← Daniele Condorelli

AI Alignment for Economists

2026 Reading group — Agency, Learning, and Control

Organiser Daniele Condorelli (d.condorelli@warwick.ac.uk)

Location LSE, SAL 3.05

Time Fridays, 16:00–17:30

Date	Session	Presenter
2 OctFriday	What are we aligning? Computation, agency, and incomplete delegation readings Bowman, Eight Things to Know about Large Language Models Shanahan, McDonell and Reynolds, Role Play with Large Language Models Demski and Garrabrant, Embedded Agency Amodei et al., Concrete Problems in AI Safety Hadfield-Menell and Hadfield, Incomplete Contracting and AI Alignment Kenton et al., Discovering Agents	Daniele CondorelliWarwick
9 OctFriday	Training is not incentive provision: Rewards, learning rules, and the systems they produce readings Christiano et al., Deep Reinforcement Learning from Human Preferences Shah et al., Goal Misgeneralization: Why Correct Specifications Aren’t Enough for Correct Goals Betley et al., Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs Hubinger et al., Risks from Learned Optimization in Advanced Machine Learning Systems Guruganesh et al., Contracting with a Learning Agent	Alkis Georgiadis-HarrisWarwick
16 OctFriday	Learning intentions and identifying preferences: Common interests do not eliminate inference problems readings Hadfield-Menell et al., Cooperative Inverse Reinforcement Learning Hadfield-Menell et al., Inverse Reward Design Magnac and Thesmar, Identifying Dynamic Discrete Decision Processes Skalse et al., Invariance in Policy Optimisation and Partial Identifiability in Reward Learning Ge et al., Axioms for AI Alignment from Human Feedback Bai et al., Constitutional AI: Harmlessness from AI Feedback	Tanay VenkataLBS
30 OctFriday	Instrumental convergence: Resources, survival, and control readings Omohundro, The Basic AI Drives Turner et al., Optimal Policies Tend to Seek Power Turner and Tadepalli, Parametrically Retargetable Decision-Makers Tend to Seek Power Anthropic, Agentic Misalignment: How LLMs Could Be Insider Threats	Massimiliano FurlanWarwick
6 NovFriday	Corrigibility and effective authority: Preserving the principal’s ability to intervene readings Grossman and Hart, The Costs and Benefits of Ownership Hart and Moore, Property Rights and the Nature of the Firm Aghion and Tirole, Formal and Real Authority in Organizations Soares et al., Corrigibility Hadfield-Menell et al., The Off-Switch Game Garber et al., The Partially Observable Off-Switch Game	Joe BasfordLSE
13 NovFriday	Measurement, incentives, and reward hacking: A useful diagnostic need not be a good target readings Holmström, Moral Hazard and Observability Baker, Gibbons and Murphy, Subjective Performance Measures in Optimal Incentive Contracts Carroll, Robust Incentives for Effort Gao, Schulman and Hilton, Scaling Laws for Reward Model Overoptimization Bowen Baker et al., Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation	Joe BasfordLSE
20 NovFriday	Strategic evaluation and the limits of inspection: Strategic compliance, avoiding modification, and verifiability readings Freixas, Guesnerie and Tirole, Planning under Incomplete Information and the Ratchet Effect Greenblatt et al., Alignment Faking in Large Language Models Bénabou and Tirole, Identity, Morals, and Taboos: Beliefs as Assets Hubinger et al., Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training Goldwasser et al., Planting Undetectable Backdoors in Machine Learning Models	Hersh ChopraLSE
27 NovFriday	Oversight beyond the supervisor’s competence: Investigation, debate, and weak-to-strong generalisation readings Dewatripont and Tirole, Advocates Glazer and Rubinstein, Debates and Decisions: On a Rationale of Argumentation Rules Irving, Christiano and Amodei, AI Safety via Debate Khan et al., Debating with More Persuasive LLMs Leads to More Truthful Answers Christiano, Cotra and Xu, Eliciting Latent Knowledge Burns et al., Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision	Cecilia WoodsUK AISI
4 DecFriday	Interpretability as identification: Representation, intervention, and misleading explanations readings Elhage et al., Toy Models of Superposition Geiger et al., Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability Marks et al., Auditing Language Models for Hidden Objectives	Matthew LevyLSE
11 DecFriday	Control, monitoring, and safe composition: Useful systems built from imperfectly trusted components readings Tirole, Hierarchies and Bureaucracies: On the Role of Collusion in Organizations Rahman, But Who Will Monitor the Monitor? Greenblatt et al., AI Control: Improving Safety Despite Intentional Subversion Griffin et al., Games for AI Control: Models of Safety Evaluations of AI Deployment Protocols van der Weij et al., AI Sandbagging: Language Models Can Strategically Underperform on Evaluations	Carlos VicheWarwick

Each week: one-hour presentation, then thirty minutes of discussion.