The Limits of Good Will
Ten Lectures on Strategic Interaction, Cooperation, and the Architecture of Decision
About this course
“Talk to each other.” “Punish the cheaters.” “Let the market decide.” “Decentralize everything.” Every social problem has its piece of popular wisdom — a common-sense answer that seems to dissolve the problem before it has been properly stated. This course asks what happens to those common-sense answers when they are confronted with the experimental and computational record.
The answer is almost never that the intuition is wrong. The answer is usually that it is bounded. Cheap talk works — under six specific conditions. Graduated sanctions work — within a specific ladder. Polycentric governance works — until the externalities cascade. The interesting object of each lecture is not the intuition itself, nor its refutation, but the perimeter inside which it survives.
The course is also the third altitude of the BDPD ecosystem: the research papers run the laboratory experiments, The Forest of Humbaba card game lets players experience the dynamics intuitively, and these ten lectures sit in the middle, where the empirical record of sixty years of social-dilemma research meets the questions that the new generation of AI agents forces us to ask.
The ten lectures
Each lecture follows the same five-section pattern: good-will intuition → what the classics actually say → what survives → the BDPD angle → mini-challenge. The four narrative blocks below tell a coherent story: the easy answers fall first, the institutional ones turn out to be fragile, the technical ones depend on measuring the right thing, and at the end the deeper problem of time emerges.
Block A — Communitarian answers
1. Just Talk to Each Other — Cheap talk and cooperation in social dilemmas. Sixty years of experiments confirm that talking solves social dilemmas — a +45 percentage-point effect, robust and replicated. So why does the same intervention produce almost nothing when the agents around the table are not human?
2. We Are Enough Among Ourselves — Olson and group size. Olson argued that small groups cooperate and large groups don’t. The argument has fed half a century of policy claims about subsidiarity, community-based management, and the magic of locality. What happens when the “group” includes members who can clone themselves on demand?
3. Decentralize Everything — Polycentric governance and cascades. Ostrom built her career on the empirical case that many local centres of decision beat one centralised authority. Then the externalities cascade across the boundaries — and the centres that look most autonomous turn out to depend most on what their neighbours do.
Block B — Institutional answers
4. Let’s Punish the Cheaters — Sanctions and altruistic punishment. Costly punishment is the standard remedy for free-riding, and it works, modestly, when humans wield it. What does a graduated sanctions ladder look like when the punished do not feel cost — and when the punishers can compute optimal evasion strategies in milliseconds?
5. Democratic Voting — Arrow and the limits of aggregation. Arrow proved that no voting system can satisfy a small list of fairness criteria that every voter would endorse separately. The impossibility theorem is a fixture of social choice; what changes when the voters are language-model agents with introspectable preference orderings?
6. The Market Knows — Externalities, Pigou and Coase. Pigou said tax the externality. Coase said let the parties bargain. Both arguments have powered modern environmental and regulatory policy for decades. Both presuppose a kind of agent that may not be the agent now showing up at the negotiating table.
Block C — Technical answers
7. We Need a Rule — Goodhart and normative rigidity. Goodhart’s Law says every measure becomes a target. The folklore of public administration treats it as a quirk to be managed. BDPD finds that rule-bound governance is not just inefficient under endogenous dynamics — it is systematically beaten by signal-adaptive alternatives. Which rules survive that test?
8. More Information Equals Better Decisions — Informational cascades. The intuitive case for transparency, openness, and information abundance is that decisions get better when actors know more. The information-cascade literature shows that the opposite can hold: rational agents who see each other’s choices converge on the wrong one and stay there. AI agents make this faster, not better.
9. If We Meet Again, We Cooperate — The folk theorem and its bad equilibria. The folk theorem of repeated games says almost any outcome — including full cooperation — can be sustained as an equilibrium if the players are patient enough. The same theorem says almost any outcome — including full mutual exploitation — can also be sustained. The brochure version of the folk theorem leaves out the second half.
Block D — The final limit
10. We’ll See the Danger Coming — Leading vs lagging signals and the Seneca effect. Bardi’s Seneca effect: slow up, fast down. Most regulatory systems wait for the fast-down signal to act, which is by construction too late. BDPD shows that the binding question is not which lever the regulator pulls but which signal triggers them — and the answer reorders sixty years of policy advice on climate, fisheries, and any system with hidden inertia.
How to read this course
The lectures can be read in order or stand-alone. Each is self-contained: skipping previous blocks loses the narrative thread but not the technical prerequisites. The natural prerequisite profile is a final-year undergraduate or master’s student in social sciences (economics, political science, psychology), computer science or data science with openness to social-science material, or philosophy with a quantitative orientation. The non-substitutable prerequisites are an introductory course in game theory and an introductory course in statistics.
Each lecture is paced for two academic hours of structured study — roughly one hundred effective minutes. Embedded mini-challenges at the end of each lecture assume access to a local BDPD installation (see docs/experiments/reproduce.md); readers who cannot execute them will still follow the lecture, but will miss the experience the ladder — from research paper to classroom to tabletop — is designed to deliver.
Two structural notes before you begin. (i) The type of mini-challenge varies across the course. Lectures whose BDPD angle is empirical — where the platform has already executed the relevant experiment — close with a runnable challenge against the local install. Lectures whose BDPD angle is speculative or extrapolative — where the experiment has been designed but not yet run — close with a thought-experiment or design exercise. Both styles are deliberate; the lecture states its type in the opening sentences of the BDPD-angle section. (ii) The length of the BDPD-angle section scales with the type of angle. Empirical lectures may run roughly twice the length of speculative ones in that section, because they have numerical results to report. This is a structural feature of the five-section template, not a defect in either kind of lecture.
A two-week intensive workshop covering all ten lectures plus mini-challenges runs to roughly thirty-two hours of student time, equivalent to two or three ECTS.
The full project — code, paper sources, card-game artwork, and the documentation site — is available on GitLab. Code is AGPL-3.0; papers, course material, and game rules are CC BY 4.0; AI-generated card art is CC0.
Transparency and authorship. This document was authored by Roberto Brunelli with the support of frontier large language models; model outputs were source-grounded by prompt, validated by extensive triaging with different models and finally integrated under author supervision.