Slowing the pace of advanced AI development is necessary to reduce its safety risks

What's this about?

People disagree about whether we need to slow AI work to make it safer.

The key question is whether slowing all AI work is needed, or if other steps can help.

What supporters say

• Strong AI can help with cyber attacks and may trick people, creating real risks.
• AI skills may grow faster than safety tests can check, leaving rule makers unready.

What critics say

• We may lower risk with focused rules, such as watching who gets huge computer power.
• More AI work may help make tools that spot threats and defend people.

How to read this

The number of points on each side does not show who is right; strong proof matters more.

The bottom line

The risks seem real, but the facts do not show that all AI work must slow down.

The best path may use more tests, watchful checks, and focused rules instead of one broad stop.

The claim that advanced AI development must be slowed to make it safer rests on a precautionary argument: powerful systems may be advancing faster than researchers and regulators can understand or control them. But the available evidence does not show that slowing all development is necessary.

The case for

The strongest argument for slowing is that AI capabilities may be moving faster than safety testing and government oversight. The International AI Safety Report describes rapid progress, uncertainty about dangerous abilities and important weaknesses in today’s evaluations. Giving researchers and regulators more time could reduce the chance that systems with poorly understood behavior are deployed before safeguards are ready.1

That argument is precautionary rather than conclusive. The report supports more time for testing, monitoring and governance, but it does not identify a particular slowdown that would be required. Nor does it show that reducing the overall speed of development would work better than other ways of controlling risk.

There are also more concrete concerns about what advanced systems might do. Cybersecurity research points to ways frontier AI could strengthen offensive operations, while a study of language models reports experiments relevant to deceptive behavior. These capabilities could create serious dual-use and control risks, especially if systems behave in ways that are difficult to predict or supervise.2

Slowing development or deployment could limit exposure while safeguards are being tested. But the evidence establishes that these risks are plausible more clearly than it establishes that a broad slowdown is the necessary response.

The case against

The main challenge to the claim is that risk can potentially be reduced through targeted controls rather than by slowing every area of advanced AI. A proposal for oversight of compute providers would monitor access to the large amounts of computing power needed for frontier systems and impose accountability on high-risk users. The Nova Premier evaluation offers an example of model-specific testing for critical risks before or alongside deployment.3

These approaches do not prove that targeted oversight will be enough. They do, however, weaken the stronger claim that development must be slowed in every case. Controls could focus on the most dangerous systems, uses or actors while allowing lower-risk research to continue.

Continued development may also bring safety benefits of its own. Cybersecurity analysis finds that more capable AI could help attackers, but could also support threat detection, incident response and the discovery of software weaknesses.4 The evidence does not show whether these defensive gains outweigh the dangers of faster progress, so this remains an important consideration rather than a decisive rebuttal.

The evidence base also has limits. It does not provide a controlled comparison showing whether slowing development reduces harm more effectively than regulation, deployment restrictions or compute oversight. Researchers have not established a clear link between development speed and the likelihood or severity of safety failures, or identified a point at which slowing becomes essential. Some findings rely on preprints and analyses, while one evaluation covers only a single model and may have developer-related affiliations.

The bottom line

The evidence favours stronger safeguards and precaution in selected high-risk cases, but not a general slowdown of advanced AI development. Confidence is high that advanced AI presents credible safety concerns and that evaluation, monitoring and governance need to improve. Confidence is lower that slowing the overall pace is necessary, because no comparative evidence shows it works better than targeted alternatives.123

The conclusion could differ depending on a system’s capabilities, its intended use and the effectiveness of available controls. Slowing particularly risky development or deployment may be justified. But the current record does not establish that all advanced AI progress should be slowed. The central unanswered question is whether the risk reduction from slower development would outweigh the safety and wider social benefits that controlled continued development might deliver.4