OCOPD OpenAI, Anthropic and Google DeepMind explore common AI safety standards
AI safety standards are becoming a subject of direct coordination among three of the most prominent frontier-model developers: OpenAI, Anthropic and Google DeepMind. Representatives of the companies have been meeting on a working-group basis since at least July to discuss a shared industry standards body focused on testing and auditing advanced AI systems, according to reporting by TechCrunch and Yahoo News.
The discussions continued as recently as the previous week, but no binding agreement has been reached and industry-wide support is not yet in place. That leaves the effort at an early but consequential stage: the companies involved set much of the pace for the market in general-purpose AI models, coding agents and other systems with broad economic and security implications.
The central concept under discussion is a common framework for evaluating powerful models before public release. Possible elements include independent evaluations, pre-release safety reviews and standardized risk assessments. Such a body could give companies a shared vocabulary for identifying and disclosing risks, rather than relying solely on separate internal policies that can differ in scope, test methodology and publication practices.
A proposed U.S.-led body modeled on financial oversight
The talks were reportedly catalyzed by a July essay from Google DeepMind founder Demis Hassabis. He proposed a U.S.-led standards organization modeled on the Financial Industry Regulatory Authority, the Wall Street oversight body that operates under Securities and Exchange Commission supervision.
In that model, a public-private partnership would assess advanced models prior to their release. Industry would provide funding, while independent technical experts would carry out evaluations. The proposal points to an approach that is neither purely voluntary corporate self-governance nor a conventional government regulator.
OpenAI chief executive Sam Altman has expressed support for an AI testing and auditing body. Anthropic chief executive Dario Amodei has also publicly urged coordination around AI testing and auditing. The three-company conversations began before Amodei published a September 12 essay on those issues, indicating that the effort was already under way rather than being a direct response to his public intervention.
Competitive pressures complicate voluntary coordination
A shared standards body would face a basic commercial challenge: the companies are simultaneously partners in safety discussions and direct competitors in the race to develop and deploy more capable models. Common testing thresholds could improve comparability and potentially reduce pressure to shortcut safety work in pursuit of product launches. But participants would have to determine which model capabilities require review, who can inspect confidential systems and how findings would be communicated without exposing proprietary information.
The talks also raise antitrust questions. Coordinated standards among leading competitors can be useful when they improve interoperability or safety, but the companies would need to avoid arrangements that improperly constrain competition or amount to collective control over market access. Reports indicate those concerns remain unresolved.
For smaller AI developers and enterprise customers, a credible common regime could eventually lower uncertainty. Enterprises deploying AI in sensitive workflows often need clearer evidence about model behavior, misuse risk and the limits of a provider’s safeguards. Conversely, requirements designed principally by the largest labs could become de facto entry barriers if the testing process is expensive, opaque or inaccessible to competitors.
Policy backdrop: safety debate meets U.S.-China competition
The industry negotiations are taking place amid a sharper policy divide in Washington. TechCrunch reported that President Donald Trump’s team has played down safety concerns while emphasizing the need for the United States to maintain its pace against China. The companies’ discussions are continuing without a settled federal consensus on how advanced-model oversight should work.
That distinction matters. A voluntary standards body could influence future regulation by establishing testing practices that lawmakers or agencies later adopt. It could also remain a limited industry mechanism, with uneven participation and no authority to prevent a company from releasing a model.
The immediate significance is therefore not a completed safety pact, but the recognition by OpenAI, Anthropic and Google DeepMind that individual commitments may not be sufficient in a market where speed is a strategic advantage. Whether the talks produce a durable institution will depend on participation beyond the initial group, credible independence for evaluators, and agreement on consequences when a model fails a safety assessment.