Human–AI Coevolution

Course description

The rise of socio-technical systems, where humans interact with AI assistants and recommenders, poses risks of unintended consequences. Navigation apps like Google Maps may cause congestion by directing too many drivers to the same route; profiling and targeted ads can reinforce biases and inequality; and generative AI chatbots risk declining quality as synthetic data increasingly drives retraining. These issues stem from Machine Learning models trained on human behaviour, creating feedback loops that shape future choices and data. This course equips students to analyse and model these loops, designing experiments to assess their societal impact and foster responsible AI development.

Learning outcomes

• Design experiments to assess the impact of AI algorithms in human–AI ecosystems.
• Measure and interpret the impact of AI algorithms on individuals and society.
• Model feedback loops between humans and AI algorithms.

Assessment

The assessment (“esame”) is optional. Students who wish to take the exam must submit, within two weeks after the end of the course, a short project proposal outlining a possible empirical or simulation-based study aimed at assessing the impact of AI systems in a given ecosystem (e.g., social media, online retail, urban mapping, generative AI).

Students are not required to perform the full analysis. Instead, the goal is to formulate a well-structured and theoretically grounded proposal that:

1. Identifies the dimension(s) of impact to be investigated (individual and/or collective);
2. Describes the experimental or simulation design, including data requirements, methodology, and evaluation metrics;
3. Argues for the feasibility of the study, explains why the proposed analysis is scientifically and socially relevant, and dicusses the ethical aspects related to the study;
4. Summarises key findings from existing literature that motivate and support the proposal.
5. Prepares a data access request for Very Large Online Platforms (VLOPs) or Very Large Online Search Engines (VLOSEs) under the EU Digital Services Act (DSA), following the procedure outlined at https://data-access.dsa.ec.europa.eu/home

The project will be discussed during a short oral presentation, in which the student will justify the design choices and demonstrate the potential contribution of the proposed study.

Readings

1. Pedreschi et al., Human-AI Coevolution, Artificial Intelligence, Volume 339, February 2025, 104244. DOI
2. Pappalardo et al., A survey on the impact of AI-based recommenders on human behaviours: methodologies, outcomes and future directions, arXiv:2407.01630. arXiv
3. Huszár et al., Algorithmic amplification of politics on Twitter, PNAS 119(1) e2025334119 (2021). DOI
4. Cornacchia et al., How routing strategies impact urban emissions, SIGSPATIAL ’22, Article 42, pp. 1–4 (2022). DOI

Course syllabus (sessions)

1. Introduction to Human–AI Coevolution. Overview of the course topic, delivered as a lecture with interactive discussions.
2. The Human–AI Feedback Loop (FL). Exploration and formalisation of feedback loops in ML models, with examples from social media, online retail, urban mapping, and chatbots.
3. Types of Experiments. Overview of experimental methods to evaluate feedback-loop impacts, illustrated with examples from diverse human–AI ecosystems.
4. FL effects: Social Media. Case studies on platforms like Facebook, Twitter, and YouTube.
5. FL effects: Online Retail. Real-world examples on platforms like Amazon and Spotify.
6. FL effects: Urban Mapping. Effects on platforms such as Google Maps and Airbnb.
7. FL effects: Generative AI. Challenges posed by AI-generated chatbots like LLaMA and ChatGPT.
8. Tutorial: How to prepare a data access request for VLOPs or VLOSEs under the EU Digital Services Act (DSA)
9. Open Challenges in Human–AI Coevolution. Unresolved technical, legal, and political issues in measuring human–AI coevolution.
10. Practice Session. Hands-on activity: designing experiments to assess impacts of AI recommenders (exam preparation).
11. Exam / Presentations. Student presentations and evaluation (for those who need the make the exam).

Calendar of Lessons