OpenAI, Anthropic and Google DeepMind Confirm Weeks of AI Safety Talks

OpenAI has been discussing AI safety with rivals Anthropic and Google DeepMind for several weeks, the company’s global policy chief Chris Lehane told reporters on Tuesday. The disclosure confirms that the leading frontier-model companies have been coordinating privately as policymakers debate how to manage increasingly capable systems.

The talks come amid a widening split in Washington. Anthropic CEO Dario Amodei has urged the industry to slow the pace of frontier AI development to reduce catastrophic risks, while President Donald Trump and his AI adviser David Sacks have dismissed those concerns and argued that tighter restrictions could weaken the United States against China. For AI builders and enterprise buyers, the dispute could determine how model evaluations, deployment controls and independent oversight develop over the next phase of the market.

Why the talks surfaced now

Lehane’s comments, reported by TechCrunch after earlier coverage from Bloomberg, followed a public essay from Amodei calling for closer cooperation among AI companies. OpenAI CEO Sam Altman, Google DeepMind CEO Demis Hassabis and SpaceXAI’s Elon Musk were among industry figures reported to have supported the call to action.

Altman has also said OpenAI would work with Anthropic to embed third-party evaluators in its operations. Those evaluators would be intended to monitor whether advanced models are being developed and deployed safely, although the available reporting does not specify the scope, authority or operating standards of the proposed reviews.

The private discussions had been hinted at before OpenAI’s confirmation. TechCrunch reported that Altman recently acknowledged private conversations with other AI leaders. The Information separately reported that OpenAI, Anthropic and Google DeepMind were working toward an industry standards body, though that account said Altman told OpenAI staff such an effort might need to proceed without U.S. government support.

The talks therefore appear to involve more than informal executive contact. They may include shared safety practices, external evaluation and a possible mechanism for setting standards across competing labs. The companies have not publicly released a detailed framework, and TechCrunch said it contacted OpenAI, Anthropic and Google for comment.

Safety coordination meets antitrust concerns

Cooperation among direct competitors creates a legal and commercial complication. Altman and others have acknowledged that joint discussions could raise antitrust concerns if regulators conclude that safety coordination is being used to limit competition, slow product releases or divide markets.

Amodei’s proposal included a narrow government waiver that would protect certain safety-related coordination. According to TechCrunch, Lehane said the companies do not need such a waiver. That position leaves open a central question: what safeguards can the labs discuss together without crossing into conduct that could be interpreted as suppressing competition?

The distinction matters because the companies are not merely academic research groups. OpenAI, Anthropic and Google DeepMind compete for developers, enterprise contracts, infrastructure capacity and influence over the standards that may govern advanced models. A common testing protocol could improve comparability and reduce duplicated work, but it could also give the largest labs disproportionate control over the rules applied to the rest of the industry.

For companies building on top of these models, the outcome could affect documentation, audit requirements and access to higher-risk capabilities. A voluntary industry process may move faster than legislation, but it could also lack consistent enforcement or independent authority.

Evidence, claims and the policy divide

The strongest confirmed fact in the reporting is Lehane’s statement that OpenAI has been in safety discussions with Anthropic and Google DeepMind for weeks. The broader standards-body effort remains based on reporting by The Information rather than a public announcement from the companies.

The benefits of third-party evaluation are also prospective rather than demonstrated in the source material. No independent results, evaluator names, testing criteria or enforcement process were provided. References to catastrophic risk reflect the position of Amodei and other AI-safety advocates, not an established measurement showing that a specific model or deployment has reached such a threshold.

Hassabis has previously called for the United States to establish a new standards body with the ability to screen advanced models and coordinate industrywide slowdowns if risks increase. That proposal contrasts with the position attributed to Trump and Sacks, who have characterized existential-risk warnings as exaggerated or politically motivated and emphasized competition with China.

The policy conflict gives the private talks unusual significance. If the U.S. government is unwilling to create a formal oversight mechanism, the frontier labs may try to build one themselves. But a private body led by the companies being evaluated would face questions about independence, transparency and conflicts of interest.

What the talks could mean for AI builders

For model developers, the immediate implication is a possible shift toward shared evaluation practices. Common tests could make it easier to compare systems on dangerous capabilities, monitor changes between model versions and identify when a deployment needs stronger controls. They could also create new compliance costs, particularly for smaller firms that lack the safety teams and infrastructure of the largest labs.

For enterprise product teams, independent verification could become part of vendor diligence. Buyers may eventually want evidence that a model has undergone external testing before permitting it to handle sensitive data, operate tools or make decisions in regulated workflows. However, there is no confirmed timetable for such requirements, and no agreement described in the reporting has yet established a common certification.

The commercial stakes are equally direct. A slowdown coordinated around safety findings could affect release schedules, API availability and the pace at which developers receive new capabilities. Conversely, if companies are unable to coordinate because of antitrust risk, each lab may continue setting its own thresholds and evaluation methods. That could preserve competition but leave customers with inconsistent information about model risk.

The issue also extends beyond the three firms. An industry standards body would need to account for open-model developers, cloud providers, independent researchers and companies deploying models in specialized settings. Rules created only by the biggest frontier labs could be difficult for the wider ecosystem to adopt.

What to watch next

The next concrete signal will be whether OpenAI, Anthropic or Google DeepMind publicly identifies the participants, scope and governance of the safety talks. Details about third-party evaluators—including who appoints them, what they can inspect and whether their findings will be published—would distinguish a substantive oversight plan from a broad commitment.

Lawmakers’ response to the FRONTIER Act will also matter. Lehane said OpenAI supports a provision requiring leading frontier labs to admit “independent verification organizations” into their companies. Legislative language, enforcement powers and the definition of a covered lab could determine whether the proposal becomes a real obligation or remains a policy position.

Finally, developers should watch for a formal standards-body announcement, shared evaluation benchmarks and evidence that the process includes organizations outside the three largest participants. Without those signals, the talks remain an important disclosure about coordination, but not yet an operational safety regime.

Creati.ai perspective

The significance of this story is not simply that three AI companies are talking. It is that the companies appear to recognize a need for common safety processes while operating in a market where cooperation can create legal and competitive risks. That tension will shape whether oversight emerges as legislation, voluntary standards or a patchwork of company-specific practices.

For builders and enterprise buyers, the practical lesson is to treat safety commitments as incomplete until they produce inspectable evidence: defined tests, independent access, published findings and clear consequences for failure. The talks may be a useful starting point, but their credibility will depend on whether the participants can make oversight genuinely independent without turning a safety forum into a gatekeeping mechanism for the AI market.