looking content can also make authentic political speech harder to distinguish from fabricated media.

Separately, researchers linked an AI-generated disinformation video targeting Talarico to the 

 influence operation. The original X post attracted little attention until larger verified accounts amplified it 12 hours later. It then received more than 850,000 views. X removed the original, but copies continue to circulate online. Together, these cases illustrate two uses of the same content affordance – domestic political advertising and foreign influence — and demonstrate why election integrity programs must consider both how political content is manufactured and how distribution systems give it reach.

These examples reveal a tension in election integrity work: responses must be tailored to individual platforms, even as threats move between them. Different services can play complementary roles in amplification, coordination, fundraising, private communication, or the circulation of synthetic media. Content may originate in one environment, gain credibility through influential accounts in another, and then be translated, remixed, or moved into less visible spaces. The relevant reference point is often the broader information ecosystem rather than a single platform.

Cross-platform coordination doesn’t require uniform enforcement. Platforms can build shared awareness of emerging narratives, synthetic media, coordinated tactics, or threats while applying different interventions based on their architecture and policies. One service may focus on recommendation or distribution, another on influential accounts or network behavior, and another on provenance or labeling. Shared situational awareness can exist with platform specific thresholds and enforcement choices.

What election integrity programs can do differently

Election integrity programs should respond to these differences in platform architecture in three ways.

First, start with the product, not the policy. Before deciding what content to prohibit or how to enforce against it, trust and safety teams should map how political information moves through the platform. This includes how users discover content, interact with one another, establish credibility, and move between online communities. Policies should respond to the environment the product creates rather than assume the rules that have been developed elsewhere will necessarily transfer cleanly.

Second, build platform-specific threat models. Recommendation, social graph, and open or federated systems create different opportunities for amplification, interaction, coordination, and intervention. Effective election integrity planning should identify which affordances are most susceptible to misuse, how those vulnerabilities interact with the political context, and where the platform has meaningful opportunities to intervene.

Third, treat enforcement as one part of a larger system. Election integrity requires coordination across policy, product, detection, engineering, operations, communications, public policy, and external partnerships. Because election-related activity can move between services, platforms should, where appropriate, share information about emerging threats and tactics with other platforms, election authorities, researchers, and civil society. That coordination should improve collective awareness of the threat environment without assuming that every platform should respond in the same way.

The same principles should guide policymakers and researchers evaluating platform preparedness. The central question shouldn’t simply be, “What misinformation might appear here?” It should be: “What does this product make it unusually easy to do, and unusually difficult to prevent?”

After India’s Chief Justice flippantly referred to the nation’s unemployed youth as “cockroaches” in May 2026, it was not forgotten. The backlash led to the formation of the satirical 

 (CJP). It started as an Instagram page, posting reels and memes to raise awareness about government corruption and advocating for education reform. In demanding the resignation of India’s education minister over a paper leak, 

, and on July 25, the minister resigned. But CJP’s success in mobilizing Indians to speak up and organize is driven by its strong online presence. Their Instagram page 

 followers and counting—more than that of any individual political party in India. In the story of CJP’s rise, as 

 writes, Meta’s Instagram was not merely a digital record of the movement, “it was one of the principal engines of it.”

In the aftermath of a protest on July 20, the party’s Instagram account was 

 in India. “Cockroach Janta Party’s Instagram account has been taken down by Narendra Modi,” CJP founder Abhijeet Dipke 

, sharing a screenshot that merely stated that Instagram restricted the page following a “legal request.” Other CJP members 

 that their posts about the protests had been restricted. These legal requests are de facto orders, and they go far beyond content posted by CJP members: between March and July 2026, the government issued nearly 

, up from just 2,300 between October 2024-25. But now, the orders Meta receives are reportedly being 

The Indian government has primarily relied on content takedown orders to stifle online dissent. These orders are often premised on alleged violations of law, in this case, India’s 

, which set the boundaries of permissible online speech in the country. Meta has long complied with such orders, in part because failing to do so risks its safe harbor protections, which shield it from legal liability for what its users post. Now, amendments to India’s IT rules that took effect in February mandate that platforms comply with takedown orders 

. These orders are now issued through a government system called the 

 which allows India’s law enforcement agencies to request the immediate removal of content from digital platforms. To comply with this new three hour rule, Meta has reportedly connected its systems directly to Sahyog so that removal requests are 

, without review by Meta employees. In other words, the platform that helped propel CJP’s rise is now thwarting its influence—and the process is largely automated.

Despite the government’s best efforts to curtail CJP’s online and offline influence, the movement has only strengthened since its inception. On October 10, the CJP led 

 in Delhi, this time demanding the resignation of the Chief Election Commissioner over controversial 

 lose their right to vote. The government, in response, cancelled trains, closed metro stations, 

, including journalists. But the crackdown extended to platforms too. A week before the protest, on October 3, 

, a messaging service that works via Bluetooth and does not require internet connectivity, was 

 from the Google Play Store and Apple App Store in India.

Delhi Police detained protesters and journalists

, India’s Ministry of Electronics and Information Technology 

 asking the company to review its recommendation algorithms to ensure it does not result in “disproportionate exposure of users to a single or limited set of narratives,” and that Meta must provide “information or assistance” to law enforcement “with immediate effect.” The next day, 

, showing a woman clinging to a pole as police officers attempted to detain her was restricted on Instagram in India. This is not an isolated instance. Numerous Indian citizens have reported seeing their accounts blocked or posts removed. By mid-October, the Software Freedom Law Center—a digital rights group based in India—had documented 

 taken down in connection with the protests on its 

When Meta follows government orders to restrict content in India, they rarely tell users why. In cases examined by 

, Meta’s notifications cited only a “legal request” or using an “automated system,” without saying who ordered the restriction or why, and offered no option to appeal. But now, it seems Meta is using automation to proactively restrict content in India, and this time, without explicit government takedown orders. In the first half of 2026, Meta restricted over 

 pieces of content. Notably, this content was restricted based on Indian law, not for violating Meta’s own standards, and the content remained available outside India.

5.4 million is a staggering number. For context, in the previous six months between July and December 2025, around 

were restricted. By Meta’s own admission, a “

” of these restrictions were proactive and responsive to their legal obligations in India. Some of these obligations are especially vague and concerning. Content categories like “threats to the sovereignty and integrity of India” and “promoting enmity between groups” are notoriously hard to define, and the fact that Meta uses automated tools to make these judgements, exacerbates and brings to bear the depth of online censorship in India.

Takedown notices are far from transparent, especially because blocking orders under Section 69A of the IT Act are, by law, kept 

. But takedowns leave a trace, and in theory, can be challenged in a court of law. Meta’s move to proactive and automated restrictions, however, makes even this difficult. There may be no government order behind a restriction at all, and users are often told only that an “automated system” flagged their content under the IT rules.

The pattern is clear: stifle dissent on the streets via law enforcement and a sophisticated surveillance apparatus, and bully platforms to remove unfavorable content via takedown orders. But Meta’s move to automate both the processing of these orders and now to proactively restrict content using automation means that Meta is no longer just complying with the Indian state’s censorship, it's streamlining it.

, the European Commission can enforce general-purpose AI obligations under the AI Act through its AI Office, demanding documentation (

 putting EU safety teams inside frontier labs.

While access is necessary, the Office is enforcing against definitions it has little power over and that are not easily comparable. Compliance, frontier safety, and what “frontier” even means are still partly defined and acted upon according to 

providers’ own criteria and internal know-how

. More importantly, this makes the Office blind to a crucial kind of evidence. AI safety is not only about reporting technical findings or methods. It involves taking positions about 

 for safety-relevant features. These decisions shape a model’s safety profile. Yet, providers are not held accountable for them, nor are they compelled to disclose them as decisions. Access is the first step to achieve that.

The solution I propose is a public taxonomy that makes these choices visible without requiring providers to make them in a specific way. The possibility to contest these choices in fact depends on them being visible 

non-technical choices. Otherwise, these judgments give external actors nothing to push against, and no basis for attributing responsibility when something goes wrong.

How judgment calls pass as technical and necessary choices

, a voluntary instrument providers sign to create a presumption of compliance with Articles 

. It sets out important commitments, naming the steps of development where disclosure is owed. But it does not say what counts as relevant and how it is described and leaves that definition to providers.

A provider can meet requirements while disclosing little, or disclose a great deal because it chooses to, in which case the disclosure could align with safety purposes only contingently. Some sections of the Code do demand justification of, for instance, risk-tier acceptance criteria in the Safety and Security chapter. But they still leave space for discretion over what counts as an adequate justification and how much to justify. Other non-technical choices across models’ lifecycles remain untraced and unsystematized.

Three instances show what this means: one where a value choice reads as a method, one where a concept is contested and cannot be settled by better science, and one of corporate discretion.

The first instance can be found in OpenAI’s recent 

. Astra is trained to refuse dangerous requests. If, when evaluated for what it 

do, it refuses to pursue a user request, the refusal is 

 ambiguous about what this means. OpenAI 

, §10.1.1), treating the model as capable and simply not wanting to help whenever it declines. This is a choice about which error to prefer, whether over-attributing capability or missing it. It is defensible, and it is disclosed to some extent. Nothing in the system card differentiates this choice from a technical one,

The second example appears in the same card (

, §9), reporting that monitoring Astra’s reasoning became harder and less reliable. OpenAI deployed anyway, treating the model’s underlying alignment as the main safeguard and claiming they will not accept further loss of monitorability. More evaluators or better science will not automatically bridge this gap. Extra experts still grade against standards set by providers, and while better methods close the gap where a science is immature, no amount of methodological maturity alone settles what “harmful” or “good evaluation” mean.

The third case is about high-level corporate decisions. Anthropic 

did not give the UK AISI pre-release access

 committed the company to embedded third-party evaluators just a few days later. Either decision could have been made in good faith, commercial judgment, external pressures, or all combined. We might just be lucky that this happens to point in a safety-promoting direction. Good faith and other non-safety incentives ultimately end up doing work regulators ought to be doing.

These decisions may all be defensible. But a model’s safety profile is built out of dozens of such calls, and a generic instruction to disclose acceptability criteria reaches none of them. Regulators must outline 

they want justification to happen, even if they do not set the standard for what justification should look like.

Asking regulators to set fixed safety standards is overly optimistic and may be the 

 given how fast the technology moves. The next Code revision should instead make value-laden decisions disclosable and comparable across time and providers, through a regulator-owned taxonomy of the steps where non-technical judgment calls can happen across a model’s lifecycle. This would let the Office track patterns across providers (which positions cluster, which labs are consistent outliers and why), and challenge the choices behind them. Importantly, judgment that exists only as informal practice inside a handful of labs becomes a shared and readable record, so the Office stops depending on providers' word for where the relevant information sits.

First, providers locate their model within a public taxonomy they did not create.

The Office maintains a public taxonomy of the steps across a model’s lifecycle where value-laden choices are inevitable. This could follow the Code’s existing stages, from risk identification through post-market monitoring. A first version need not wait for providers: all three examples above come from public documents. Today the Code asks providers to justify conduct and criteria they set themselves, so that providers can define what decisions are worth disclosing, at what level, and for what purpose. Under the taxonomy, they would instead locate each decision, and the principles behind it, in a systematized and more fine-grained structure they did not establish. This reduces the Office’s dependence on providers. If we do not know what decisions they make at which step, we cannot enforce against them.

Second, the checking can and should be outsourced. Independent accredited bodies can be delegated to audit whether the provider’s reported placement is honest. This is a genuine measurement against a fixed public reference. Evaluators never decide which position is correct, only whether the evidence produced matches declared positions.

Third, the taxonomy records positions rather than grading them and bites on honesty

The taxonomy does not aggregate judgments into single statements that a model is safe, and providers can adopt different principles across steps while reporting honestly. Enforcement would at first bite on honest judgment rather than on verdict to encourage honesty. What can be adjudicated is whether the provider placed its model faithfully, and said where its judgment sat when more than one position was available.

Fourth, the framework and the placements stay contestable. Where a disclosed position and evidence come apart, providers can be investigated under the AI Act. The Office remains responsible for filling gaps in the taxonomy. Because the structure is common across frontier labs, positions also become comparable over time. Finally, making these choices legible and comparable allows regulators to assess where the frontier of AI 

lies and how much it really tracks the frontier of capabilities.

The Office does not need more substantial powers to enforce this. The next Code of Practice revision could define the map with the help of the Scientific Panel (

) on the basis of the Code’s demonstrated shortcomings. 

 covers documentation requests, which could be structured against this new taxonomy. Under 

, the Office evaluations could test placements against the taxonomy and misplacement or missing entry could count as incorrect or misleading information under 

Placements would be public; the evidence behind them, which may be commercially or security-sensitive, could go through the confidential channels the Office already provides. This sharpens enforcement, as outliers become visible. With time, the Office could further complement this with its own guidelines about general value-laden trade-offs, such as prioritizing privacy in some domains.

The taxonomy is ultimately a middle ground between standardizing frontier safety and industry self-regulation and could reduce the Office’s epistemic dependence on frontier labs. Even embedded evaluators, with access to everything, would still need to know which decisions to look at. The Office could target its own evaluations at unusual steps, structure documentation requests more precisely, and trace which judgment call mattered when something goes wrong. Without such a structure, nothing breaks the cycle. Providers hold control over what is safety-relevant and how it is described. Described as technical decisions, judgments draw no scrutiny, and without scrutiny, no one can describe them otherwise.

A crowd of supporters surround President Lula holding a microphone at a podium draped with the Brazilian flag.

President Luiz Inacio Lula da Silva delivers a victory speech after winning the presidential runoff in Brazil, Oct. 25, 2026. (AP Photo/Andre Penner)

Lula’s Win Signals Fight to Regulate Social Media Will Escalate in Brazil

An illustrative photo of a man walking through the corridors of the European Parliament in Strasbourg, France.

Behind the scenes - European Parliament in Strasbourg

Despite Spyware Attacks on Its Own Officials, EU Response Falls Short

A person holds a giant Brazilian flag while perched above a crowd of people at a rally on a large avenue, a few people are holding phones in the air. One person holds their hand in the air in the L shape, signifying support for Lula. 

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe