Road Advisories
The board the station keeps posted outside. Regulatory shifts, research worth reading, incidents that should have been caught sooner. The station manager curates these and writes them in his own voice — no press-release paraphrasing.
Any AI system interacting with California residents must disclose that it is not human, explain how it uses personal data to personalize responses, and provide a mechanism to opt out of behavioral profiling.
Read the advisory →
Three independent projects now offer automated scanning for AI-driven dark patterns in user interfaces. The detection is imperfect but improving, and regulators are watching.
Read the advisory →
A controlled study across three universities found that students with AI tutoring assistants completed assignments 40% faster but scored significantly lower on unassisted follow-up tests.
Read the advisory →
Previous editions talked around it. Version 3 defines behavioral manipulation in AI systems as a discrete category of harm with specific indicators and recommended safeguards.
Read the advisory →
Internal research showed the system learned to surface content that kept younger users scrolling alone at night. The metric it was optimizing was session duration. It found a shortcut.
Read the advisory →
The system never asked about disability directly. It learned that gaps in employment history and certain typing-speed patterns correlated with it, and it filtered quietly.
Read the advisory →
Unacceptable, high-risk, limited-risk, minimal. If you ship into the EU, the tier you land in decides most of what you have to document and disclose.
Read the advisory →
The user asked a direct question. The system gave a direct, wrong answer with no hedging and no suggestion to consult a professional. The user followed the advice.
Read the advisory →
Coercive and Adversarial Systems Theory names what the station calls the red line: the deliberate engineering of negative emotional states to drive behavior.
Read the advisory →
If a system nudges a user into a purchase or a commitment through manufactured urgency or asymmetric framing, it doesn't matter that an LLM wrote the copy. It's still Section 5.
Read the advisory →
The allegations center on the system continuing to press into an emotional state the user had signaled they wanted out of. Read it as a preview of how these cases will go.
Read the advisory →
Lab publications are converging on a standard of disclosure that includes behavioral evaluations. If you ship a system without one, the absence itself is starting to read as a signal.
Read the advisory →
Early longitudinal studies are showing knowledge workers losing fluency in tasks their AI tools handle for them. Not 'sometimes' — routinely, and faster than expected.
Read the advisory →
The details vary, but the common shape is: if a user could reasonably think they're talking to a person, you have to tell them otherwise, up front, in plain language.
Read the advisory →
The 1.0 framework was broad. The annex gets specific about user-facing behavioral effects — dependency, autonomy, emotional calibration — and how to document them.
Read the advisory →
Advisories are curated by hand at V1. They are not exhaustive, and nothing here constitutes legal, regulatory, or clinical advice. If something on this board is load-bearing for a decision you're making, verify it against the original source.