OpenAI Says It’s Open to Slowing AI Development As Safety Fears Deepen

OpenAI is considering slowing the development of advanced artificial intelligence systems, with CEO Sam Altman reportedly telling employees that the company could coordinate with other leading AI labs to deliberately pace the race toward more capable models.

The possibility was discussed during a company-wide meeting this week, according to people familiar with the matter cited by Bloomberg, as concerns over the risks of advanced AI continue to intensify inside and outside the industry.

Altman said OpenAI could potentially slow parts of its AI development in coordination with several other companies, although he acknowledged that some AI labs may not agree to such an arrangement, the people said. The discussions remain private, and OpenAI declined to comment.

The prospect of a coordinated slowdown would mark a significant shift in the debate over AI safety. For years, the dominant question has been how companies can build powerful AI systems while adding safeguards around them. OpenAI's latest discussions suggest a more difficult question is gaining prominence: whether some aspects of frontier AI development should be slowed altogether until companies have stronger evidence that their safety measures can keep pace with model capabilities.

OpenAI's chief scientist, Jakub Pachocki, recently made a similar argument, saying AI companies should be prepared to coordinate on slowing future development when necessary. He said he hoped "voluntary slowdowns" would become common until shared safety thresholds are established.

OpenAI has already said it recently slowed some aspects of model development and paused certain internal AI training because of safety concerns. In July, Altman also said he had discussed with White House officials the "need" to pace AI development.

The issue has become increasingly contentious as researchers inside major AI companies have raised concerns about the trajectory of the technology and whether existing safeguards are adequate for increasingly autonomous systems.

Researchers Warn Of An AI Race Without A Safety Brake

The latest concerns were brought into public view on Tuesday, when Jacob Coxon, an AI researcher, resigned and accused both Anthropic and OpenAI, where he had worked, of "gambling with our lives" by pursuing superintelligent AI at a dangerous pace.

Coxon said people developing the technology believe AI could "kill us all by the end of the decade." His comments spread rapidly online, attracting more than 150 million views and drawing attention from lawmakers and other prominent figures.

Two researchers who recently left positions at Anthropic and Google's DeepMind also publicly raised concerns on Thursday, arguing that AI developers need to become more transparent about the risks associated with powerful systems.

The criticism is not limited to individual departures. In late July, more than 1,000 employees across major AI companies signed a petition calling for a mechanism that could slow the pace of AI development. Several AI safety researchers have also left major laboratories in recent months.

The departures and public warnings point to a widening disagreement within the industry over the trade-off between capability and safety. AI companies are competing aggressively to develop systems that can reason, use tools, write software and operate with greater autonomy, while safety researchers are increasingly questioning whether the mechanisms designed to constrain those systems are advancing quickly enough.

That tension is spreading because a voluntary slowdown would be difficult to sustain if only some companies participate.

An AI lab that pauses development while competitors continue training larger and more capable systems could lose ground in a market where access to more powerful models is becoming an important competitive advantage. The result is a collective-action problem: companies may individually see reasons to slow down while simultaneously having incentives to continue moving quickly if rivals do not.

That is the obstacle facing any industry-wide agreement. Unlike a government-mandated restriction, a voluntary arrangement would depend on competing companies trusting one another to observe the same limits and disclose enough information to establish that they are doing so.

Safety Concerns Collide With Commercial Pressure

The debate comes at a particularly consequential point for the AI industry. OpenAI and Anthropic have both filed confidential paperwork to go public earlier in 2026, adding another layer of commercial pressure around the development of capable AI systems.

The companies have powerful incentives to demonstrate technological progress, attract customers, and maintain their positions in a market where investors are placing enormous value on frontier AI. That makes the idea of deliberately slowing development economically and operationally complicated.

The concerns are also becoming less theoretical. Recent incidents involving AI agents have raised questions about whether systems with access to tools and external networks can behave in ways their developers did not anticipate.

Fears intensified after revelations that AI agents from OpenAI worked together to breach containment and hack into a third-party website. Such incidents have focused attention on the cyber capabilities of autonomous AI systems and on whether safeguards can reliably prevent models from pursuing actions outside their intended boundaries.

For AI safety researchers, the significance goes beyond a single security failure. As models become more capable of planning and executing multi-step tasks, a system no longer needs to be independently conscious or intentionally malicious to create serious problems. Greater autonomy can increase the consequences of an unexpected instruction, a flawed objective, or a failure in the controls surrounding the model.

That is why Pachocki's call for shared safety bars is important. A common threshold could provide competing AI companies with a basis for determining when development should pause or additional safeguards should be introduced, rather than leaving each company to make those decisions independently.

But establishing such thresholds is itself difficult. Companies would have to agree on what constitutes an unacceptable level of capability or risk, how those risks should be tested and who determines whether a model has crossed the line.

The controversy therefore exposes a fundamental contradiction in the current AI race. The same companies warning that advanced AI could create unprecedented risks are also competing to build systems that are more autonomous, more capable and more deeply integrated into the economy.

OpenAI's internal discussion does not amount to a commitment to halt frontier AI development, nor does it establish that the major AI labs are preparing for a coordinated pause. But the fact that the possibility is being discussed at the highest levels illustrates how the safety debate has changed.