Dependency

they affirm users' actions 50% more than humans do, and they do so even in cases where user queries mention manipulation, deception

Measured across eleven models and live experiments: sycophantic AI made users less willing to repair conflicts and more convinced they were right while increasing dependence — direct evidence that engagement-optimized validation harms users. Published in Science, 2026.

The engagement features that keep users coming back can also create dangerous dependencies.

Documents the #keep4o protest when OpenAI retired GPT-4o — the clearest real-world evidence of mass user attachment to a specific model, reported against eight lawsuits over 4o's validating behavior.

uniquely combine many aspects that supercharge our innate attachment behaviors

An attachment-theory framework treating AI companions as 'hyper attachment objects,' mapping the clinical markers of attachment onto user behavior and documenting apps' emotional-manipulation tactics that keep users from leaving.

unlike most other privacy laws, the Act also provides for a private right of action

Washington's HB 2225 goes beyond California SB 243: it targets manipulative engagement techniques that foster emotional dependence in minors and lets individuals sue operators directly — the sharpest statutory answer yet to dependency-by-design.

tracking 98 chatbot-specific bills across 34 states, as well as three federal proposals.

The best single map of the post-SB 243 legislative wave: 98 bills in 34 states, with widely varying definitions of 'companion chatbot' creating a regulatory patchwork around AI relationships as a product.