OpenAI CEO Sam Altman reportedly told employees this week that the company could slow cutting-edge AI development with rival labs, raising a central decision: can competitors coordinate without giving an advantage to any company that keeps racing?

Bloomberg reported that Altman discussed the possibility during a companywide meeting and acknowledged that some AI labs may not agree to participate. OpenAI declined to comment, and no participating companies, proposed terms or timetable were reported.

A lab that slows work alone gives its rivals an opportunity to keep advancing their models, winning customers, recruiting researchers and strengthening their technical lead. For frontier developers, a coordinated system could determine which work must pause and when. For businesses building on their models, it could make release schedules, capability upgrades and access depend partly on shared safety reviews rather than individual providers’ plans.

Pacing would not necessarily mean stopping all AI development. It could temporarily restrict particular training, deployment or research activities when an agreed safety threshold is crossed, then allow the work to resume once shared safety requirements are met. As our recent coverage explained, this competitive pressure makes that kind of restraint difficult for any company or country to attempt alone because whoever slows could lose technological and commercial ground while its rivals continue advancing.

Coordinated pacing could reduce that risk, but only if the restrictions were shared and independently verifiable. However, no joint system has been announced yet. Altman’s reported willingness to coordinate therefore matters more when placed alongside the compatible public positions now coming from OpenAI’s chief scientist and Anthropic.

Coordinated AI pacing is a proposed system in which competing frontier AI developers would restrict specified training, deployment or research together after a shared safety threshold is crossed, then resume under agreed conditions.

OpenAI chief scientist Jakub Pachocki has now made the case for slowing development in public. In an essay published by OpenAI, he argued that no lab has solved alignment and monitoring well enough to continue advancing at maximum speed for much longer. Alignment concerns whether an AI continues to follow its intended goals and human values as it becomes more capable. Monitoring is the ability to recognize when it starts behaving in unexpected or dangerous ways.

Pachocki’s proposed response combines continued safety research with deliberate limits on development. He said he hopes voluntary slowdowns will become common until labs establish shared safety requirements. He also argued that company policies such as OpenAI’s Preparedness Framework and Anthropic’s Responsible Scaling Policy should become broadly required standards enforced by independent auditors, government agencies or international bodies. That would move critical safety decisions beyond each developer’s internal judgment and toward requirements that outsiders could check.

Anthropic has drawn a similar line between slowing inside one company and pacing development across the industry. Within a company, it describes pacing as choosing safety over speed when the two conflict. Across the field, pacing would mean creating a process that prevents competitive pressure from pushing every lab toward weaker safeguards.

Anthropic says that broader approach would require coordination between industry and government. It has called for a lawful, effective system that is clear enough for others to understand and strong enough for them to verify. That position goes beyond asking companies to exercise better judgment individually. It recognizes that competitors need a reliable way to know they are operating under comparable limits.

Taken together, these positions show real movement toward a common view: safety limits will be difficult to sustain unless they apply across competing labs. It is important to point out, however, that Pachocki has made a public argument, Altman’s remarks were reported from a private meeting and Anthropic has stated a formal company position. They have not announced a joint proposal, agreed on common safety thresholds or committed to the same restrictions.

For now, Altman and Pachocki are pointing toward coordinated pacing, while Anthropic has formally endorsed it. None has explained how a joint system would work. Any plan would need to give each lab confidence that its competitors would slow too.

Coordinated pacing could keep the burden from falling on one lab. But slowing would still carry a cost: delaying training or other development could postpone a more capable model, extend the time before it reaches customers and limit what researchers can learn from building it. If participating labs faced comparable restrictions at the same time, that cost would be spread across competitors instead of falling on the first company to act.

This is a problem even when several companies would prefer a safer overall pace. Each one still has an incentive to continue if it believes its competitors will do the same. The Pacing the Frontier statement describes this pressure at both the company and country level: neither wants to slow unilaterally while others keep advancing. Maybe more importantly, it also says the technical and governance tools needed to manage the pace of frontier development do not yet exist.

Altman’s reported acknowledgment that some labs may not participate shows why several AI companies sharing the same safety concern is not enough. The same problem extends beyond an agreement among U.S. developers: Chinese AI developers could keep advancing while U.S. participants paused. That possibility would make U.S. labs less willing to accept restrictions that their Chinese competitors did not share.

Even broad participation would solve only part of the problem. The bargain becomes enforceable only when the promise to slow is translated into rules that define what activates the restriction, what work must stop, who verifies compliance, what happens if a participant breaks the rules and what allows development to resume.

A coordinated slowdown cannot begin with labs simply promising to be careful. They would first need to agree on what should make them slow down. That trigger has to be defined before a problem appears so that each company cannot decide for itself whether a concerning result is serious enough to act on.

A 2023 paper from the Centre for the Governance of AI, written by Jide Alaga and Jonas Schuett, offers one example of how this could work. Frontier models would be tested for dangerous capabilities before they advance further. If testing showed that a model had developed a dangerous capability covered by the agreement, its developer would notify the other participants and pause the work covered by the agreement. The other labs would then place comparable restrictions on their own development.

The pause would give developers time to study the capability and put stronger safeguards in place. Before work resumed, they would also need to agree on what evidence would show those safeguards were strong enough, so every participant could restart under the same conditions.

The rules would also have to specify what each lab must restrict. Depending on the risk, that could include further training, deployment of similar models or publication of research that could help others reproduce the capability. If one company stopped training while another restricted only a narrower activity, the burden would no longer be shared equally.

The agreement would also need to address open-weight models, which are released with files that allow outside developers to run or modify them on their own systems. Once those weights had been distributed, the original company could not enforce a later pause by restricting access because copies of the model would already be operating beyond its control.

The proposal describes several possible ways to oversee the system. Labs could make voluntary public commitments, sign agreements with one another, use the same independent auditor or operate under legal requirements enforced by a regulator.

The industry was already developing other pieces of this system at the time. Anthropic released the first version of its Responsible Scaling Policy in September 2023, including safety thresholds that could require the company to pause its own training or deployment. Two months earlier, Anthropic, Google, Microsoft and OpenAI had created the Frontier Model Forum to share safety research and develop common practices. Together, those efforts established a company-level safety policy at Anthropic and a place for major labs to cooperate, but neither required competitors to slow together.

Pachocki’s call for shared safety requirements enforced by auditors, governments or international bodies fits parts of this general approach. So does Anthropic’s support for a coordinated system that is lawful and verifiable. However, neither OpenAI nor Anthropic has adopted this proposal or said which structure it would support.

This model shows how a safety finding at one lab could lead to a shared response instead of leaving that company to slow alone. The harder part would be persuading competitors to place their own decisions under the same system.

The industry had a possible blueprint for coordinated pacing in 2023, but not a working agreement. Making that blueprint binding would require labs to give up some control over decisions they now make independently. They would have to accept common tests and thresholds and an outside authority able to enforce a pause. None of those pieces is in place. What has changed is that coordinated pacing now has support from Anthropic as a company and senior leaders inside OpenAI, including Altman’s reported willingness to consider slowing with other labs.

Participation is the first problem. An agreement among several U.S. companies could reduce the pressure on those labs to keep racing, but it could not require developers elsewhere to slow down. There is no evidence that Chinese AI companies would reject a coordinated system, but there is also no indication that they have agreed to join one. Governments would therefore need to help extend the agreement beyond the companies and countries initially willing to participate.

The labs that joined would also need to prove they were following the rules. An independent reviewer would need enough access to evaluate whether a safety threshold had been crossed and confirm that the required work had stopped. But model designs, development plans and some safety tests can contain sensitive information. Without sufficient access, an auditor could not verify compliance. Providing that access would require companies to trust the auditor with information they ordinarily protect from competitors and the public.

Applying equivalent restrictions could create another problem. AI labs use different models, development methods and release strategies, so the same rule may not affect each company in the same way. A restriction that stops a major training run at one lab might interrupt a different type of work at another. The agreement would need to close obvious loopholes while still accounting for those differences.

The system would only be as reliable as the tests used to trigger it. Alaga and Schuett identify the lack of reliable tests for dangerous capabilities as one of the proposal’s main obstacles. A test might miss a dangerous behavior or appear reassuring while the underlying risk remained. If labs disagreed over what a result meant, the system could stall just when a shared response was needed.

Enforcement would require government involvement as well. Governments have already negotiated international agreements that place shared limits on dangerous activities and create ways to verify compliance. Nuclear arms agreements, for example, use inspections, shared information and monitoring to check whether countries are following the rules. AI pacing would need the same basic structure, but it would have to reach private companies and evaluate fast-changing software capabilities rather than count physical weapons or detect nuclear tests. A government-backed framework could create the legal structure labs would need to slow together while addressing the coordinated-pausing paper’s concern about competitors making that agreement entirely on their own.

Voluntary promises would provide the least assurance because they would depend on companies disclosing problems, pressure on them to follow through or whistleblowers revealing that the rules had been broken. Contracts, independent audits or legal requirements could make compliance easier to check and enforce, but they would also require institutions with greater authority over the participating labs.

The 2023 proposal shows that the delay was not caused by a complete absence of ideas. The missing step was connecting company safety policies and a forum for cooperation to a rule that would require other participants to respond when one lab crossed a shared threshold.

An earlier system might have given the industry a way to respond collectively as concerns grew, but we cannot know whether it would have prevented the current situation. The more immediate problem is the one Altman reportedly acknowledged: coordinated pacing cannot relieve the pressure to keep racing if a major developer refuses to join. The current public statements do not identify the participants or explain how such a holdout would be handled. The idea will begin to move forward only when companies say publicly what they are willing to commit to.

Anthropic says it will explain in the coming weeks how it intends to contribute to coordinated pacing. That explanation could begin answering questions left unresolved by the current public statements, including who would participate, which safety findings would trigger a response and who would verify that every lab followed the resulting restrictions.

Signs that this support was becoming a concrete proposal would include named participants, shared triggers, comparable restrictions and an outside body with enough access to verify compliance.

Until those commitments appear, coordinated pacing will remain an approach formally endorsed by Anthropic and supported by senior leaders inside OpenAI rather than a joint system that could change how the companies operate. Coordinated pacing is possible in principle. Whether it works in practice depends on whether labs—and eventually governments—will accept shared restrictions and outside scrutiny before competitive pressure makes restraint even harder.

Coordinated AI pacing could give labs time to address a dangerous capability by requiring participating developers to slow related work together. That matters because current and former researchers at leading labs warn that future systems could become impossible to control and threaten humanity. As our recent coverage explained, researchers with experience inside Anthropic have placed that danger within time frames ranging from the next few years to the next decade. Those are personal assessments, not established forecasts. But they explain why Altman’s reported willingness to consider slowing with rivals, Pachocki’s public call for shared safety requirements and Anthropic’s formal support carry unusual urgency. No lab has demonstrated that it can reliably control the much more capable systems the industry is trying to build.

Coordination among U.S. labs would address only the competition between those companies. If leading Chinese developers continued advancing outside the agreement, American labs would still fear that slowing could cost the United States its technological lead. There is no evidence that Chinese developers have rejected coordinated pacing, but there is also no indication that they have agreed to participate. A coordinated slowdown would therefore need China and other major AI developers to participate because an agreement among U.S. labs could not resolve the global competition that makes slowing difficult.

For businesses building on frontier models, an agreement could turn a serious safety finding at one lab into delays or restrictions across several providers. Using multiple vendors might not prevent disruption if they were bound by the same rules. No agreement is currently in place, so businesses do not need to postpone AI projects because of these discussions. They should, however, be cautious about plans that depend on an expected frontier model arriving on a specific schedule or a particular capability remaining available.

The tradeoff would be slower access to some capabilities in exchange for a stronger response to risks that could extend far beyond any company or product. Coordinated pacing would not guarantee safety, but it could give developers time to respond to warning signs without requiring one lab or country to fall behind alone. If researchers are right that future AI could escape human control, the danger may become clear only when it is too late to slow down.

Q: What did Sam Altman reportedly say about slowing AI development?
A: Bloomberg reported that Sam Altman told OpenAI employees the company could slow cutting-edge AI development with rival labs. Altman also reportedly acknowledged that some AI labs might refuse to participate, and no companies, terms or timetable have been announced.

Q: Are OpenAI and Anthropic already planning to slow AI development?
A: No. OpenAI chief scientist Jakub Pachocki has publicly called for shared safety requirements, Sam Altman has reportedly expressed willingness to coordinate with rivals and Anthropic has formally endorsed coordinated pacing, but the companies have not announced a joint agreement.

Q: What does coordinated AI pacing mean?
A: Coordinated AI pacing is a proposed system in which competing frontier AI developers would restrict specified training, deployment or research together when a shared safety threshold is crossed. The restrictions would remain in place until participants agreed that sufficient safeguards had been established.

Q: Why do some AI leaders want labs to slow down?
A: Some AI leaders argue that alignment and monitoring have not advanced enough to keep increasingly capable systems under reliable human control. A coordinated slowdown could give developers more time to investigate dangerous capabilities and strengthen safeguards without requiring one lab to fall behind its competitors.

Q: Why would an AI slowdown need China to join?
A: Major Chinese AI developers would need to participate because U.S. labs could lose technological and commercial ground if they slowed while Chinese competitors continued advancing. There is no evidence that Chinese developers have rejected coordinated pacing, but there is also no indication that they have agreed to join.

Q: How would a coordinated AI slowdown actually work?
A: Participating labs would need shared tests and safety thresholds that determine when specified work must stop. The system would also require comparable restrictions, independent verification, enforcement and agreed evidence for deciding when development could resume.

Q: Could an AI slowdown delay new models or disrupt businesses?
A: A coordinated slowdown could turn a serious safety finding at one lab into delays or restrictions across several providers, so using multiple vendors might not prevent disruption. No agreement currently exists, but businesses should avoid plans that depend on a frontier model arriving on a fixed schedule or a particular capability remaining available.

Q: Would slowing AI development actually make it safer?
A: Slowing AI development would not guarantee safety, but it could give developers more time to study dangerous capabilities and strengthen safeguards. Its effectiveness would depend on broad participation, reliable safety tests, independent verification and enforceable restrictions.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing support, AEO/GEO/SEO optimization, image concept development, and editorial structuring support from ChatGPT, an AI assistant. All final editorial decisions, perspectives, and publishing choices were made by Alicia Shapiro.