🚨 Yesterday, the 'White House Accord on Super Intelligence' was signed. What was supposed to be about AI safety and accountability was inst…

🚨 Yesterday, the "White House Accord on Super Intelligence" was signed. What was supposed to be about AI safety and accountability was instead a PROBLEMATIC endorsement of self-regulation. My comments: - The agreement is voluntary and not legally binding. Trump called it “morally binding.” In practice, nothing happens if a company decides not to comply. - The language used is vague, and it is unclear what successful compliance means. Terms like “robust internal controls,” “alignment,” and “operating as intended” are undefined and lack measurable thresholds. Vague agreements are destined to be useless. - An “independent external auditor” is not as independent as it seems. Companies can choose their evaluators; they pay them, and no official body oversees these evaluations. They are incentivized to choose evaluators aligned with their own goals. - An “independent committee of the board of directors” is still a committee within the company, with no real independence from the company’s ideology, plans, and strategy. - No transparency, no public oversight. The document says that auditors will assess controls and board committees will receive reports. What about the public? What if the evaluators and the board disagree? What if there are conflicts of interest or mistakes? We will never know (but we will still be subjected to the harms posed by their reckless practices). - The expression “behaves as intended” is effective when anchored in external standards, concrete goals, and specific outcomes, usually set out by legally binding rules and frameworks. The company will define “intended behavior” as whatever fits its needs. It does not mean that its technology is acceptably safe. - The agreement says that participating companies will meet regularly to “establish standards and best practices.” So companies will define for themselves what counts as AI safety, risk mitigation, and accountability. This is essentially self-regulation, and companies, as private entities created to generate profit, will naturally prioritize their own interests, not the interests of the public. - These companies compete against each other. If a company's evaluator says its AI model is too risky, they will ask for a second or third opinion until they get the desired outcome (releasing the model). They have no incentive to put themselves in a position to suffer economic losses. One way to change this incentive structure is through legally enacted rules that require all companies to comply with the same requirements and standards, with meaningful monetary penalties for noncompliance. - This agreement will not materially change anything. It endorses self-regulatory practices that are vague, not legally binding, and ineffective at meaningfully improving AI safety or holding companies accountable. Incidents such as the OpenAI/Hugging Face hack would still have happened under this accord. 👉 READ my full article (and join my newsletter's 101,200+ subscribers) below.

Source: Gary Marcus (X) | 2026-09-30

Related