Experimental Evidence for the Propagation and Preservation of Machine Discoveries in Human Populations - Code and Data

Code and data accompanying the paper *"Propagation and preservation of AI-discovered problem-solving strategies in human culture"* by Levin Brinkmann, Thomas F. Eisenmann, Anne-Marie Nussberger, Maxime Derex, Sara Bonati, Valerii Chirkov and Iyad Rahwan. The study asks under which conditions strategies discovered by intelligent machines are adopted, understood and preserved by human populations. Human participants solved a reward-network task in a cultural transmission experiment, seeded with solutions from a deep reinforcement-learning agent; an agent-based model extends the findings beyond the tested conditions. This archive contains everything needed to reproduce the results: - **Data** — de-identified experimental data at the participant and move level, both raw (as downloaded from the experiment interface) and processed with derived variables; manually coded written strategies; the reward networks together with machine and heuristic solutions; trained model weights and training metrics; and agent-based-model outputs. - **Algorithm** — PyTorch code for training the deep Q-network machine player on the reward-network task. - **Experiment** — the React frontend and Flask backend used to run the online experiment (a static version of the interface is available at https://center-for-humans-and-machines.github.io/reward-network-iii). - **Analysis** — Jupyter notebooks for data processing and figures, R scripts for the statistical analyses, and the agent-based model. Installation and usage instructions are in `README.md`, with further details in the per-directory READMEs. Tested with Python 3.10.