Last week, more than 200 leading researchers and policy figures urged artificial intelligence (AI) companies to invite independent and outside evaluators into their firms to assess and report on risks. They did so less than a week after the CEOs of the world’s leading AI companies near-simultaneously endorsed a pause in AI development and endorsed the idea of bringing in evaluators. “We believe that all frontier AI companies should embed evaluators to independently assess AI risks,” they wrote in an open letter, “including evaluating the systems themselves and any significant incidents of real-world harm.”
But the letter’s phrase “any significant incidents of real-world harm” does not say what counts as harm, and could be used to justify government censorship or, worse, the manipulation of AI models to generate ideological content to users rather than serve as truth-seeking devices.
And now, two of the letter signers, deeply alarmed by the risks of AI, have gone on the record in exclusive interviews with Public to warn that AI regulation could lead to “state control” and “ideological domination.”
At the same time, Public has identified at least four letter signers who worked within the censorship industrial complex and advocated greater censorship. This conflict within the AI evaluator community could become more important as governments around the world begin to regulate AI.
The push for AI safety comes at a moment when Barack Obama and liberal-Left leaders in Britain, France, Germany, Australia, and the European Union have called for AI regulation and, at the same time, issued new calls to “fight misinformation” and “hate speech” through new censorship laws.
Already, the European Commission designated ChatGPT a very large online search engine under the Digital Services Act (DSA) on August 31, which gives OpenAI until January 2027 to begin assessing and mitigating systemic risks including those to “electoral processes and public security,” both of which governments, NGOs, and social media firms have used as justification for censoring legal speech.
Britain’s government promised in February to “force all AI chatbot providers to abide by illegal content duties in the Online Safety Act or face the consequences of breaking the law.” The law allows censorship based on “the risks of harm to individuals in the United Kingdom presented by illegal AI-generated content.”
Australia’s government has introduced new censorship legislation that would cover any service that lets users “generate material by means of artificial intelligence”
Ideological control of the models is not the same as taking down content. But it could be even more pernicious, directing users away from and toward particular content and framings, without them knowing it.
The two letter signers who are outspoken against censorship have significant credibility within the AI evaluator community. They have both done high-level work, one for the US government and the other for Anthropic, which is currently leading the AI race, on risk, safety, and regulation.