If two AI labs know they're neck-and-neck, does that knowledge push them to race faster and cut more safety corners?

This explores why AI developers who are close to each other in capability, and can see that they are, may be more tempted to skip safety work. The corpus doesn't treat 'proximity' as a single variable. It does have a sharp game-theoretic answer to part of the question, plus several real-world examples of the same dynamic.

The clearest result is surprising: information can make the race worse. In Armstrong, Bostrom and Shulman's model of an AI development race, teams that know exactly how capable their rivals are become more likely to skip safety measures Does revealing rivals' capabilities reduce AI racing disaster risk?. The intuition is that a leader who can see a rival close behind knows that every unit of caution could cost the win, so it cuts precautions. When capabilities are private, no team knows how close the others are, and that uncertainty works as a safety buffer. In the model, disaster becomes likely only when every team happens to be weak at once. Being close is dangerous, but being close and knowing it is worse. That cuts against the common instinct that more transparency between labs always helps.

The broader version is a collective action problem. Each company invests less in safety because a rival that spends less could pull ahead, even though every company would be better off if all of them were careful Can AI companies cooperate on safety despite competitive pressure?. That note names five conditions that can tip things back toward cooperation: trust, shared gains, low exposure, low advantage from defecting, and shared losses. A close race weakens several of them at once. When rivals are evenly matched, a small edge in speed looks decisive, and mutual trust becomes harder to sustain. This is why some argue companies can't fix the problem on their own. The Future of Life Institute calls for binding government limits backed by hardware verification Can companies alone manage the risks of AI systems?. But governments are competitors too. Amodei's proposal to slow the pace of AI development was dismissed by both Washington and Beijing within days Can AI safety pacing work without government cooperation?. Proximity between nations reproduces the same trap one level up.

The same pattern shows up outside AI labs, which suggests it's structural rather than specific to them. Hiring has its own escalation loop: applicants use AI to mass-apply and slip past filters, and employers filter harder in response Are job applicants and employers locked in an escalating AI arms race?. Hoel describes a 'dark forest' in research, where AI can finish a half-shared idea faster than the person who had it, so people stop sharing work in progress Does AI scooping force researchers to hide work in progress?. That is the Armstrong result showing up in culture: once being visible to close competitors becomes risky, people respond by hiding information. Secrecy might lower the racing pressure, but it also erodes the trust that cooperation depends on.

The takeaway you may not have expected: the standard fix of 'more transparency between labs' can backfire in a close race. What reduces risk is changing the payoffs (shared losses, smaller rewards for winning, external enforcement), not just letting everyone see the scoreboard.