Deep and shallow thinking in the long run

GAICIL Insights

Picture a market where some players think one move ahead and others think ten.¹ Intuition says the deep thinkers should dominate. Deep and shallow thinking in the long run, by Heinrich Harald Nax and Jonathan Newton, finds that intuition holds only briefly. Studying cognitive hierarchies has produced new insights for equilibrium analysis, and this paper asks what happens to those insights over the long run.

A provocation up front: much of the anxiety about superintelligent negotiators may be aimed at the wrong target. Over time, the rules of the game, and how risky each move is, may decide a large share of the outcome, whoever sits at the table.

The result is surprising. Short-run behavior is highly sensitive to variation in strategic reasoning abilities, yet that variation is not replicated in the long run. When generalized risk dominant strategy profiles exist, they emerge in the long run independently of players' strategic reasoning abilities. Those abilities may be arbitrarily low or high, heterogeneous across players and evolving over time. GR-dominant strategies emerge when behavioral rules similar to best response are played against a large class of conjectures about other players, including conjectures that others play level k best responses.

In other words: given time, the structure of the game beats the cleverness of the players.

That lands squarely in the age of artificial intelligence, where AI agents and humans interact with very different reasoning depths. An AI system may reason many steps ahead; a human may reason a few. The paper implies that in a broad class of settings, this gap will not alter long-run outcomes. The strategies that prevail will be generalized risk dominant whether the population includes deep-thinking AI or shallow-thinking humans.

That offers a measured reassurance. Fears that sophisticated AI agents will systematically outmaneuver humans may be overstated for long-run outcomes in the settings the paper describes. If risk dominance pins down the long-run convention, individual reasoning depth matters less than the game's structure. Payoffs and risks become the decisive levers, and those are levers institutions can pull.

The short run is another story, and policy should not ignore it. Variation in reasoning abilities does affect short-run outcomes, so AI agents with deeper reasoning may gain advantages before the system converges. Anyone worried about transitional effects, such as algorithmic trading's impact on market dynamics, should watch short-run behavior even while trusting long-run robustness. A long run that arrives eventually is cold comfort to those squeezed on the way there.

The design implication is elegant. If risk dominance decides long-run outcomes, institutions can steer outcomes by changing the risk properties of strategies. Make cooperative or efficient strategies less risky, through insurance or guarantees, and they can become generalized risk dominant and therefore likely to prevail. Crucially, this approach is robust to the cognitive heterogeneity of participants, including any mix of human and AI agents. Policy does not need to know how smart the players are.

The authors point to alternative learning models and to collecting and interpreting empirical data in light of the theory. Experiments pairing human and AI agents of varying reasoning depth would test whether the predicted long-run robustness holds in practice, a natural laboratory for the coming decade.

Stepping back, the structure of strategic environments can dominate the cognitive characteristics of players. Deep and shallow thinkers converge to the same long-run conventions when risk dominance applies. As AI agents join human participants in markets and social interactions, designers and policymakers should concentrate on the properties of the games being played. Get the structure right and good long-run outcomes follow, regardless of who, or what, is doing the thinking.