Pith. sign in

REVIEW 4 major objections 4 minor 40 references

The paper argues that unsafe AI development in a two-player race is driven by the opponent's behaviour and by falling behind, not by the level of private risk or by people's measured risk preferences.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 00:46 UTC pith:MYZUDO6F

load-bearing objection A genuinely new behavioral AI-race experiment with a robust opponent effect, but the headline causal story runs ahead of the paper's own non-exogenous regressions; worth refereeing seriously, with a required revision. the 4 major comments →

arxiv 2607.26034 v1 pith:MYZUDO6F submitted 2026-07-28 cs.AI cs.CYcs.GTecon.GNq-fin.EC

Falling Behind Drives Unsafe Development in an Idealised AI Race Experiment

Elias Fern\'andez Domingos , The Anh Han This is my paper
classification cs.AI cs.CYcs.GTecon.GNq-fin.EC MSC 91-0591A2091A22
keywords AI racerisk-takingbehavioural experimentrepeated gameevolutionary game theorycompetitive pressureconditional cooperationtechnological risk
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that, in a two-player technological race, unsafe development is driven by the dynamics of competition—the opponent's previous actions, the participant's position in the race, and early momentum—rather than by the size of the private risk or by people's measured risk preferences. Two pre-registered hypotheses failed: raising the maximum private risk from 60% to 90% did not reduce unsafe choices, and an Eckel–Grossman risk-preference measure did not predict them. Exploratory analyses show that participants were substantially more likely to choose Unsafe after their opponent had done so, that being ahead made Unsafe less likely, and that falling behind made it more likely, with first-round Unsafe choices forecasting later behaviour. A four-strategy evolutionary model reproduces the treatment pattern and shows how conditional Unsafe play can persist under competitive race incentives. If the paper is right, race structure itself creates pressure toward unsafe AI development even among risk-averse actors, so governance should target competitive pressure and reciprocal behaviour rather than only individual risk attitudes.

Core claim

Central claim: when two players repeatedly race toward a technology prize, an individual's decision to develop unsafely is determined less by how much private risk unsafe development carries (10%, 60%, or 90%) or by that person's elicited risk preference, and more by the strategic state of the race. The most consistent predictor is the opponent's previous round: after an opponent chooses Unsafe, the focal player is markedly more likely to choose Unsafe. Being ahead in race steps lowers the probability of Unsafe, while falling behind raises it; choosing Unsafe in the first round predicts continued Unsafe play. The reduced evolutionary model recovers the qualitative treatment differences using

What carries the argument

The load-bearing object is the indefinitely repeated two-player race: Safe development advances one step, Unsafe 1.5 steps with higher round payoffs but accumulating private risk up to a maximum; the race lasts at least five rounds and then ends with 20% probability each round, and only winners or ties face the risk lottery. Around this game the paper builds a reduced evolutionary model with four strategies—Always Safe, Always Unsafe, Conditionally Safe (Safe first, then copy opponent), and Conditionally Antisocial Safe (Unsafe first, then copy opponent)—whose expected payoffs are computed analytically for unconditional pairs and by Monte Carlo for conditional pairs, and whose long-run frequ

Load-bearing premise

The central interpretive claim—that falling behind and observing an opponent's Unsafe move cause unsafe choices—requires reading the lagged behavioural coefficients as causal, but the paper states that those lagged regressors are not exogenous, so the measured associations may reflect unobserved strategic history or selection into race positions rather than a direct effect.

What would settle it

Conduct the same race with a scripted or randomly assigned opponent: for example, a yoked design where the opponent's previous action is determined by an independent schedule rather than by the participant's own history. If the opponent-effect and falling-behind effect disappear when opponent behaviour is exogenous, the causal reading fails. A simpler check within the current data would be to include the full preceding action history, not just the one-round lags; if the lagged coefficients shrink toward zero, the apparent momentum is the trace of longer-run history rather than a per-round trig

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Raising the private cost of risky development from 60% to 90% did not lower unsafe play, so making risk more expensive to the individual developer is not by itself a sufficient brake under race dynamics.
  • Unsafe behaviour is contagious: after one player chooses it, the rival is more likely to choose it too, implying a single aggressive actor can shift the race toward unsafe play.
  • Being behind in the race increases unsafe choices, so reducing the payoff to catching up—or protecting lagging competitors—would plausibly dampen unsafe development more than individual risk warnings.
  • Elicited risk preferences did not predict behaviour, so interventions aimed at changing individual risk attitudes are unlikely to change race outcomes.
  • The evolutionary model indicates that at the highest private-risk level a conditionally safe strategy (start Safe, then copy the opponent) becomes dominant, so safety can persist if actors can reciprocate safe behaviour.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the causal story: script or randomly assign the opponent's first-round and mid-game actions so they are exogenous to the participant's own history; if the opponent-effect and catch-up effect survive, the 'fear of falling behind' mechanism is causal, and if they vanish it is selection.
  • Because the stage payoffs form a Deadlock rather than a Prisoner's Dilemma (T > P > R > S), the social dilemma only appears once the terminal private risk is included. This suggests that whether competitive races generate unsafe outcomes may hinge on whether unsafe choices impose collective or externalised risk; a natural extension would replace private risk with shared or societal risk and look f
  • The opponent-copying behaviour is the same family as Tit-for-Tat, so one testable governance idea is to make 'safety-first' behaviour visible and credibly committed in the first rounds of a race, potentially shifting early momentum from aggressive to safe play.
  • The paper found no difference between the 60% and 90% maximum-risk treatments, so the model's equilibrium shift toward Conditionally Safe at 90% is a prediction to be probed at even higher values or with longer horizons before drawing firm policy conclusions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper reports a framed, two-player 'AI race' experiment in which participants repeatedly choose Safe or Unsafe development under an uncertain horizon, with the maximum private setback risk set to 10%, 60%, or 90%. The pre-registered hypotheses—that the 0.6 vs 0.9 risk contrast changes Safe choices and that elicited risk preferences predict Unsafe choices—are not supported. The authors then report exploratory panel-logistic regressions suggesting that Unsafe choices are associated with the opponent's previous action, with falling behind in race steps, and with a first-round Unsafe choice, but not robustly with the participant's own previous action. A reduced evolutionary model with four strategies (AS, AU, CS, CAS) is introduced, and the authors claim it reproduces the treatment-level pattern and shows how conditional Unsafe behaviour can be favoured by race dynamics. The paper concludes that unsafe AI development can emerge from interaction history, opponent behaviour, and fear of falling behind, rather than from risk preferences alone.

Significance. If the empirical associations are robust, the paper provides novel behavioural evidence that competitive race dynamics—opponent copying and lagging behind—can shape risky technology choices independently of static risk attitudes. The authors are admirably transparent: they report that the pre-registered hypotheses failed, disclose the exploratory status of the headline results, provide effect sizes for null comparisons without post-hoc power fallacies, and include a robustness check for a mid-race dropout. The evolutionary model is analytically explicit and the treatment-dependent equilibria (AU/CAS at low and intermediate risk, CS at high risk) are a useful theoretical illustration. However, the strength of the contribution is currently undercut by the gap between the acknowledged associational nature of the regressions and the causal language used in the abstract and discussion, and by the fragility of the key race-position result.

major comments (4)
  1. [Abstract; Section 2; Section 4; Section 6.3] The manuscript states in Section 6.3 and Section 4 that a_i^{t-1}, a_{-i}^{t-1} and ΔS^{t-1} are 'not exogenous regressors' and that the coefficients 'should be read as conditional associations rather than as estimates of the causal effect of one round's action on the next.' Yet the Abstract and Results claim Unsafe behaviour is 'driven by' interaction history and 'emerges from' fear of falling behind. Without pair-level or participant-level fixed effects, stable unobserved pair differences (competitiveness, beliefs, strategic sophistication) can create the same correlations without any social influence. This is the reflection/correlated-effects problem, and it is load-bearing for the headline conclusion. Please add fixed-effects robustness specifications or re-frame the central claim as associational, with the mechanism presented as a model-based interpretation.
  2. [Section 2, Table 1] The race-position effect is not robust across the reported specifications. The coefficient on ΔS^{t-1} is not significant in model 3 (β=-0.238, p=0.113) and only marginally significant in model 6 (β=-0.296, p=0.048). These are exploratory analyses with no correction for the six models and many interaction terms, and the pre-registered hypotheses failed. The abstract's statement that 'falling behind increases [Unsafe play]' rests on one marginal coefficient. Please report a specification curve or sensitivity analysis, state the uncorrected and corrected p-values, and soften the claim to a hypothesis-generating association.
  3. [Section 6.4; Figure 3] The four-strategy model is not independent of the regressions: the choice of CAS (starts Unsafe) is motivated by the first-round momentum effect, and CS is motivated by the reciprocal opponent-response effect. The best-fit parameters β=0.01 and μ=0.05 are then selected to reproduce the treatment-level Unsafe frequencies. Consequently, 'the model reproduces the treatment effect' and 'Together, the experiment and model show...' overstate the status of the model: it is a calibrated illustration of one possible mechanism, not an independent confirmation. Please state explicitly that the model is fitted to the same data and that its agreement does not validate the empirical associations.
  4. [Section S1.1.2, Table S1 vs Section 2, Table 1] There is an unresolved inconsistency about the effect of the participant's own previous action. Table S1 reports the mixed-effects model with own lagged action as the most robust predictor (β=-0.620, p<0.001), whereas Section 2 states that the coefficient on a_i^{t-1} is small and insignificant across the Table 1 specifications. This discrepancy needs reconciliation; if the negative own-lag effect is real in the pre-registered mixed model, the claim that 'later Unsafe play is not explained by simple action persistence alone' is not robust to specification choice.
minor comments (4)
  1. [Section S1.1.3] The instruction error about the expected number of rounds (stated as 10 rather than 9) is honestly disclosed in the supplement; it would be helpful to note in the main text as well, since participants' beliefs about the horizon could matter for the interpretation of risk-taking.
  2. [Section 3; Section S3.2] The Discussion compares CS/CAS to Tit-for-Tat in repeated Prisoner's Dilemma experiments, but the stage-game ordering is T>P>R>S (a Deadlock game), which the authors acknowledge only in S3.2. The analogy would be clearer if the Discussion explicitly noted that the repeated structure converts the game into a social dilemma through the terminal risk.
  3. [Table 1 caption] The caption says models (1)-(3) exclude a_i^1 and (4)-(6) include it; this is clear, but the main text alternately refers to 'initial-action bias' and 'first-round action'. Standardizing the terminology would avoid confusion.
  4. [Figure 2C] The winner/loser correlation in Figure 2C is descriptive and may reflect the mechanical fact that both players' Unsafe frequencies are correlated within pairs. Please make explicit that this panel is not a causal analysis of race outcomes.

Circularity Check

1 steps flagged

Model 'reproduction' of the treatment effect is partly calibrated to the experimental medians, but the central experimental finding stands independently.

specific steps
  1. fitted input called prediction [Section 2, 'A reduced evolutionary model reproduces the behavioural pattern'; Figure 3B; Methods 6.5]
    "The best-fit point (pink cross) is obtained when fixing a higher mutation rate, μ=0.05, and weak selection, β=0.01, which, in combination, capture a higher amount of behavioural noise expected in the experiment... Figure 3B compares the experimental median frequency of Unsafe choices with the model predictions. The model captures the qualitative pattern across maximum private-risk treatments, including the relatively small difference between the two higher-risk treatments and the stronger contrast with the low-risk treatment."

    The model's β and μ are selected as the 'best-fit' parametrisation against the same three experimental median Unsafe frequencies that Figure 3B then reports as 'model predictions' (MAE=0.175). The claimed 'reproduction' of the treatment effect is therefore a description of calibration to the target data rather than an independent out-of-sample prediction. Moreover, the strategy set itself is explicitly motivated by the same regressions: CS and CAS copy the opponent's previous action, and the Safe/Unsafe first-round distinction is motivated by the first-round effect in Table 1. Thus the model's 'interpretation' partly repackages the empirical effects it was built from, although it does add the independent claim that such conditional strategies are evolutionarily viable under the payoff stru

full rationale

The central empirical result — Unsafe choices track the opponent's previous action, race position, and first-round momentum — is established directly by the cluster-robust panel regressions in Table 1 and the descriptive patterns in Figure 2; it does not depend on the evolutionary model. The model is transparently presented as being motivated by those regression results and is then fitted at β=0.01, μ=0.05 to match the experimental medians, so its 'reproduction' of the treatment pattern is partly a calibration exercise rather than a genuinely independent prediction. However, the paper acknowledges the exploratory status of both the analyses and the model, and the model's payoff-based phase transitions provide some independent content. There is no load-bearing self-citation or imported uniqueness theorem, and the pre-registered null results are reported honestly. Overall the circularity is moderate and confined to the model's fitted 'prediction', not the core experimental finding.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The model depends on two fitted evolutionary parameters (β, μ), a standard population size Z, and hand-chosen stage-game parameters calibrated to the experiment. The strategy set is ad hoc, derived from the same data patterns the model is claimed to explain. There are no new physical entities or independent external handles.

free parameters (4)
  • β (selection intensity) = best fit β = 0.01; reference β ≈ 2
    Chosen to fit the model's predicted Unsafe frequency to the experimental medians (Fig 3B). The treatment ordering is robust across a range of β, but the quantitative match relies on this fit.
  • μ (mutation rate) = best fit μ = 0.05; reference μ = β/Z
    Chosen together with β as the best-fit point; the authors describe it as capturing 'behavioural noise' in the experiment (Fig 3B, S3.3).
  • Z (population size) = 100
    Standard finite-population size used in the evolutionary dynamics; the stationary distribution depends on Z, though treatment ordering is reported as robust (Fig S6).
  • Stage-game parameters b=4, c=1, B=100, sU=1.5, sS=1 = b=4, c=1, B=100, sU=1.5, sS=1
    Chosen by hand to reproduce the experimental payoff matrix and race prize; they are model inputs, not fitted to behavioral outcomes (S3.1.2).
axioms (5)
  • ad hoc to paper The four-strategy set AS/AU/CS/CAS adequately approximates the behavior observed in the experiment.
    The strategies are motivated by the same regression results the model is used to explain; S3.1.4 explicitly calls them 'reduced-form proxies for a richer behavioural rule'.
  • domain assumption Fermi pairwise imitation with mutation is an appropriate model of how human strategy choices evolve in this population.
    Standard evolutionary game theory assumption; not independently validated against learning data in this paper (Methods 6.5).
  • domain assumption The experimental race captures the strategically relevant features of real AI development races.
    The paper frames the task as an 'idealised AI race' and discusses generalization in Limitations (Section 4); external validity is assumed.
  • domain assumption Private setback risk only applies to race winners/ties, matching how participants were instructed.
    This payoff rule is part of the experimental design (Instructions 3/4) and is carried into the model; if real-world risks are collective rather than private, conclusions may not transfer (Section 4).
  • domain assumption Eckel-Grossman gamble choice is a valid continuous measure of risk preference.
    Standard measure, cited; the null result for risk preferences is interpreted against this operationalization (S1.3.2, S2.2).

pith-pipeline@v1.3.0-alltime-deepseek · 27734 in / 12117 out tokens · 121219 ms · 2026-08-01T00:46:16.725537+00:00 · methodology

0 comments
read the original abstract

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 treatment-specific 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.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 22 canonical work pages

  1. [1]

    , Bloom , N

    barticle Aghion , P. , Bloom , N. , Blundell , R. , Griffith , R. , Howitt , P. : Competition and innovation: An inverted-u relationship . The Quarterly Journal of Economics 120 ( 2 ), 701 -- 728 ( 2005 ) 10.1162/0033553053970214 barticle

    work page doi:10.1162/0033553053970214 2005

  2. [2]

    , Brundage , M

    barticle Askell , A. , Brundage , M. , Hadfield , G.K. : The role of cooperation in responsible AI development . arXiv preprint arXiv:1907.04534 ( 2019 ) 10.48550/arXiv.1907.04534 https://arxiv.org/abs/1907.04534 arXiv:1907.04534 [cs.CY] barticle

    Pith/arXiv arXiv doi:10.48550/arxiv.1907.04534 1907

  3. [3]

    , Bostrom , N

    barticle Armstrong , S. , Bostrom , N. , Shulman , C. : Racing to the precipice: A model of artificial intelligence development . AI & Society 31 ( 2 ), 201 -- 206 ( 2016 ) 10.1007/s00146-015-0590-y barticle

    work page doi:10.1007/s00146-015-0590-y 2016

  4. [4]

    , Dziuda , W

    botherref Bueno de Mesquita , E. , Dziuda , W. , Polborn , M. : The AGI race and existential risk. Technical Report 35276, National Bureau of Economic Research (May 2026). 10.3386/w35276 botherref

    work page doi:10.3386/w35276 2026

  5. [5]

    , Hinton , G

    barticle Bengio , Y. , Hinton , G. , Yao , A. , Song , D. , Abbeel , P. , Darrell , T. , Harari , Y.N. , Zhang , Y.-Q. , Xue , L. , Shalev-Shwartz , S. , : Managing extreme AI risks amid rapid progress . Science 384 ( 6698 ), 842 -- 845 ( 2024 ) barticle

    2024

  6. [6]

    : An introduction to linear mixed-effects modeling in R

    barticle Brown , V.A. : An introduction to linear mixed-effects modeling in R . Advances in Methods and Practices in Psychological Science 4 ( 1 ), 2515245920960351 ( 2021 ) 10.1177/2515245920960351 barticle

    work page doi:10.1177/2515245920960351 2021

  7. [7]

    , Scher , A

    botherref Barnett , P. , Scher , A. : AI governance to avoid extinction: The strategic landscape and actionable research questions. arXiv preprint arXiv:2505.04592 (2025) botherref

    Pith/arXiv arXiv 2025

  8. [8]

    , Schotter , A

    barticle Bull , C. , Schotter , A. , Weigelt , K. : Tournaments and piece rates: An experimental study . Journal of Political Economy 95 ( 1 ), 1 -- 33 ( 1987 ) 10.1086/261439 barticle

    work page doi:10.1086/261439 1987

  9. [9]

    , \' O h \' E igeartaigh , S.S

    bchapter Cave , S. , \' O h \' E igeartaigh , S.S. : An AI race for strategic advantage: Rhetoric and risks . In: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society , pp. 36 -- 40 . Association for Computing Machinery , New York, NY, USA ( 2018 ). 10.1145/3278721.3278780 bchapter

    arXiv 2018

  10. [10]

    , Santos , F.C

    barticle Cimpeanu , T. , Santos , F.C. , Pereira , L.M. , Lenaerts , T. , Han , T.A. : Artificial intelligence development races in heterogeneous settings . Scientific Reports 12 , 1723 ( 2022 ) 10.1038/s41598-022-05729-3 barticle

    work page doi:10.1038/s41598-022-05729-3 2022

  11. [11]

    , Schonger , M

    barticle Chen , D.L. , Schonger , M. , Wickens , C. : otree—an open-source platform for laboratory, online, and field experiments . Journal of Behavioral and Experimental Finance 9 , 88 -- 97 ( 2016 ) barticle

    2016

  12. [12]

    , Fr \'e chette , G.R

    barticle Dal B \'o , P. , Fr \'e chette , G.R. : On the determinants of cooperation in infinitely repeated games: A survey . Journal of Economic Literature 56 ( 1 ), 60 -- 114 ( 2018 ) barticle

    2018

  13. [13]

    , Eckel , C.C

    barticle Dave , C. , Eckel , C.C. , Johnson , C.A. , Rojas , C. : Eliciting risk preferences: When is simple better? Journal of Risk and Uncertainty 41 ( 3 ), 219 -- 243 ( 2010 ) 10.1007/s11166-010-9103-z barticle

    work page doi:10.1007/s11166-010-9103-z 2010

  14. [14]

    , Santos , F.C

    botherref Domingos , E.F. , Santos , F.C. , Lenaerts , T. : Egttools: Evolutionary game dynamics in python. Iscience 26(4) (2023) botherref

    2023

  15. [15]

    , Grossman , P.J

    barticle Eckel , C.C. , Grossman , P.J. : Forecasting risk attitudes: An experimental study using actual and forecast gamble choices . Journal of Economic Behavior & Organization 68 ( 1 ), 1 -- 17 ( 2008 ) 10.1016/j.jebo.2008.04.006 barticle

    work page doi:10.1016/j.jebo.2008.04.006 2008

  16. [16]

    , Kval y , O

    barticle Eriksen , K.W. , Kval y , O. : Myopic risk-taking in tournaments . Journal of Economic Behavior & Organization 97 , 37 -- 46 ( 2014 ) 10.1016/j.jebo.2013.10.004 barticle

    work page doi:10.1016/j.jebo.2013.10.004 2014

  17. [17]

    , Park , A

    barticle Emery-Xu , N. , Park , A. , Trager , R. : Uncertainty, information, and risk in international technology races . Journal of Conflict Resolution 68 ( 10 ), 2019 -- 2047 ( 2024 ) 10.1177/00220027231214996 barticle

    work page doi:10.1177/00220027231214996 2019

  18. [18]

    , Gioia , F

    barticle Filippin , A. , Gioia , F. : Competition and subsequent risk-taking behaviour: Heterogeneity across gender and outcomes . Journal of Behavioral and Experimental Economics 75 , 84 -- 94 ( 2018 ) 10.1016/j.socec.2018.05.003 barticle

    work page doi:10.1016/j.socec.2018.05.003 2018

  19. [19]

    , Tirole , J

    barticle Fudenberg , D. , Tirole , J. : Preemption and rent equalization in the adoption of new technology . The Review of Economic Studies 52 ( 3 ), 383 -- 401 ( 1985 ) 10.2307/2297660 barticle

    work page doi:10.2307/2297660 1985

  20. [20]

    , Avin , S

    barticle Gruetzemacher , R. , Avin , S. , Fox , J. , Saeri , A.K. : Strategic insights from simulation gaming of AI race dynamics . Futures 167 , 103563 ( 2025 ) barticle

    2025

  21. [21]

    , Pagliero , M

    barticle Genakos , C. , Pagliero , M. : Interim rank, risk taking, and performance in dynamic tournaments . Journal of Political Economy 120 ( 4 ), 782 -- 813 ( 2012 ) 10.1086/668502 barticle

    work page doi:10.1086/668502 2012

  22. [22]

    , Struth , L

    barticle G \"u rtler , O. , Struth , L. , Thon , M. : Competition and risk-taking . European Economic Review 160 , 104592 ( 2023 ) 10.1016/j.euroecorev.2023.104592 barticle

    arXiv 2023

  23. [23]

    , Heisey , D.M

    barticle Hoenig , J.M. , Heisey , D.M. : The abuse of power: The pervasive fallacy of power calculations for data analysis . The American Statistician 55 ( 1 ), 19 -- 24 ( 2001 ) 10.1198/000313001300339897 barticle

    work page doi:10.1198/000313001300339897 2001

  24. [24]

    , Lenaerts , T

    barticle Han , T.A. , Lenaerts , T. , Santos , F.C. , Pereira , L.M. : Voluntary safety commitments provide an escape from over-regulation in AI development . Technology in Society 68 , 101843 ( 2022 ) barticle

    2022

  25. [25]

    , Moniz Pereira , L

    barticle Han , T.A. , Moniz Pereira , L. , Santos , F.C. , Lenaerts , T. : To regulate or not: A social dynamics analysis of an idealised AI race . Journal of Artificial Intelligence Research 69 , 881 -- 921 ( 2020 ) 10.1613/jair.1.12225 barticle

    work page doi:10.1613/jair.1.12225 2020

  26. [26]

    , Mazeika , M

    botherref Hendrycks , D. , Mazeika , M. , Woodside , T. : An overview of catastrophic AI risks. arXiv preprint arXiv:2306.12001 (2023) botherref

    Pith/arXiv arXiv 2023

  27. [27]

    , Pereira , L.M

    barticle Han , T.A. , Pereira , L.M. , Lenaerts , T. , Santos , F.C. : Mediating artificial intelligence developments through negative and positive incentives . PLOS ONE 16 ( 1 ), 0244592 ( 2021 ) 10.1371/journal.pone.0244592 barticle

    work page doi:10.1371/journal.pone.0244592 2021

  28. [28]

    : Market structure and innovation

    barticle Loury , G.C. : Market structure and innovation . The Quarterly Journal of Economics 93 ( 3 ), 395 -- 410 ( 1979 ) barticle

    1979

  29. [29]

    : Conditional logit analysis of qualitative choice behavior

    bchapter McFadden , D. : Conditional logit analysis of qualitative choice behavior . In: Zarembka , P. (ed.) Frontiers in Econometrics , pp. 105 -- 142 . Academic Press , New York ( 1974 ) bchapter

    1974

  30. [30]

    , Gruji \'c , J

    barticle Montero-Porras , E. , Gruji \'c , J. , Fern \'a ndez Domingos , E. , Lenaerts , T. : Inferring strategies from observations in long iterated prisoner's dilemma experiments . Scientific Reports 12 ( 1 ), 7589 ( 2022 ) 10.1038/s41598-022-11654-2 barticle

    work page doi:10.1038/s41598-022-11654-2 2022

  31. [31]

    , Steels , L

    barticle M \"u ller , V. , Steels , L. , Szathm \'a ry , E. : Evolvable AI : Threats of a new major transition in evolution . Proceedings of the National Academy of Sciences 123 ( 17 ), 2527700123 ( 2026 ) 10.1073/pnas.2527700123 barticle

    work page doi:10.1073/pnas.2527700123 2026

  32. [32]

    , Sliwka , D

    barticle Nieken , P. , Sliwka , D. : Risk-taking tournaments: Theory and experimental evidence . Journal of Economic Psychology 31 ( 3 ), 254 -- 268 ( 2010 ) 10.1016/j.joep.2009.03.009 barticle

    work page doi:10.1016/j.joep.2009.03.009 2010

  33. [33]

    : The Precipice: Existential Risk and the Future of Humanity

    bbook Ord , T. : The Precipice: Existential Risk and the Future of Humanity . Hachette UK , London ( 2020 ) bbook

    2020

  34. [34]

    : Competitive development of dangerous technologies

    botherref Polborn , M.K. : Competitive development of dangerous technologies. Technical report, Vanderbilt University (April 2025) botherref

    2025

  35. [35]

    , Schitter , C

    barticle Palan , S. , Schitter , C. : Prolific.ac—a subject pool for online experiments . Journal of Behavioral and Experimental Finance 17 , 22 -- 27 ( 2018 ) 10.1016/j.jbef.2017.12.004 barticle

    doi:10.1016/j.jbef.2017.12.004 2018

  36. [36]

    , Graham , J

    botherref Saeri , A.K. , Graham , J. , Noetel , M. , Slattery , P. , Ah-King , D. , Aittokallio , E. , Akindehin , I. , Mahdi , A.A. , Alhajjar , E. , Lipcsey , R.A. , et al.: Prioritization of risks from artificial intelligence: A delphi study of 272 international experts. arXiv preprint arXiv:2606.04490 (2026) botherref

    Pith/arXiv arXiv 2026

  37. [37]

    , Sommer , F

    barticle Schedlinsky , I. , Sommer , F. , W \"o hrmann , A. : Risk-taking in tournaments: an experimental analysis . Journal of Business Economics 86 ( 8 ), 837 -- 866 ( 2016 ) 10.1007/s11573-016-0813-9 barticle

    work page doi:10.1007/s11573-016-0813-9 2016

  38. [38]

    , Trager , R.F

    botherref Stafford , E. , Trager , R.F. , Dafoe , A. : Safety not guaranteed: International races for risky technologies. Technical report, Centre for the Governance of AI (November 2022) botherref

    2022

  39. [39]

    , Ven , J

    barticle Spadoni , L. , Ven , J. , Willems , T. : The effect of competition on risk taking in contests . Games 9 ( 3 ), 72 ( 2018 ) 10.3390/g9030072 barticle

    work page doi:10.3390/g9030072 2018

  40. [40]

    : Racing with uncertainty: a patent race experiment

    barticle Zizzo , D.J. : Racing with uncertainty: a patent race experiment . International Journal of Industrial Organization 20 , 877 -- 902 ( 2002 ) 10.1016/S0167-7187(01)00087-X barticle

    work page doi:10.1016/s0167-7187(01)00087-x 2002