Human-AI coevolution SES Lab’s Journal Club Calendar Momoha Hirose Posted on 06/18/2025 

Paper Information 1 Title: Human-AI coevolution Authors: Pedreschi et al. Publication: Artificial Intelligence, 339, Article 104244 (2025) DOI: https://doi.org/10.1016/j.artint.2024.104244 Pedreschi, D., Pappalardo, L., Ferragina, E., Baeza-Yates, R., Barabási, A.-L., Dignum, F., Dignum, V., Eliassi-Rad, T., Giannotti, F., Kertész, J., Knott, A., Ioannidis, Y., Lukowicz, P., Passarella, A., Pentland, A. S., Shawe-Taylor, J., & Vespignani, A. (2025). Human-AI coevolution. Artificial Intelligence, 339, 104244. 

The interaction between users and AI, especially recommender systems, creates a powerful and potentially endless feedback loop. Background | The Core Concept: Human-AI Feedback Loop 2 The Process: User’s choices on online platforms produce Training Data. This data is used to feed and re-train Recommenders. Recommenders provide Suggestions to users. These suggestions influence the next set of users' choices, which in return generate more data for re-training. 1. 2. 3. 4. 

While human-technology coevolution is not new, AI-based recommenders have simultaneously magnified five key aspects: Pervasiveness: Integrated into all major online platforms, from social media to mapping services. Persuasiveness: Highly personalized due to abundant data on individual choices, making them highly effective in capturing users' preferences. Traceability: Human choices and AI suggestions leave a massive digital trace. Speed: AI can be re-trained and provide suggestions at an unparalleled speed, often with little to no human oversight. Complexity: Fosters a huge volume of interactions between massive numbers of users and products, escalating system complexity. Background | What Makes this Coevolution Unprecedented? 3 

The paper establishes human-AI coevolution as the cornerstone for a new field of study at the intersection of AI and complexity science. The study of this coevolution enriches existing AI debates by adding a new dimension, which the authors define as Society-Centred AI. Proposing a New Perspective: Society-Centred AI 4 

The feedback loop produces outcomes at different levels: i.e., individual, item, model, and systemic Systemic Outcomes: Polarization, echo chambers, inequality, concentration, and segregation. Examples of human-AI ecosystems and its systemic outcomes: Social Media: Recommenders can create filter bubbles and amplify political polarization. Online Retail: Can increase sales volume but also decrease overall product diversity, amplifying the success of popular items. Urban Mapping: Individually optimal routes can lead to collective chaos, increasing traffic congestion and emissions. Content Generation (LLMs): Training on AI-generated content can lead to "model collapse" or "autophagy," a loss of linguistic diversity. Outcomes of Human-AI Coevolution 5 

Open Challenges & Conclusion 6 The paper outlines challenges at three levels: Scientific: Measuring the feedback loop and investigating bi-directional causality. Legal: Ensuring data access for researchers and transparency. Socio-political: Addressing the concentration of "the means of recommendation". The authors conclude that progress requires systematic feedback-loop measurement together with coordinated scientific, legal, and socio-political efforts, so as to steer human–AI coevolution toward society-centric AI and avert the negative externalities of an uncontrolled feedback loop.