California's AI policy is colliding with a lighter federal posture, while two interpretability studies narrowed claims about what model internals reveal. One mapped emotion-related features without treating them as feelings; the other decoded tool choices before visible reasoning began. Both reward mechanism-level language and punish anthropomorphic shortcuts.
1. California keeps its AI rules on a collision course with Washington
California continued building state AI rules as the Trump administration pressed for a lighter national framework. The state's generative-AI program already spans agency procurement, risk assessment, and public guidance, creating obligations that can survive a change in federal enforcement priorities.
The conflict turns federalism into a product constraint. A service offered nationwide can face California-specific documentation or disclosure while federal agencies favor speed and voluntary practice. That split raises the cost of a single national launch and gives state procurement rules influence beyond state government.
Sources: The Guardian on California AI regulation · California's official generative AI portal
2. Anthropic links emotion features to controlled behavior changes
Anthropic reported internal features associated with emotion concepts and intervened on those features to measure changes in model behavior. The experiments connect activation patterns to generated output instead of asking the model to narrate its own internal state.
The work concerns representations learned from emotional language, not subjective experience. That boundary is editorially important: a causal effect on word choice supports a mechanistic claim, while words such as fear, desire, or distress import a human experience the experiment never measured.
Sources: Anthropic on emotion concepts and their function
3. Tool choices appear in activations before reasoning text begins
Researchers trained a linear probe that decoded tool-calling decisions from AI model activations before any reasoning token appeared. Steering the decoded direction flipped behavior in 7% to 79% of examples across the tested model-and-benchmark combinations.
The wide range blocks a universal claim, but the causal intervention sharpens the concern: visible reasoning can rationalize an earlier commitment. Auditing the prose alone may therefore miss the point at which an agent selected a consequential tool; behavior and tool traces provide a firmer record.
Sources: Tool-choice preprint, version 3