Artificial intelligence can develop solutions that humans would be unlikely to discover on their own. But what happens when machines do more than solve individual problems and actually discover entirely new solution strategies? And can such discoveries become part of human knowledge?
A team from the Center for Humans and Machines at the Max Planck Institute for Human Development, together with researchers from the Toulouse School of Economics and Humboldt University of Berlin, investigated this question experimentally for the first time. The results show that, under certain conditions, humans can not only adopt strategies discovered by machines but also pass them on and preserve them across multiple generations.
The study was motivated by the observation that modern AI systems are increasingly developing strategies that surprise human experts. This phenomenon became well known through AI systems’ performance in board games such as Go and chess, where they discovered unusual yet highly successful ways of playing. Until now, however, it was unclear which conditions are required for such machine discoveries to have a lasting influence on human thinking and behavior.
The researchers argue that three conditions must be met: the strategy must initially be difficult for humans to discover, it must remain learnable, and its advantage must be clearly recognizable. Only then can it spread within a community and be maintained over time.
“Human cultural evolution depends on knowledge being transmitted across many individuals and generations,” says lead author Levin Brinkmann, a research scientist at the Center for Humans and Machines at the Max Planck Institute for Human Development. “If machines are now discovering novel strategies, the question is whether these can also become part of the knowledge humans pass on. That is exactly what we investigated.”
Testing cultural transmission in the laboratory
To test their hypothesis, the researchers conducted a behavioral experiment involving 1,155 participants. Participants were asked to earn as many points as possible in specially designed reward-based tasks. The task was structured so that the optimal strategy initially required participants to accept small early losses in order to achieve much larger gains later. This ran counter to a common human tendency: avoiding short-term losses whenever possible.
The participants were organized into groups that learned from one another across multiple generations. Some groups consisted entirely of humans. In other groups, AI agents were also present at the beginning. These agents had previously completed the task using a learning algorithm and had discovered the optimal strategy.
Participants could observe successful solutions from the previous generation and use them to guide their own decisions. This made it possible to investigate whether a strategy discovered by machines would be adopted by humans, passed on, and maintained over the long term.
Humans learn from successful role models
The results were clear: in groups without AI, the optimal strategy was almost never discovered. Out of 600 participants, only one person independently found the best solution. In contrast, in groups that included AI, the strategy discovered by the machine frequently spread and was preserved across generations. In nine out of fifteen human-machine groups, this strategy remained in use throughout the experiment.
Participants showed a strong tendency to follow the most successful role models. Most chose the person or agent with the best results as a model for their own behavior. In this way, the strategy discovered by the machine was able to spread and be preserved over time.
A new perspective on the role of AI
The findings suggest that intelligent machines could become more than mere tools for automation or computation. If machines discover problem-solving strategies that humans can understand and pass on, they could contribute to the long-term expansion of human capabilities.
“People often discuss whether AI will make humans more dependent,” says Thomas Eisenmann, co-lead author of the study and a research scientist at the Center for Humans and Machines at the Max Planck Institute for Human Development. “Our results point to another possibility: if machine discoveries are comprehensible, they can empower people to develop new skills, preserve them over time, and expand our cultural repertoire.”
This study provides empirical support to the concept of ‘machine culture’. Many of today's most advanced AI capabilities stem from machines learning human knowledge and culture. Large language models, for example, which are trained on vast collections of human-written text, can process and generate fluent text on nearly any topic. The study's results show that this is not a one-way street: humans can similarly bootstrap their capabilities from machine discoveries. The consequence is the beginning of the age of machine culture, in which humans and machines are joined in a single cultural evolutionary process — each a source of innovation for the other, and each capable of building on what the other discovers.
At a glance :
Nature Communications
Propagation and preservation of AI-discovered problem-solving strategies in human culture
18-Aug-2026