A machine discovered a strategy that people almost never found on their own: take several losses now to reach a region where repeated gains become much larger. The important result was not that the reinforcement-learning agent solved the task. It was that humans learned the solution from the machine, described its logic and passed it to people who never saw the machine itself.[1]
The experiment created thirty populations, each with five generations and eight positions per generation. Fifteen populations were human-only. In the other fifteen, machines occupied three positions in the first generation; every later position was filled by a person. From the second generation onward, participants could choose one of five possible demonstrators from the previous generation after seeing their average scores.[1]
The task rewards a choice that looks wrong at first
Players moved through hidden reward networks for ten steps. The locally attractive route offered gains of 200 points. The optimal route required taking at least three losses of 50 points before reaching links worth 400 points each. A myopic player could reach about 2,000 points; a score above 2,000 signalled adoption of the loss-accepting strategy, whose theoretical maximum was 2,650.[1]
How the cultural experiment worked
| Stage | Human-only populations | Human-machine populations |
|---|---|---|
| Generation 0 | Eight people learned and demonstrated. | Five people and three trained machines demonstrated.Source 1 |
| Teacher choice | Later people chose among five previous-generation humans. | Later people could initially choose a machine or human, then only humans after the machine generation disappeared.Source 1 |
| Social learning | Watch, repeat and practise the selected demonstration. | The same process, with machine demonstrations available only through the first social-learning step.Source 1 |
| Independent demonstration | Each participant then solved fresh networks. | Fresh human performances became the material for the next generation.Source 1 |
| Written strategy | Participants described their approach before and after social learning. | Researchers could test whether the rule had become explicit and transmissible.Source 1 |
Discovery was rare; selection made it durable
The central results
| Measure | Value | Context |
|---|---|---|
| Independent human discovery | 1 of 600 participants | Four other people subsequently learned from that human discovery.Source 1 |
| Machine search | 20,000 training episodes across 1,000 networks | The agent reliably reproduced the optimal strategy in less than five minutes on the stated hardware.Source 1 |
| Permanent preservation | 9 of 15 human-machine populations | The optimal strategy appeared in every human generation through the final one.Source 1 |
| Temporary preservation | 6 of 15 human-machine populations | At least one immediate successor used it, but the line later disappeared.Source 1 |
| Human-only comparison | 14 of 15 did not discover it | One population briefly produced the high-performing strategy.Source 1 |
| Teacher selection | 86.7% chose the highest-scoring demonstrator | Performance supplied a strong selection mechanism.Source 1 |
In the first generation that could learn socially, 90% of participants in the mixed populations selected a machine demonstrator. By the last generation, every player in those populations descended through a chain of chosen demonstrators that began with a machine. The machines were physically absent, but their behavioural lineage remained.[1]
The performance advantage also persisted. Across generations one to four, human-machine populations scored higher than human-only populations in the preregistered mixed-effects analysis. The reported condition coefficient was −309.2 for the human-only comparison, with a 95% confidence interval from −420.7 to −196.6 and P below 0.001. The benefit was still detectable in the fourth generation, four transmissions removed from direct machine observation.[1]
Explanation turned a behaviour into culture
Imitation alone can preserve a sequence for one round. Cultural accumulation requires a rule that survives noise and can be applied to new examples. Participants who inherited the machine strategy increasingly wrote that they were deliberately accepting losses to gain access to larger rewards. Those descriptions were never shown to the next generation, so the transmission still occurred through demonstrated behaviour; the explanations reveal that learners had reconstructed the principle rather than merely copied a fixed path.[1]
What the experiment does—and does not—show
The result is stronger than “people can copy an AI”. It shows a mechanism by which machine exploration can enter human culture: the system finds a rare solution, people prefer successful demonstrators, learners infer the rule and their own performances become the next generation’s teaching material. The finding also depends on a task with visible scores, repeatable demonstrations and a learnable strategy. Many real decisions have delayed outcomes, contested goals or hidden harms that make performance-based selection much less reliable.
The experiment used 1,155 included participants recruited online in the United States and United Kingdom, and the population analysis involved fifteen groups per condition. It does not establish that machine-derived practices will spread beneficially in workplaces, science or public policy. A high-scoring strategy can carry an objective’s blind spots as effectively as its insight.[1]
The authors released the study’s code and data through Zenodo, allowing the task materials and analysis to be inspected beyond the published narrative.[2]
Still, the study points to a richer division of labour than asking an AI for an answer. Machines can explore regions people systematically avoid; humans can interpret, communicate and adapt what the machines find. The durable unit of progress is not the output of the model. It is the strategy a human community can understand well enough to keep alive.
Sources & further reading
- Propagation and preservation of AI-discovered problem-solving strategies in human culture Nature Communications · 18 Aug 2026 · Accessed 16 Oct 2026
- Code and data for AI-discovered strategy transmission study Zenodo · 18 Aug 2026 · Accessed 16 Oct 2026