OpenAI Updates Safety Framework, Downgrades Mass Manipulation and Disinformation from Critical Risk Status

In a significant policy shift, OpenAI announced it will no longer assess its artificial intelligence models before release for risks related to persuasion or manipulation capabilities—risks that could potentially influence elections or enable sophisticated propaganda campaigns.

The company revealed this change in an update to its “Preparedness Framework” published yesterday. Instead of pre-release evaluation for these risks, OpenAI will rely on its terms of service to restrict AI usage in political campaigns and lobbying, while monitoring how users interact with its models after deployment.

This marks a substantial departure from previous policy. OpenAI has also revised its stance on high-risk models, indicating willingness to release AI systems it judges to present “high risk” provided the company has implemented appropriate safeguards. Even more surprisingly, the framework now states OpenAI might consider releasing a model presenting “critical risk” if a competing AI lab has already released similar technology.

Previously, the company had maintained it would not release any AI model presenting more than “medium risk.”

The updated framework has divided AI safety and security experts. Some have praised OpenAI for voluntarily publishing its revised guidelines, noting improvements such as clearer risk categories and stronger emphasis on emerging threats like autonomous replication and safeguard evasion.

Others expressed significant concerns. Steven Adler, a former OpenAI safety researcher, criticized the framework for no longer requiring safety tests of fine-tuned models. “OpenAI is quietly reducing its safety commitments,” he wrote on X, though he acknowledged appreciating the update overall.

The removal of persuasion from the Preparedness Framework’s danger categories sparked particular alarm among experts. Shyam Krishna, research leader in AI policy and governance at RAND Europe, noted, “OpenAI appears to be shifting its approach. Instead of treating persuasion as a core risk category, it may now be addressed either as a higher-level societal and regulatory issue or integrated into OpenAI’s existing guidelines.”

Courtney Radsch, a senior fellow at several prominent policy institutions, called the framework “another example of the technology sector’s hubris,” arguing that downgrading persuasion concerns “ignores context – for example, persuasion may be existentially dangerous to individuals such as children or those with low AI literacy or in authoritarian states and societies.”

Oren Etzioni, former CEO of the Allen Institute for AI, questioned OpenAI’s motivations: “Downgrading deception strikes me as a mistake given the increasing persuasive power of LLMs. One has to wonder whether OpenAI is simply focused on chasing revenues with minimal regard for societal impact.”

Not all experts were critical. One AI safety researcher, speaking anonymously, suggested addressing disinformation risks through terms of service might be reasonable, noting that persuasion risks are difficult to evaluate pre-deployment and more ambiguous compared to other dangers like facilitating biological weapons development or cyberattacks.

The policy change parallels recent developments in European regulation. Some Members of the European Parliament have raised concerns that the latest draft of the proposed code of practice for EU AI Act compliance similarly downgraded mandatory testing for disinformation risks to a voluntary consideration.

Research has shown AI chatbots can be highly persuasive, though this capability isn’t inherently dangerous. Cornell University and MIT researchers found that dialogue with chatbots effectively helped people question conspiracy theories.

Another controversial aspect of the updated framework is OpenAI’s statement that “If another frontier AI developer releases a high-risk system without comparable safeguards, we may adjust our requirements.” Critics view this as signaling a potential race to the bottom in safety standards.

Max Tegmark, president of the Future of Life Institute, warned, “The race to the bottom is speeding up. These companies are openly racing to build uncontrollable artificial general intelligence—smarter-than-human AI systems designed to replace humans—despite admitting the massive risks this poses.”

AI critic Gary Marcus interpreted the statement as revealing that “what really governs their decisions is competitive pressure—not safety. Little by little, they’ve been eroding everything they once promised.”

Miranda Bogen, director at the Center for Democracy & Technology’s AI governance lab, acknowledged the value of transparency in risk management practices while expressing concern about “moving the goalposts” as risks become more apparent. She also criticized the framework’s focus on ‘frontier’ models, noting companies have used narrow technical definitions to avoid publishing safety evaluations for powerful recent models.

OpenAI released its GPT-4.1 model yesterday without a safety report, claiming it was not a frontier model, underscoring Bogen’s point that “voluntary commitments only go so far” in ensuring responsible AI development.