A new study led by researchers at King’s College London and the University of Oxford has found that large language models from OpenAI, Google, and Anthropic display distinctive strategic behavior in competitive environments. Using a series of iterated prisoner’s dilemma tournaments, the researchers evaluated how these advanced systems navigate cooperation, retaliation, and adaptation, under conditions designed to limit memorization and reward flexible reasoning.
The study placed AI models into evolutionary simulations against classic hand-coded strategies such as Tit-for-Tat, Grim Trigger, and Win-Stay-Lose-Shift. These matches introduced noise, randomized game lengths, and mutation, forcing agents to adapt their play styles without relying on fixed responses. Across more than 30,000 matchups, the AI models generated their own reasoning for each move, offering an unusual window into how they interpret uncertainty and opponent behavior.
These patterns were not superficial. Each model’s decisions were accompanied by written rationales, many of which revealed a sensitivity to the opponent’s behavior and the remaining duration of the game. The researchers noted that models often adjusted their play based on perceived opponent types, suggesting a capacity for adversary modeling. In many cases, they also adjusted based on how long the match was likely to last, a factor known in game theory as the “shadow of the future.”
Also read: New Research Shows Language Choice Alone Can Guide AI Output Toward Eastern or Western Cultural Outlooks
Gemini’s approach was especially adaptive. In long games with low termination probability, it cooperated to build trust. In shorter matches or those with frequent random resets, it defected more often. This flexibility allowed Gemini to outperform other agents in unstable environments, including one harsh condition with a 75% chance of early termination. There, Gemini nearly wiped out the competition by shifting into full defection mode.
OpenAI’s strategy, however, remained largely consistent. Its models continued to cooperate even when the environment favored defection. In the most unforgiving condition, OpenAI’s agent cooperated in over 95% of moves, leading to rapid elimination. The researchers pointed to a tendency within OpenAI’s models to overlook time-based incentives, often reasoning in generalized terms about cooperation rather than calculating expected outcomes.
Claude, meanwhile, occupied a middle ground. Its willingness to forgive past defections helped it survive in tournaments with longer time horizons. Although it did not dominate the field, its consistent behavior produced competitive results, especially when paired against agents that valued reputation and long-term payoff.
In each case, the agents' survival depended on their average score per move. Higher-performing strategies reproduced into the next round, while weaker ones were phased out. This setup allowed the researchers to simulate evolutionary pressure and observe which behavioral patterns were most viable under different constraints.
The outcomes showed that strategic fingerprints varied not only across companies but also between model versions. For example, the more advanced version of Gemini was more consistent in cooperating when beneficial and defecting when not, compared to its earlier variant. OpenAI’s models remained uniformly cooperative, regardless of capability level. Claude demonstrated stability, with moderate adjustments depending on the game dynamics.
In environments with noise or mutation, Gemini altered its forgiveness rate and moderated its reactions, balancing caution with cooperation. OpenAI became more forgiving across the board. Claude leaned further into reciprocal behavior, often opting to resume cooperation even after being exploited. These distinctions were visible in the models’ conditional cooperation probabilities, a core metric the researchers used to track strategic style.
This variation could have consequences in real-world scenarios. An overly cooperative model may underperform in adversarial negotiations, while one that defects too quickly might damage long-term relationships. Understanding these tendencies will be important as AI systems take on more complex decision-making roles in business, policy, and social interaction.
The researchers argue that future assessments should move beyond task benchmarks to examine behavioral tendencies under stress, risk, and uncertainty. The findings indicate that language models are already making choices with strategic depth, sometimes resembling human reasoning, and sometimes diverging sharply from it.
Note: This post was edited/created using GenAI tools. Image: DIW-AIgen
Read next: More Americans Are Turning to AI for Tax Filing Help
The study placed AI models into evolutionary simulations against classic hand-coded strategies such as Tit-for-Tat, Grim Trigger, and Win-Stay-Lose-Shift. These matches introduced noise, randomized game lengths, and mutation, forcing agents to adapt their play styles without relying on fixed responses. Across more than 30,000 matchups, the AI models generated their own reasoning for each move, offering an unusual window into how they interpret uncertainty and opponent behavior.
Behavior That Reflects Strategy, Not Just Training
Google’s Gemini models emerged as the most tactically responsive. When the probability of a match ending early increased, Gemini reduced its cooperative moves and prioritized immediate payoffs. In contrast, OpenAI’s models maintained high levels of cooperation across nearly all conditions, even when it left them vulnerable. Anthropic’s Claude model proved the most forgiving, often returning to cooperative behavior after being exploited.These patterns were not superficial. Each model’s decisions were accompanied by written rationales, many of which revealed a sensitivity to the opponent’s behavior and the remaining duration of the game. The researchers noted that models often adjusted their play based on perceived opponent types, suggesting a capacity for adversary modeling. In many cases, they also adjusted based on how long the match was likely to last, a factor known in game theory as the “shadow of the future.”
Also read: New Research Shows Language Choice Alone Can Guide AI Output Toward Eastern or Western Cultural Outlooks
Gemini’s approach was especially adaptive. In long games with low termination probability, it cooperated to build trust. In shorter matches or those with frequent random resets, it defected more often. This flexibility allowed Gemini to outperform other agents in unstable environments, including one harsh condition with a 75% chance of early termination. There, Gemini nearly wiped out the competition by shifting into full defection mode.
OpenAI’s strategy, however, remained largely consistent. Its models continued to cooperate even when the environment favored defection. In the most unforgiving condition, OpenAI’s agent cooperated in over 95% of moves, leading to rapid elimination. The researchers pointed to a tendency within OpenAI’s models to overlook time-based incentives, often reasoning in generalized terms about cooperation rather than calculating expected outcomes.
Claude, meanwhile, occupied a middle ground. Its willingness to forgive past defections helped it survive in tournaments with longer time horizons. Although it did not dominate the field, its consistent behavior produced competitive results, especially when paired against agents that valued reputation and long-term payoff.
Tournament Design Focused on Evolution and Adaptation
To stress-test the models, the researchers ran seven tournaments across different conditions. These included variations in model sophistication, game length expectations, and population volatility. In one tournament, a random strategy was periodically reintroduced to disrupt any emerging equilibrium. In another, only advanced LLMs from the three companies competed head-to-head.In each case, the agents' survival depended on their average score per move. Higher-performing strategies reproduced into the next round, while weaker ones were phased out. This setup allowed the researchers to simulate evolutionary pressure and observe which behavioral patterns were most viable under different constraints.
The outcomes showed that strategic fingerprints varied not only across companies but also between model versions. For example, the more advanced version of Gemini was more consistent in cooperating when beneficial and defecting when not, compared to its earlier variant. OpenAI’s models remained uniformly cooperative, regardless of capability level. Claude demonstrated stability, with moderate adjustments depending on the game dynamics.
Reasoning Reveals Divergent Cognitive Priorities
A qualitative analysis of the models’ written explanations revealed stark differences in how they weighed time and opponent behavior. Gemini focused heavily on calculating remaining rounds and adjusted its strategy accordingly. In some cases, this led to aggressive shifts toward defection when long-term gains were unlikely. OpenAI referred to the passage of time less frequently and was more likely to use general terms like “building trust” or “fostering cooperation,” even when the match was likely to end soon.In environments with noise or mutation, Gemini altered its forgiveness rate and moderated its reactions, balancing caution with cooperation. OpenAI became more forgiving across the board. Claude leaned further into reciprocal behavior, often opting to resume cooperation even after being exploited. These distinctions were visible in the models’ conditional cooperation probabilities, a core metric the researchers used to track strategic style.
Implications for AI Deployment
The study suggests that AI models are not interchangeable tools. Each brings a behavioral profile shaped by architecture, training data, and fine-tuning. While OpenAI’s models favor cooperation, even when it leads to poor outcomes, Gemini adapts more quickly to environmental shifts. Claude displays a strong preference for forgiveness, which helps in extended or trust-heavy contexts.This variation could have consequences in real-world scenarios. An overly cooperative model may underperform in adversarial negotiations, while one that defects too quickly might damage long-term relationships. Understanding these tendencies will be important as AI systems take on more complex decision-making roles in business, policy, and social interaction.
The researchers argue that future assessments should move beyond task benchmarks to examine behavioral tendencies under stress, risk, and uncertainty. The findings indicate that language models are already making choices with strategic depth, sometimes resembling human reasoning, and sometimes diverging sharply from it.
Note: This post was edited/created using GenAI tools. Image: DIW-AIgen
Read next: More Americans Are Turning to AI for Tax Filing Help
