Should AI Companies Coordinate to Slow Down AI Development?

Quick Answer

• A blanket slowdown across the entire AI industry would be difficult to define, coordinate, and enforce.
• There is a stronger case for companies to coordinate around specific high-risk capabilities and shared safety thresholds.
• Major frontier AI developers already participate in joint safety initiatives involving evaluations, information sharing, and risk-management frameworks.
• A realistic system would allow ordinary development to continue while permitting delays when predefined risk thresholds are crossed.

Calls to “slow down AI” sound simple until someone has to define what slowing down actually means. Should companies stop training larger models? Delay public releases? Restrict certain capabilities? Pause research entirely? The phrase can describe very different policies.

A more useful debate focuses on coordination. Frontier AI companies face some risks that no individual company can manage well if competitors are operating under completely different standards. At the same time, freezing broad areas of AI research could delay useful advances without necessarily stopping the most dangerous development elsewhere.

The practical question is therefore whether companies can establish shared conditions under which development or deployment should temporarily slow, while allowing lower-risk research and applications to continue.

1. Why Some AI Risks Create a Case for Coordination

Coordination becomes most relevant when frontier models develop capabilities that could create serious public-safety or security risks beyond the company operating them.

AI companies normally have powerful incentives to develop better products quickly. Competition attracts customers, investors, researchers, and computing resources. That system works reasonably well when the consequences of moving too quickly mainly fall on the company itself.

Frontier AI creates a different problem when a new capability could increase risks involving advanced cyberattacks, biological or chemical misuse, highly autonomous systems, or theft of powerful model weights. The potential consequences may extend far beyond the developer that created the model.

That creates a coordination problem. One laboratory may want to conduct extensive safety testing, but if competitors release comparable capabilities earlier, the cautious company pays the commercial cost of restraint while everyone may still face the external risk.

Shared safety standards can reduce that pressure by giving multiple developers similar expectations about evaluations, safeguards, incident reporting, and circumstances that could justify delaying deployment.

2. Why a Blanket AI Slowdown Would Be Difficult

Not all AI development carries the same risk. A universal slowdown would have to distinguish dangerous frontier capabilities from ordinary research, efficiency improvements, medical tools, accessibility systems, and countless other applications.

“AI development” covers an enormous range of activity. Improving a small language model that runs locally on a laptop is not necessarily comparable to training a frontier system that develops substantially stronger capabilities in cybersecurity or autonomous research.

A broad industry pause would also face an enforcement problem. Participation could be voluntary for some companies while other developers, open-source groups, startups, or organizations in other jurisdictions continued working. The companies that complied most strictly could lose competitive ground without guaranteeing that overall technological progress slowed.

There is also an opportunity cost. AI systems can contribute to scientific research, software development, accessibility, education, healthcare research, cybersecurity defense, and productivity. Delaying every form of development treats potential benefits and potential hazards as if they were identical.

That is why much of the emerging safety architecture focuses on risk-based thresholds rather than a permanent industry-wide speed limit.

3. AI Companies Are Already Coordinating on Safety

The industry has already moved beyond completely independent safety programs. Major frontier developers participate in shared initiatives covering evaluations, safety frameworks, security practices, and information sharing.

The Frontier Model Forum brings together major developers including Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI. Its work includes identifying frontier AI safety and security practices, supporting research, developing shared understanding of risk evaluations, and facilitating information exchange.

In March 2025, forum members entered an information-sharing agreement covering safety-relevant vulnerabilities, threats, and capabilities of concern. The initiative has since developed mechanisms for exchanging information while addressing intellectual-property and antitrust considerations.

International coordination has developed in parallel. The 2024 AI Seoul Summit produced voluntary Frontier AI Safety Commitments from numerous major technology companies. Those commitments centered on safety frameworks for severe risks and included the principle that companies should identify thresholds at which risks could become intolerable unless sufficient mitigation is available.

This is coordination, but it is not an agreement to make AI generally advance more slowly. It is closer to building common traffic rules before discovering that every laboratory brought its own interpretation of what a red light means.

4. A Risk-Threshold System Could Be More Practical Than a General Pause

Instead of slowing every project, companies can define measurable capability or risk thresholds that trigger stronger safeguards, restricted deployment, additional testing, or temporary delays.

Frontier safety frameworks are designed around this idea. Developers evaluate models for capabilities associated with severe risks, define thresholds in advance, and connect those thresholds to specific mitigation requirements.

A model below a defined threshold might proceed through ordinary development and deployment. A model approaching a more serious threshold could require stronger access controls, additional evaluations, improved cybersecurity, or restricted availability. At still higher risk levels, development or deployment could be delayed until adequate safeguards exist.

This approach has an important advantage over an arbitrary calendar-based pause: the restriction is connected to what the system can actually do. It also creates a framework that can evolve as evaluation methods improve.

The weakness is measurement. Frontier capabilities can be difficult to evaluate reliably, and developers may disagree about thresholds, testing methods, or whether particular mitigations are sufficient. A risk-based system therefore depends heavily on credible evaluations and some degree of external scrutiny.

5. Coordination Needs More Than Voluntary Promises

Shared commitments are useful only if companies can verify risks, communicate incidents, update safeguards, and respond when competitive incentives push against caution.

The hardest coordination problem is not writing principles. It is deciding what happens when a laboratory discovers a dangerous capability shortly before a major product release while competitors are preparing their own models.

Information sharing can help companies identify vulnerabilities and emerging threats earlier. Common evaluation methods can make safety decisions more comparable. Independent research and external testing can reduce reliance on companies evaluating themselves. Incident-reporting systems can also help the industry learn from failures rather than allowing the same weakness to appear repeatedly.

Coordination also has to avoid becoming a mechanism for established companies to block smaller competitors. Safety rules should focus on demonstrable risks and capabilities rather than simply imposing requirements that only the richest laboratories can afford.

The most credible model is therefore neither unrestricted competition nor an indefinite universal pause. It is a system in which developers compete normally but operate under increasingly demanding safeguards as measurable risks rise.

Key Takeaways at a Glance

• A universal AI slowdown would struggle with definitions, enforcement, international competition, and widely different levels of risk.
• Coordination is more practical when focused on frontier capabilities that could create severe public-safety or security risks.
• Major AI developers already share some safety information and participate in common frontier-safety initiatives.
• Predefined risk thresholds can provide a reason to delay a particular model without freezing unrelated AI development.
• Effective coordination depends on reliable evaluations, information sharing, accountability, and safeguards against misuse of safety rules.

Approach	Main Advantage	Main Problem
Industry-Wide Pause	Creates time for safety work	Difficult to define and enforce globally
Shared Evaluations	Makes risks more comparable	Evaluation science is still developing
Risk Thresholds	Targets restrictions at dangerous capabilities	Companies may disagree about thresholds
Information Sharing	Spreads safety lessons and threat intelligence	Requires trust and protection of sensitive information
Independent Testing	Adds external scrutiny	Access and testing standards must be established

The Better Question Is When AI Development Should Slow Down

The debate does not have to be reduced to “accelerate everything” versus “stop AI.” Those are convenient slogans, which is precisely why humans keep dragging complicated technical problems toward them.

A more precise framework separates ordinary AI progress from capabilities that could create severe risks. Companies can continue competing while agreeing on evaluations, security practices, incident reporting, information sharing, and thresholds that require stronger safeguards.

Under that model, slowing down is not the objective itself. It becomes one possible safety response when evidence shows that a particular system is advancing faster than the available protections can manage.

Sources

Frontier Model Forum • Information Sharing, Incident Reporting, and Incident Response for Frontier AI Risks

Frontier Model Forum • Membership and Frontier AI Safety Requirements

UK Department for Science, Innovation and Technology • Frontier AI Safety Commitments, AI Seoul Summit 2024

UK Government • The Bletchley Declaration

Frontier Model Forum • Frontier Mitigations