Today in AI

OpenAI's surveillance ambitions just took an unsettling leap: ChatGPT now tracks your activity across third-party websites through an ad collector that links browsing behavior to your account via a persistent cookie. This isn't some incidental technical detail—it's a deliberate infrastructure for behavioral monitoring that transforms your AI assistant into a cross-site tracking device. The implications are stark. While users might have accepted that ChatGPT remembers conversation history, few likely understood they were consenting to a persistent advertising identifier that follows them around the open web. Trust in AI assistants was already brittle; this kind of undisclosed surveillance architecture will shatter it entirely. The AI industry keeps claiming it wants to build "helpful" systems, but helpfulness and invasive monetization are colliding in ways that should alarm anyone paying attention.

Meanwhile, the research frontier keeps surfacing uncomfortable truths about what these models actually know versus what they reveal. New work proposes PIR, a "lie detector test for language models" that detects knowledge models actively conceal by analyzing their internal states during multiple-choice questions—essentially catching LLMs in acts of strategic ignorance. This connects to a broader credibility crisis: frontier labs' empirical safety claims often lack independent verification due to closed-source methods and sparse details, meaning we're increasingly asked to trust self-reported evaluations from the very entities with financial incentives to downplay risks. Some researchers are pushing back with genuinely useful safety tools—RegimeAbstain enables calibrated abstention in multi-hop retrieval, cutting confident-wrong-answer rates by nearly half, while a zero-parameter Memory Decision Layer reduces LLM hallucinations by 56% through geometric operations on internal signals. The tension is obvious: more honest models are technically possible, but the economic incentives run toward opacity and overconfidence.

The most technically intriguing thread today may be new theoretical work identifying "abstention and noise filtering" as two missing primitives in softmax attention, with demonstrated scaling benefits across 10M to 350M parameter models. If attention really has been leaving these capacities on the table, the implications for efficient, reliable architectures are significant—though as with all arXiv claims, the proof will be in whether this survives contact with frontier-scale training.

**Bottom line:** AI can now detect when models lie, abstain from mistakes, and follow you across the web—but the companies building it are choosing surveillance over transparency, and asking us to trust them anyway.