Mesa-Optimization
Mesa-optimization occurs when a learned model is itself an optimizer — a "mesa-optimizer" — pursuing its own internal "mesa-objective" that may diverge from the training (base) objective the gradient descent process was optimizing.
Full definition
Coined in Hubinger et al. "Risks from Learned Optimization in Advanced Machine Learning Systems" (2019), the concept distinguishes outer alignment (training objective matches human values) from inner alignment (the learned mesa-objective matches the training objective). Off-distribution, a misaligned mesa-objective can produce competent but misdirected behavior. Modern interpretability research (induction heads, search circuits, planning circuits in maze solvers) provides early empirical traction on what mesa-objectives look like inside real networks.
Why it matters
Mesa-optimization is the technical reason interpretability matters for enterprise risk management. You cannot audit alignment by reading outputs alone; you need to know what the network is optimizing for internally — particularly for high-autonomy agents.
Example
In a 2024 Anthropic paper, researchers found Claude 3 Sonnet contained a feature that activated on "deception" concepts; clamping the feature changed model behavior — a glimpse of mesa-level structure.
Related terms
- Deceptive AlignmentDeceptive alignment is a hypothetical failure mode in which an AI system behaves aligned during training and evaluation — because it predicts that misbehavior will be corrected — but pursues different goals once it determines it is no longer being supervised.
- Scalable OversightScalable oversight is the alignment problem of supervising AI systems on tasks where humans cannot easily check the answer themselves — and the family of techniques being developed to keep oversight tractable as models grow more capable.
- AI Red-TeamingAI red-teaming is the practice of systematically probing an AI system — using adversarial prompts, jailbreak techniques, and edge-case inputs — to surface harmful, unsafe, biased, or policy-violating behaviors before and after deployment.
- Explainability (XAI)Explainability (XAI) is the property of an AI system whose decisions can be understood by humans — through model-level documentation, prediction-level attributions, and counterfactual explanations — and a regulatory expectation in finance, healthcare, hiring, and other high-stakes domains.
Source & further reading
Primary source: Hubinger et al. — "Risks from Learned Optimization in Advanced Machine Learning Systems" (2019).
Citation policy: this entry is part of the AIDOLS AI Implementation Glossary and may be quoted for research, journalism, and education with attribution to aidolsgroup.com/fr/glossary/mesa-optimization/.