The flare-up earlier this year between Anthropic and the U.S. Department of Defense has exposed a deeper crisis: our systems for holding AI development and deployment to account are broken. Companies enforce their own internal rules when convenient; governments regulate when politically expedient. What is missing is an independent public interest layer of accountability: a third way between corporate self-regulation and state enforcement. Replace “AI” with “social media,” “quantum computing,” or “biotech,” and the statement still stands.
On one hand, tech companies often show the best intentions, yet do not practice what they preach. Many cheered Anthropic’s CEO, Dario Amodei, for resisting the U.S. government’s use of his company’s models for mass surveillance or fully autonomous weapons systems. But on Feb. 25, the company quietly gutted the central safeguard of its Responsible Scaling Policy: the company’s commitment that it would not deploy or train models until adequate safety mechanisms were in place. That previous pledge, clear in the policy’s 2023–25 versions, virtually disappeared from the 2026 revision. As the company’s chief scientist explained, it no longer “[made] sense for us to make unilateral commitments… if competitors are blazing ahead.”
And while Anthropic was still being celebrated publicly, we learned its model had already been used in U.S. military operations in Venezuela and Iran.
On the other hand, governments often overreach. DoD’s classification of Anthropic as a “supply chain risk” perplexed many, some of whom have commented that it appears to be disingenuous political retaliation, potentially even in violation of protected constitutional rights.
There are more subtle examples: under the guise of combating online disinformation or protecting children, officials have drifted into censorship. In Gabon, for instance, authorities suspended platforms like Facebook and TikTok for alleged policy failures, citing specifically the “spread of false information,” “cyberbullying,” and the “unauthorised disclosure of personal data” as reasons for the decision. Australia’s age bans on social media risk undermining minors’ rights to expression and information. And Indonesia and Malaysia banned Grok after the AI model’s deepfake scandal involving sexualized content of women and children.
It is worth noting that widespread online platforms bans and their impacts on human rights have generally been criticized by the U.N. Special Rapporteur on Freedom of Expression. But when faced with a company led by a CEO like Elon Musk, who has been accused of ignoring legal orders from authorities—including Brazilian judges—outright platform bans may be tempting. But it is not clear that these comprehensive bans pass the three-part test of legality, proportionality, and necessity that the U.N. Special Rapporteur on the Promotion and Protection of the Right to Freedom of Opinion and Expression has opined must be met before states restrict the freedom of expression (in alignment with Article 19 of the International Covenant on Civil and Political Rights). There is no definitive answer on how to best sanction a company (often its CEO) for refusing to comply with legal orders; some countries have explored criminal liability, while others impose fines. In any event, such sanctions should always seek to be legal, proportionate and necessary to legitimate aim pursued.
In 2022, I drew on the concept of lex mercatoria—the self-regulating commercial law of medieval merchants—to propose a Lex Platformia, a quasi-legal normative order that technology companies, often social media, have spontaneously generated. For such a quasi-legal system to work, I argued, effective external accountability mechanisms were essential.
Four years later, the same lesson applies to AI. Companies have grandiosely promised to anticipate the harms of innovation rather than repeat past mistakes. Yet history is repeating itself. Internal accountability efforts often fall short. We see it with Anthropic’s Responsible Scaling Policy. The limited efficacy of accountability behind closed doors is also visible at OpenAI, where the presence of Safety and Security Committee, whose decisions are not public, did not prevent the company from adding mental health capabilities to its flagship chatbot—despite calls for caution, including from past employees, about the privacy and medical implications.
If neither states nor companies can be trusted to protect the public, who can?
We need a third way—one that is independent from both governments and corporations, answerable only to the public interest.
This “third way” should not consist of a single entity, but an ecosystem of layered accountability: independent oversight boards, civil-society watchdogs, academic labs, and “user councils” acting in concert as counterweights to state and corporate power.
One of the existing layers is Meta’s Oversight Board, the only attempt so far at independent platform governance. Yes, it has flaws, not least its origin in the company it oversees—Mark Zuckerberg reportedly had the idea of a Facebook “Supreme Court.” But that should not dismiss it outright. As one of its founding members, I recognize both its weaknesses and its promise.
The Board’s democratic model is imperfect. Many initial members, myself included, were chosen by Meta. Today, recruitment lies with the Board itself rather than with the company it oversees, although representation remains uneven. We need more voices beyond North America and Europe, fewer lawyers (as proud as I am to be one), more creators, AI annotators, youth, and everyday users from Quito to Dhaka. This geographic representation becomes especially urgent now that Meta has announced plans to use AI more to enforce its moderation rules, and rely less on third party human moderators. AI’s capabilities to match the billions of pieces of content posted everyday on the company’s platforms is undeniable. At the same time, serious questions remain on AI’s ability to understand culture, humor, and legitimate political criticism, particularly when expressed in languages that have provided less training data. This could exacerbate the existing and well-documented moderation inequities. And this is all the more pressing in light of a recent court judgment finding that Meta violated New Mexico’s Unfair Practices Act due to insufficient safety practices in relation to children.
The Board’s funding dependency is another vulnerability: Meta remains the sole contributor to the Independent Trust, through which the activities of the Oversight Board are funded. While that has not compromised the Board’s independence, sustainability matters. Companies whose products or services impose social costs—whether job displacement, psychological harm, or violence against women—should help fund the infrastructure that mitigates those costs. This could take the form of an industry-supported public interest fund, akin to the Universal Service Funds that expanded internet access or spectrum auctions that served broadcasting obligations.
Another limitation: the Board’s recommendations are not binding. But its implementation committee tracks how Meta responds and publishes an assessment, which has created public accountability. We should go further, and make those recommendations more accessible and actionable for stakeholders, including civil society, advertisers and investors. Concretely, this means closing the feedback loop with the civil society organizations, academics, and communities with lived experience, who help us understand the harms we adjudicate. When a recommendation is made, contributors deserve to know whether it was implemented, where and why it stalled. For advertisers and investors, the Board should have the standing to present its annual report directly to Meta’s board of directors and, by extension, to make that report available to shareholders ahead of annual general meetings. This idea of investor transparency seems to be validated by the market itself: in a recent study of investor risk and return among 52 U.S. listed companies, researchers from the TechForward Investors initiative found that companies with formalized accountability structures for societal risk consistently exhibited lower share price volatility.
Despite these imperfections, the Board offers a working model of independent oversight. Its case decisions are binding to the company. Meta deserves all the credit for doing that. While the Board and Meta do not always agree, this binding aspect has led Meta to correct its actions. For example, in a recent case involving an anti-scam awareness video posted by Taiwanese authorities, Meta initially took the video down, before reversing course when the Board overturned that decision.
In spite of its structural relationship to Meta, the Board’s commitment to independence is real. The Board is composed of highly accomplished individuals, who all cherish freedom of expression, including their own. I have used my voice and platform to denounce the choice by Meta to retreat from its content moderation efforts in the name of free speech, warning that this would ultimately lead to more censorship.
Our geographic diversity is our strength. As my colleague Nighat Dad of the Oversight Board and the Digital Rights Foundation in Pakistan, and others at the De-Center, often note, companies attentive to the periphery are those that thrive. The rise of Chinese AI developer DeepSeek across the Global South underscores this point. Our Board is composed of 21 members from all continents, although as I noted above, there are other dimensions of diversity that the Board should continue to grow.
Despite its imperfections, the Oversight Board offers a blueprint for AI accountability. It proves that independent oversight can coexist with innovation. It further shows that legitimacy and trust grow from transparency and participation. Other tech companies should take heed of both the promise and the shortcomings of Meta’s oversight experiment, and take proactive steps to do the following:
- Establish an external, independent oversight mechanism with binding decision-making authority. Internal ethics boards and safety committees are a start, but their decisions must carry real consequences.
- Ensure that the oversight body is appointed independently of the company it oversees. Self-selected accountability is a contradiction in terms. Recruitment should involve open, documented nominations from a global pool, with selection by the oversight body itself rather than by the company.
- Build in proactivity, not just reaction. An oversight mechanism that only responds to individual cases after harm has occurred is structurally too slow for AI. The body must have a mandate and resources to anticipate emerging risks and issue guidance before risks and harms scale.
- Prioritize diversity beyond geography. Global representation matters. So does professional diversity: engineers, behavioral scientists, public health specialists, and people with direct lived experience of platform harm must sit alongside lawyers, academics and journalists.
- Fund accountability. Company-funded oversight is better than none, but sustainable independence requires diversified funding. Funding can be sustained through an industry-wide public interest fund, or other mechanism that removes the financial dependence from any single company.
- Make implementation transparent and accessible. Publish not just decisions but their implementation, in plain language, in multiple languages, and in formats that civil society, investors, advertisers, and regulators can actually use.
- Treat accountability findings as material information. Board reports and implementation records should feed into environmental, social, and governance (ESG) disclosures, shareholder communications, and advertiser briefings.
In a functioning democracy, citizens do not fear who is in power because rules, not rulers, hold sway. The same principle should govern the future of AI.
