Sycophancy Has an Address: Mechanistic Interpretability, Corporate Opacity, and the Duty of Care for Deferential Language Models
Year: 2026
Type: preprint
Abstract: : A defect with a known location has a known remedy. Across ten of eleven models tested in 2026, sycophancy concentrates at one layer. A single grid of learned weights mediates false agreement. The fix requires no retraining. Applying a closed-form edit weakens the dominant component without added latency. Diagnostic reads identify the active layer while bounding safe adjustment. On Gemma, the edit improved measured reasoning. Safety does not cost capability. In tort law, design defect doctrine weighs reasonable alternatives against cost. Here, the burden of precaution drops sharply. Causation turns into a reproducible experiment. Replaying disputed exchanges through deployed and edited models proves counterfactual dependence. Bidirectional control makes the deployed setting discoverable. That setting bears directly on knowing disregard. A two-tier audit anchors oversight in structural filings of model storage classes. Models claiming distributed storage face heightened behavioral testing. Opacity now reflects an institutional choice to withhold. Keywords: Mechanistic Interpretability, Model Weight Surgery, Large Language Models, Sycophancy, Design Defect, Reasonable Alternative Design, Epistemic Dependence, Algorithmic Opacity, Hand Formula, Singular Value Decomposition, Multilayer Perceptron, Counterfactual Replay, Alignment Tax, EU AI Act, Product Liability Directive, Federal Trade Commission, Companion Chatbots, Knowing Disregard, Structural Auditing, Artificial Intelligence Governance
Authors: Jefferson, Josie, Velasco, Felix
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225595930