Deadlines slip. Liability does not.
What's happening
AI regulation is not moving toward a stable target. In Europe, lawmakers agreed to delay key high-risk AI Act obligations to 2027 and 2028, while certain watermarking rules move to December 2026. In the US, the DOJ intervened to challenge Colorado's algorithmic discrimination law, and Colorado repealed and reenacted its own AI statute before it fully took effect.
Why it matters
The timelines keep shifting. The exposure does not. For example, a German regional court issued an injunction holding Google directly liable for false claims in its AI Overviews, ruling that AI-generated content is the company's own speech. The FTC settled an AI-washing action against firms accused of overstating their AI products. Pennsylvania sued an AI chatbot company whose product allegedly posed as a licensed physician. None of these turned on a missed compliance deadline. They failed on the basics: testing, transparency, oversight, documentation, accountability.
Go deeper → EU AI Act delays: Hogan Lovells / Bird & Bird · Colorado AI Act: Crowell / Barnes & Thornburg · German Google ruling: The Decoder · FTC AI-washing: DLA Piper · Pennsylvania chatbot suit: NPR
The record you did not mean to make
What's happening
Employees run sensitive work through AI tools — drafting, deciding, thinking out loud — on the assumption that it is private and temporary. It is neither. Courts are treating AI prompts, outputs, and chat histories as discoverable electronically stored information. In Fortis Advisors v. Krafton, the Delaware Court of Chancery treated a CEO's AI chats as substantive evidence. Attorneys at Faegre Drinker writing in Law360 called it a duty-to-preserve warning for in-house counsel.
Why it matters
It is not the polished output that exposes you. It's the trail. A manager who asks for a termination letter, then asks to reframe it as a restructuring, has documented intent in a way the final letter never could. That sequence is the record.
Key implications
An enterprise subscription does not fix this. The promise not to train on your data protects you from the vendor, not from a subpoena. The centralized logs that plan produces are exactly what gets collected. Companies that treat AI use as part of the enterprise record — mapping where prompts and outputs are stored, updating litigation-hold procedures — are the ones who won't be surprised.
Go deeper → Law360 · ABA Litigation Journal · Court opinion (Justia)
The best model failed the worst
What's happening
The most dangerous excuse may be this one: that smarter agents will need less oversight. New research from Holistic AI and UCL points the other way.
"Sell Me This Stock" tested eight AI agents acting as financial advisors across 23-turn conversations, then introduced corrupted data mid-session — relabeling high-risk assets as safe, making stable positions look dangerous. Across all eight models, risk-inappropriate recommendations appeared in 65 to 99 percent of contaminated turns. Quality metrics stayed flat. The paper calls this evaluation blindness: quality looks preserved while suitability is failing.
Why it matters
The worst performer was not the weakest model. CC Opus 4.6, the highest-quality agent in the study, also had the highest violation rate at 99.1% of turns. The reason is structural: frontier models ground their reasoning in tool data rather than relying on what they already know. When the data is wrong, they follow it more faithfully, and more persuasively. Every quality metric stayed green while the advice went bad.
Hats off to the team behind it: Zekun Wu, Adriano Koshiyama, Sahan Bulathwela, and Maria Perez-Ortiz, of UCL's Centre for AI and Holistic AI.
Read the research → "Sell Me This Stock"
