Falling Behind Drives Unsafe Development in an Idealised AI Race Experiment

Technological races create tension between speed and safety: actors may gain by moving faster than competitors, even when risky development is harmful. This is prominent in debates about artificial intelligence (AI), where competitive pressure is often argued to incentivise riskier, less safety-conscious development. We study this using a framed behavioural experiment based on an idealised AI race, in which paired participants repeatedly chose between Safe and Unsafe development under an uncertain time horizon. Unsafe development gave faster progress and higher immediate payoffs but accumulated private risk up to a treatmentspecific maximum of 10%, 60%, or 90%; the race’s competitive structure was held constant, and only this maximum risk varied. Neither the pre-registered comparison between risk levels nor the role of elicited risk preferences was supported by the data. Instead, exploratory analyses motivated by the task’s repeated structure show that Unsafe behaviour is shaped less by risk preferences than by the evolving strategic state of the race: participants are more likely to choose Unsafe after their opponent does so, being ahead reduces Unsafe play while falling behind increases it, and first-round choices predict later behaviour. To interpret these effects we introduce a reduced evolutionary model with four strategies – Always Safe, Always Unsafe, Conditionally Safe, and Conditionally Antisocial Safe – which reproduces the treatment effect and shows how conditional Unsafe behaviour can be favoured by competitive race dynamics. Together, the experiment and model show that unsafe development can emerge from early behavioural momentum, opponent behaviour, and fear of falling behind, rather than from risk preferences alone, suggesting policy should focus on reducing competitive pressure and promoting cooperation in AI development rather than only individual risk.

Introduction. Rapid technological development often takes place under competitive pressure. Firms, laboratories, or states may benefit from being first to develop, deploy, or control a new technology, but the same race incentives can also create pressure to accept higher levels of risk. This tension is especially salient for artificial intelligence (AI), where rapid progress in large language models (LLMs) and their widespread adoption have intensified concerns about whether competitive dynamics encourage actors to prioritise speed over safety (Hendrycks et al. 2023; Ord 2020). Recent cycles of advanced AI development, in which frontier AI labs and states compete to develop and deploy increasingly capable systems, have been characterised as AI (arm) races toward advanced or transformative AI (Bengio et al. 2024; Gruetzemacher et al. 2025; M ̈uller et al. 2026; Saeri et al. 2026). Here, we ask how technological races shape human decisions when faster development is privately rewarding but potentially risky: when individuals face an AI race between Safe and Unsafe development, do they respond primarily to the level of risk, to their own risk preferences, or to the evolving competitive state of the race? To isolate this question, we hold the competitive structure of the race fixed – two participants always compete under the same rules – and instead manipulate the private cost of taking risks. As detailed below, the central finding is that Unsafe choices track the evolving state of this fixed competition rather than the level of risk or individual risk attitudes. Economic theory has long studied races for innovation, market entry, and technological adoption. Classic models of patent races and pre-emption show that competition can accelerate innovation and induce actors to move early in order to secure strategic advantage (Loury 1979; Fudenberg and Tirole 1985). However, while competition can stimulate innovation under some conditions, it can reduce incentives under others, as captured by the inverted-U relation between competition and innovation (Aghion et al. 2005). Additionally, in many real technological domains, speed is not merely a question of efficiency, but can trade off against risk mitigation. In this direction, tournament and contest behavioural experiments show that participants often take greater risks when risk-taking improves relative performance, when the probability of winning is low, or when the contest is more winner-take-all (Bull et al. 1987; Nieken and Sliwka 2010; Eriksen and Kvaløy 2014; Schedlinsky et al. 2016; Spadoni et al. 2018; G ̈urtler et al. 2023). At the same time, behavioural evidence indicates that risk-taking under competition depends on institutional details, relative standing, gender, prior outcomes, and whether the individual has won or lost previous contests (Filippin and Gioia 2018; Genakos and Pagliero 2012). Thus, the behavioural literature supports that competitive pressure can incentivise riskier behaviour, while also showing that the effect is context-dependent. The AI-race literature makes the safety trade-off more explicit. Formal models of AI development races argue that first-mover advantages can make unsafe development individually attractive, even when unsafe actions increase the probability of harmful outcomes (Armstrong et al. 2016; Han et al. 2020). Related work emphasises that race rhetoric and strategic competition may undermine cooperation over responsible AI development (Cave and ́Oh ́Eigeartaigh 2018; Askell et al. 2019; Barnett and Scher 2025). Evolutionary and game-theoretical analyses of idealised AI races further show how incentives, regulation, heterogeneity, and institutional interventions can alter the prevalence of unsafe behaviour (Han et al. 2021, 2022; Cimpeanu et al. 2022). More recent models of international technology races similarly analyse how uncertainty, information, and strategic competition can generate incentives for risky deployment or insufficient safety investment (Stafford et al. 2022; Emery-Xu et al. 2024; Polborn 2025; Bueno de Mesquita et al. 2026). Across these theoretical contributions, competitive pressure is often argued to incentivise faster, riskier, or less safety-conscious technological development. Despite this theoretical motivation, direct empirical evidence remains limited. Existing behavioural studies of contests and tournaments establish important mechanisms linking competition and risk-taking, but they rarely study repeated technological-race environments in which risky actions both accelerate progress and accumulate endogenous private risk (though see (Zizzo 2002) for a repeated patentrace experiment under technological and dynamic uncertainty). This leaves open the question of whether unsafe technological development is primarily driven by individual risk preferences, by the exogenous level of risk, or by the evolving strategic state of the race.

Method. The main behavioural outcome is whether a participant chose Unsafe in a given round. Before the race task, participants completed a risk-preference elicitation task based on the Eckel–Grossman method (Eckel and Grossman 2008; Dave et al. 2010). The pmax r = 0.6 and pmax r = 0.9 treatments correspond to the preregistered low- and high-risk comparison. The pmax r = 0.1 treatment was added subsequently to test the model prediction that stronger treatment differences should emerge at lower maximum private risk.

The experimental task is a repeated two-player race. In each round t, player i chooses repeated Prisoner’s Dilemma and derives the Nash equilibria of the reduced strategy game for each treatment.

We analyse a reduced strategy set motivated by the regression results:

• AU Always Unsafe: always plays U; • CS Conditionally Safe: plays S in round 1, then from round 2 copies the opponent’s previous action; • CAS Conditionally Antisocial Safe: plays U in round 1, then from round 2 copies the opponent’s previous action.

We consider a finite well-mixed population of size Z. Individuals revise strategies according to a pairwise comparison process with Fermi imitation. If the expected payoffs of strategies A and B are fA and fB, respectively, the probability that an individual using B adopts A is

Discussion. Our two pre-registered hypotheses were not supported: Unsafe behaviour did not differ significantly between the pmax r = 0.6 and 0.9 treatments, and elicited risk preferences did not predict Unsafe choices. The central empirical contribution of this paper instead comes from a set of exploratory analyses that show that Unsafe technological development in this idealised AI race is shaped by interaction history and competitive position. Participants were not simply more or less risk-seeking depending on the exogenous risk treatment or their elicited risk preferences. Instead, they responded to what their opponent had done, to whether they were ahead or behind in the race, and to the early behavioural state of the interaction. This provides behavioural evidence for a central concern in AI and technological race models: competitive pressure can make risky development attractive even when risk is individually costly. The strongest behavioural predictor of Unsafe play was the opponent’s previous action. This suggests that Unsafe development can spread through reciprocal or defensive responses. Once one competitor acts unsafely, the other becomes more likely to do the same, not necessarily because they prefer risk, but because Safe behaviour becomes strategically costly when the competitor is gaining speed. This mechanism is especially relevant in technological races, where safety decisions are rarely made in isolation and where one actor’s unsafe behaviour can change the incentives faced by others. It also highlights an important target for public policy in reducing the strategic pressure to match unsafe behaviour by improving transparency, coordination, monitoring, or credible commitments to safe development. This pattern of conditioning on the opponent’s previous action is consistent with the reciprocal strategies – Titfor-Tat-like reactive rules in particular – that dominate the strategy classifications inferred in repeated Prisoner’s Dilemma experiments (Dal B ́o and Fr ́echette 2018; Montero-Porras et al. 2022). Our Conditionally Safe and Conditionally Antisocial Safe strategies (Section S3.2) can therefore be understood as race-specific instances of this broader family of conditionally cooperative strategies. Race position further shaped behaviour. Participants who were ahead were less likely to choose Unsafe, whereas those who were behind had stronger incentives to take risks. This supports the idea that fear of falling behind can promote unsafe development. Once more, this effect does not require participants to be generally riskseeking. Rather, the race structure itself makes Unsafe development attractive when it offers a chance to recover from a disadvantage. At the same time, participants still reduced Unsafe behaviour when maximum private risk was high. The marginal distributions show a lower modal frequency of Unsafe behaviour for pmax r = 0.9, and the evolutionary model predicts a shift towards Conditionally Safe behaviour at high private risk. Thus, high risk does not eliminate Unsafe development, but it changes the strategic conditions under which Unsafe behaviour is favoured. Participants remain willing to take risks when the competitive context makes Unsafe development advantageous, but high private risk weakens the dominance of unconditional Unsafe play. The reduced evolutionary model provides a mechanistic interpretation of these results. Conditional strategies that respond to the opponent’s previous action, especially those that differ in their first-round behaviour, can reproduce the observed treatmentlevel patterns. At low private risk, Always Unsafe behaviour is favoured because the speed advantage dominates the expected cost of risk.

Conclusion. This study provides behavioural evidence that Unsafe technological development in an idealised AI race is shaped by competitive dynamics rather than by risk preferences alone. Participants conditioned their choices on the opponent’s previous behaviour and on their relative position in the race, with lagging participants showing a stronger tendency to choose Unsafe. A reduced evolutionary model showed that these patterns can be reproduced by a small set of conditional strategies, linking the experimental results to broader theoretical concerns about technological races. These findings suggest that safety failures in competitive AI development may arise not only because actors underestimate risk, but because competition changes the strategic value of taking that risk. When Unsafe development offers speed and catch-up advantages, safety becomes conditional on the behaviour of rivals. Effective governance of technological races may therefore require interventions that reduce incentives for reciprocal Unsafe behaviour and make Safe development strategically viable even under competitive pressure.

Limitations. The study has several limitations. First, the race is deliberately stylised and short relative to real technological competition. Real AI and technological races involve many actors, repeated public signals, uncertain capabilities, regulatory constraints, and reputational concerns. Second, the reduced model approximates richer distancedependent behaviour using a small number of deterministic strategies. This makes the mechanism interpretable, but it abstracts away from gradual responses to race distance and from more heterogeneous learning rules. Third, the experimental design focuses on private risk rather than collective or systemic risk. This is useful for isolating individual incentives, but many real safety failures impose externalities on other actors or on society more broadly. Fourth, the central empirical findings reported here – the roles of opponent behaviour, race position, and first-round momentum – emerged from exploratory rather than pre-registered analyses. Fifth, the three treatments manipulate only the maximum private risk while holding the competitive structure of the race – the presence of a rival and the incentive to outpace them – constant.