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

A close up of a spider in its web. Valeriia Neganova/Unsplash

A close up of a spider in its web. Valeriia Neganova/

Breaking the Open Web Won’t Stop the Bad Bots

A photo of a group of men in suits and ties, gathered outdoors, speaking and smiling with trees in the background.  

Leading frontier AI companies, their employees, AI researchers, and governments have all become increasingly vocal about the need to regulate “frontier” AI. But how do leading AI companies envision public governance? By examining the recent pronouncements of the five most vocal major AI model developers — Anthropic, Google (a subsidiary of Alphabet), Meta, Microsoft, and OpenAI — a close reader can better answer this question. .

Our research, as part of the Multistakeholder Approach to Participation in AI Governance (MAP-AI), summarized below, aims to deepen understanding of, and shape conversations around, the relationships, responsibilities, and recourse needed to build meaningful, appropriate, sustainable, and democratically aligned AI governance. This is an initial compilation and overview of relevant proposals and positions, which will be updated over time. The documentation and comparative analysis supporting this analysis is available 

A variety of proposals without clear rules

Reading through this recent flurry of “urgent” calls for regulation and the corresponding pushback from some 

, helps provide clarity on what leading companies are asking for and what is most urgently needed — namely, governance that is inclusive, grounded in human rights, and incentivizes responsible model development and deployment by clearly and appropriately assigning responsibility, liability, and accountability.

The chart below summarizes our findings, which are described in detail by the rest of this essay.

The most consequential disagreement among the proposals reviewed is about whether to slow down or stop frontier development. The call for “pacing the frontier” was most clearly stated in September by Anthropic CEO Dario Amodei in his aptly titled essay, “

.” This post echoes and elaborates upon Anthropic’s earlier “

,” which ask Congress to create a new agency 

, with the explicit power to test frontier models, oversee independent evaluations, and block a model’s release if it poses a significant risk of catastrophic harm even after mitigations. In the most recent proposal, Amodei also suggests global coordination around a standards-setting body that can control pausing or slowing down model release after a thorough risk assessment.

While OpenAI CEO Sam Altman was among others who endorsed Amodei’s September call for pacing the frontier, his company’s public positions have heretofore been more cautious on this front. Its “

Chris Lehane’s “reverse federalism”

 essay ask US states to pass matching laws that federal legislation would later absorb into a harmonized national approach. OpenAI proposes that the existing 

Center for AI Standards and Innovation (CAISI)

 be built up as the government’s central testing body, but without addressing whether it should be empowered to halt model releases. It also nods toward the need for global coordination, suggesting “a US-led international forum that establishes accepted standards, provides expert and impartial analysis of capabilities and risks, and makes the technology available to nations and companies that participate and follow the rules,” something Altman had 

Alphabet’s chief scientist and Google DeepMind co-founder Demis Hassabis has also endorsed Amodei’s call for pacing the frontier. In its more developed governance proposals, Google splits the problem in two: presenting frontier AI as a national-security question, and addressing everyday AI from a consumer-protection perspective. For frontier models, its “

” proposes a new industry-funded, government-supervised body called the Federal AI Regulation Organization (FARO), modeled on the 

North American Electric Reliability Corporation 

 his own thoughts on what FINRA-style regulation could look like. Developers would self-certify to FARO before a “significantly more capable” release, giving the regulator the ultimate power to intervene where necessary, although Google is careful to frame the approach as certification rather than a discretionary government power to block model release.

acknowledges the need for new laws to address highly capable AI foundation models, making the case for a new regulatory body, which would ensure models comply with existing laws, license the model for deployment, and license data centers for training and deployment of models based on security, export controls, and safety protocols. It also proposes “safety brakes” that let critical-infrastructure operators disengage a misbehaving system. To this end, Microsoft committed to sharing specialized knowledge about its foundation models and supporting governments in defining requirements for licensing and enforcement. Microsoft has articulated further positions on pacing and other AI safety commitments in its recently published draft “

Meanwhile, Meta rejects the premise of prerelease review. Mark Zuckerberg’s “

” letter argues that no single company or government should be trusted to set a fixed review timeline for everyone and instead proposes that AI companies provide partly trained models and technical staff to the government while training is still underway, giving officials, in his words, “an army of capable engineers” to find problems in real time rather than checking a finished product. He frames this approach as more protective of open, decentralized power than a licensing regime, and warns that delaying a release “even by a month” could cost the US its lead in the global AI race.

The companies’ proposals also converge on the high-level outline of what should be done to identify and address deployment-related risks: publish a safety framework, test models against a set of risk domains, and make disclosures about the results. A similar approach has also been articulated by a range of countries through a 

 led by the governments of Finland and Norway. Beyond this basic testing and disclosure approach, differences emerge regarding the appropriate level of independence for testing, the range of risks to address, and the extent to which transparency should be mandated. 

Anthropic gets into the most detail on how to ensure independence among assessors, proposing the development of a professional, AI-specific evaluation industry, with published qualification standards, dedicated funding so evaluators aren’t paid by the companies they assess, government ratings of evaluator quality, and random assignment so companies can’t shop for an easy reviewer. It pairs this with a 15-day window to report stolen weights, realized harm from a named risk, or a model that deceives its own developers to evade oversight.

OpenAI and Google both support independent audits in principle, and OpenAI names them as one of three things every state law should require, alongside a public safety framework and incident reporting. Google’s FARO would run annual audits through established audit firms, which would initially focus on whether a company followed its own stated process, with an eye toward more substantive reviews occurring once standards mature, alongside three procedural protections for the companies being audited (standard document sets, a chance to fix problems before the report is final, and confidential delivery).

Microsoft bases its Blueprint on the know-your-customer-approach to financial risk management, adapting it to the AI context and arguing for “certain obligations to know one’s cloud, one’s customers, and one’s content” (KY3C). This approach would tie outside testing to the proposed licensing scheme rather than a freestanding audit regime. By contrast, Meta does not reference external audits at all.

There is also substantial ambiguity across these proposals regarding what these evaluation and audit frameworks would focus on. While every company proposal reviewed recognizes “existential” harms as a necessary focus for risk management, Google, OpenAI, and Microsoft offer broader approaches that incorporate additional risks, including deepfakes, privacy concerns, as well as energy, environmental, and employment-related issues. 

Transparency is a cornerstone of responsible business conduct, enabling third parties to hold companies and governments accountable. Over the past two decades, transparency reporting has evolved into a robust, established practice on digital platforms. Anthropic outlines the most extensive approach to transparency, noting that companies should publish safety frameworks, issue a risk report every six months, release a system or model card when more capable models are released, and provide certification of compliance with evaluation requirements. OpenAI, Google, and Microsoft also support publishing some form of overarching policy outlining key practices, though what it should contain differs from company to company. Microsoft proposes annual transparency reports and supporting a national registry of high-risk AI systems. In contrast, Zuckerberg’s post is quiet on transparency reporting. 

All five companies nod to the importance of reporting security incidents, though there is disagreement and ambiguity over what would constitute a reportable incident, how soon the incident should be reported, to whom it should be reported (the public, the regulator, the data center), as well as what should be reported. Notably, Meta is the only company that explicitly calls for cooperation with law enforcement in these cases. 

Anthropic and OpenAI also both push for mandated, industry-wide whistleblower protections, while the other companies are silent on this, even though Meta’s and Microsoft’s internal safety frameworks (their Frontier AI Framework and Frontier Governance Framework, respectively) include anonymous reporting channels and anti-retaliation commitments for their own employees.

Collective action challenges sit at the core of AI governance, and the tension between competition and coordination is evident across the companies’ proposals, with one notable exception: all five companies appear eager to develop predictable, confidential processes through which they can disclose risk-related information to the US government.

As a starting point, all five US companies target their proposals to the US political and regulatory context, with minimal references to the regulatory developments in the EU and elsewhere. While this is not surprising given that all five are US-headquartered companies, it nevertheless shows that companies may not sufficiently appreciate the fact that other democratic countries are unlikely to defer completely to regulatory approaches modeled on US institutions and enforced exclusively through the US legal system. These recent calls for regulation should also be interpreted in light of 

 in which some of these same companies have actively sought to undermine regulation in the US.

Where the companies do address the international dynamics around AI governance, they articulate a shared desire for interoperability with allies, either implicitly or explicitly leaving some countries out, while juxtaposing the feasibility of these efforts with the need to stay ahead of China specifically.

Meta stands out for favoring a “balance of power that favors individuals” focused on accelerating the diffusion of superintelligence. After committing to “contribute to this by building leading personal superintelligence agents and models, including open source models, and delivering them to billions of people around the world,” Zuckerberg states that “[a]ny policy that slows American model releases — even by a month — could add significant risk to American leadership while letting foreign models race ahead,” and focuses exclusively on encouraging cooperation between frontier labs and the US government, without addressing if or how international cooperation should take place.

While Anthropic’s earlier proposals focused on global coordination among democracies, Amodei’s most recent proposal goes further and proposes areas for possible coordination with China specifically. While acknowledging the likely limits on such cooperation, he proposes several areas for possible agreement, including prohibiting certain dangerous uses of AI, mutual commitments to prerelease testing, a SALT-like treaty for slowing recursive self-improvement, and finally, and least realistically, mutual commitment to a full-scale pause.

Within the US context, the companies have staked out diverging positions on the question of “federalism” — i.e., whether activities should be regulated exclusively and uniformly through federal law, or through potentially divergent state laws. Anthropic opposes federal “preemption” (superseding) of state law unless a federal law is passed that meets or exceeds the standards set out in their proposal. Most recently, according to 

, a draft bill circulated by the company would allow preemption “related to public disclosure of risk, third-party evaluation and reporting of critical incidents.”

By contrast, OpenAI’s “reverse federalism” model treats state laws as a bottom-up bridge to an eventual national standard, reserving national-security and highly technical review for the federal government and leaving states to lead on child safety, energy, and education. Google appears to defer to federal authority on national security without taking an explicit preemption position, while Microsoft’s Brad Smith has 

 the current patchwork as “regulation without transparent or complete rules,” and Meta’s documents are silent on the question entirely.

The companies also take somewhat different approaches to the issues raised by open-weight models. Not surprisingly, given its focus on open weights, Meta bases its proposed approach on openness as an antidote to concentrated power, arguing that restricting today’s open ecosystem “would be a mistake.” Microsoft acknowledges a genuine, unresolved tension between security and openness and joined a July 2026 

 from a wide range of companies, including Google and Meta, endorsing open weight models as an important part of a strong, open ecosystem that diffuses AI across the US economy. Meanwhile, Anthropic and OpenAI are missing from that list and do not address the question in their documents, beyond Anthropic’s stated concern with detecting and reporting attempts to copy or steal proprietary weights.

All five companies converge on the idea that the US government should get some form of privileged early information about and in some cases access to frontier capabilities, whether through incident reports routed to federal agencies and national labs (Anthropic), a federal testing framework that gets tools to critical-infrastructure defenders quickly (OpenAI), early notice for models that break new ground on sensitive national-security capabilities (Google), advance notification of large training runs, comprehensive risk assessments focused on dangerous or breakthrough capabilities, and extensive prerelease testing by internal and external experts as part of licensing requirements (Microsoft), or, most dramatically, handing over partially trained models and staff during training itself (Meta).

Finally, Amodei’s recent proposal calls for proactive safety-focused cooperation between leading US AI companies and calls for the government to mediate or enable those discussions “for antitrust reasons.” The idea of issuing such an antitrust waiver, which has also been supported by OpenAI and Google, has been 

 in the US Congress and has already provoked a 

These company proposals offer useful building blocks for developing regulatory institutions, evaluation mechanisms, safety frameworks, and information-sharing mechanisms. However, they are largely company-focused, US-centric, and anchored to technical or model-level safety. They leave open fundamental questions such as who should govern AI, in whose interests, for which context, and with what degree of independence; what mechanisms should exist for accountability and participation; and how domestic and international governance and cooperation should interrelate.

, rather than working toward a comprehensive treaty and a single global regulator, it is critical to develop multiple, intersecting, and coherent mechanisms to effectively govern AI. This ecosystem approach should be polycentric and people-centered. Coherence should be built with the norms and principles established through efforts like the 

Council of Europe’s Framework Convention on Artificial Intelligence,

 UNESCO Recommendation on the Ethics of Artificial Intelligence

, and the UN Office of the High Commissioner for Human Rights’ Foundational Paper on 

Advancing Responsible Development and Deployment of Generative AI

The authors have written this analysis in their capacities working on the MAP-AI project; nothing here should be interpreted as representing institutional positions of CCG, GNI, or GNI members.

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.

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe