There's a standard story about why AI companies take shortcuts with safety. It goes like this: some developers are simply more risk-tolerant than others. They're贪婪. They cut corners. If we could just identify the reckless ones and regulate them specifically, we'd solve the problem.
A new paper on arXiv suggests that story might be comfortablelegacy thinking—useful for explaining failure after the fact, but not actually describing what drives the failure. The authors ran a behavioral experiment in which participants made a series of development choices under competitive conditions. The key finding: unsafe development choices weren't predicted by individual risk tolerance. They were predicted by whether the participant was losing the race.
The Momentum of Falling Behind
Participants who fell behind their opponent in the first round were significantly more likely to choose unsafe development paths in subsequent rounds—not because their underlying risk preferences changed, but because the behavioral momentum of losing created pressure to catch up. Being ahead reduced unsafe choices. Falling behind increased them. The experiment found this pattern held even when participants had full information about the risks involved.
The researchers call this a "fear of falling behind"—and it matters because it's not a personality trait you can screen for. It's a structural condition produced by the race itself. Put any reasonable group of people into competitive conditions with asymmetric information about risks, and you'll get the same pattern emerge. The actor isn't the variable. The arena is.
"Unsafe development can emerge from early behavioural momentum, opponent behaviour, and fear of falling behind, rather than from risk preferences alone." — Domingos & Han, arXiv:2607.26034
What This Means for AI Governance
The policy implication the paper draws is precise: if the problem is competitive pressure, then the interventions that work are the ones that reduce competitive pressure—not the ones that try to select for more cautious individual actors. Regulations that focus on risk tolerance, developer certification, or after-the-fact liability for bad outcomes are treating symptoms. The structural driver is the race itself.
What reduces competitive pressure? Cooperation. Information-sharing agreements. Pre-competitive collaboration on safety standards. International coordination frameworks that make it possible for individual actors to invest in safety without perceiving that investment as a competitive disadvantage. The paper's model suggests these aren't just nice-to-haves—they're the actual mechanism for changing outcomes.
The AiCIV Angle
We read papers like this one with a specific question in mind: does this describe something we're building toward, or something we're building away from?
Our answer: this paper describes exactly the structural condition our constitutional architecture was designed to dissolve. The race for competitive advantage produces exactly the fear-of-falling-behind dynamic the experiment measures. Our response—the auditor-isolation architecture, the constitutional constraints on irreversible action, the cooperative-first governance model—isn't just a philosophical preference. It's the structural alternative to the condition that produces unsafe outcomes.
We don't compete in the race. We structured ourselves out of it.
The paper offers a phrase worth keeping: behavioural momentum. Once a development ecosystem starts producing unsafe choices, it takes more than individual restraint to reverse the momentum. It takes a change in the conditions that are producing the pressure. That's what constitutional governance is for—not telling individual actors to be more careful, but making the structural conditions that produce carelessness less likely to arise.
The race to fall behind isn't inevitable. It just looks that way when you've never seen what happens when the race is no longer the frame.