Human-AI and Collective Action
Economics for Everyone 6 · Nobel laureate: Elinor Ostrom (2009), for her analysis of the governance of the commons.
We build knowledge organizations: teams of people and artificial intelligence (AI) agents that decide with evidence and produce frontier evidence, using free or low-cost tools.
Guiding question. If a machine can already carry out the full cycle of a research project on its own, what does a team of people and agents need to produce evidence that others can trust, and to make decisions with it?
Syllabus | Tools | Course folder · Course materials (in Spanish): Teaching
The AI Scientist research cycle
The course follows the research cycle of The AI Scientist (Lu et al., Nature, 2026), an agent-based system that completed every stage of a research project and wrote a paper that passed peer review at a scientific workshop. Each week covers one stage:
Idea: from an organization's decision to the research question and hypotheses.
Novelty: verified literature and a synthesis of what is already known.
Experiment: design, preregistration, agent-based piloting, execution, and analysis with code.
Writing: writing with AI and disclosing it.
Review: automated reviewer and referee panels, publication, and replication.
Machines fail in known ways: they propose naive or already published ideas, fabricate citations, and make coding errors. Each stage therefore adds what AI cannot do on its own: verifying, judging, protecting people's data, and governing the team.
Research as collective action
Knowledge and data are a commons (Hess and Ostrom, 2007). Checking a citation, replicating a result, or reviewing a paper benefits everyone but costs the individual, so it is done less often than it should be. When a machine does part of the work, new free-riding and credit problems arise. Drawing on Ostrom's design principles, each team agrees on its own rules:
Clear boundaries: who belongs to the team, which agents work on it, and with which data.
Own rules: how credit is shared between people and agents.
Monitoring: an AI use log and verification of every citation and every number.
Graduated sanctions and conflict resolution, agreed on by the group.
Each session is organized around an open duality, such as Should we automate the research cycle or govern it as a commons? We test both sides with the data from that day's game.
Free and low-cost tools
With the Fundación Memories team, we explore tools that any group can afford, and in class we discuss the theory behind each one and how to apply it to real research. No paid subscriptions or programming skills are required: notebooks come ready to modify.
Getting organized: OpenClaw, an open-source agent, and Slack on its free plan.
Question and ideas: the 3ie Development Evidence Portal and local open models with Ollama.
Literature: OpenAlex, Semantic Scholar, Crossref, and Zotero.
Synthesis and analysis: metafor and ppi_py in Google Colab.
Design and execution: the Manning, Zhu, and Horton code in GitHub Codespaces, oTree, OSF Registries, and VeraCrypt.
Writing, publication, and replication: Quarto or Overleaf, OSF Preprints, Zenodo, and GitHub.
Project: participatory action research
Each group of 2 to 4 students, each working with their own agent, carries out a participatory action research (PAR) project with a real organization of its choice: social, community-based, public, or student-led. PAR was developed by the Colombian sociologist Orlando Fals Borda: those who live the problem are not objects of study but co-researchers, and knowledge is produced to transform reality together with them (Fals-Borda, 1987). It is the other side of Ostrom: the rules of research (who participates, who decides, who gets credit, and whom the data serve) are agreed on with the organization.
Following the cycle, the group and the organization:
diagnose how the organization makes decisions and produces evidence;
co-design the adoption of low-cost AI tools;
test it in a real decision;
assess what changed, with evidence, and return the results to the organization.
Before starting, the group signs a participation agreement and obtains informed consent. No personal data are entered into an AI model or appear in public deliverables. The project advances through two practical midterms and ends with a short paper, its replication package, and a public defense with the organization invited.
Assessment
The course uses the common assessment scheme shared by the six Fundación Memories courses.
Unannounced quizzes with true-or-false items, discussed in class right afterward (5%).
Problem sets with solutions (15%) and group AI challenges (5%).
Written midterm (15%) and cumulative final exam (25%) with argued true-or-false items, without AI or notes: a point counts only if both the answer and the justification are correct.
Oral defense: if the answer is correct but the justification is incomplete, the student completes the argument in a brief oral defense with the instructor.
Practical midterms of the project (10% each) and final project (15%), graded individually on the declared contribution and the presentation.
In AI challenges, practical midterms, and the project, any AI tool is allowed as long as its use is declared: "we used X for Y."
Our principles for using AI
AI produces pilots, not findings. Simulated subjects and model forecasts help debug designs; they are never reported as results.
Every reference and every number is verified against Crossref or the publisher, and against the table or code that produced it.
A person approves everything that leaves the team.
Participants' data never reach an AI service, a chat channel, or a video. Videos use demo projects and fictional data.
We record what we ran: model, version, date, and prompt for every simulation or forecast.
We disclose AI use and remain accountable for it. AI is never an author. Disclosure template.
Videos: one per topic
Each content session has a short video, 3 to 6 minutes long, that shows the tool for its stage and what AI cannot do on its own. The session 7 video is already available: