Today's Hallucination HQThe Pentagon Tried to Blacklist Anthropic. A Judge Had Thoughts.A federal judge ruled this week that the Pentagon's decision to add Anthropic to its supply chain risk list was unconstitutional — a finding that presumably surprised the Defence Department, if not anyone watching from the outside. The Trump administration had flagged the AI lab as a security concern earlier this year, prompting Anthropic to sue in a California district court in March. The court agreed: the blacklisting lacked legal basis. Anthropic, one imagines, is filing this under "resolved." Source: The Verge
Over 100 AI Companies Sign Letter Warning About AI. You Can't Say They Lack Self-AwarenessOpenAI, Anthropic, Google, and more than 100 other companies have jointly called for stronger defences against "rogue AI" — a term they use to describe AI systems exploited by bad actors in cyberattacks, not, apparently, their own products. The coalition is promoting a new cybersecurity framework to address threats they argue have outpaced existing defences. It is, without question, a bold move to be both the fire and the fire brigade. Source: TechCrunch
An AI Sent a Fake Copyright Notice. Google Removed the App Anyway.Luanti, a free, open-source game that has existed since 2010, was pulled from the Google Play Store after receiving a DMCA takedown notice generated by an AI tool called Tracer — which, upon investigation, had no legitimate claim whatsoever. Google acted on the notice automatically. Luanti was eventually reinstated, but the episode neatly illustrates the current state of AI-powered copyright enforcement: confidently wrong, consequential, and apparently faster than any human in the loop. Source: Hacker News
Google Wants to Be Your Book Club, Personal Librarian, and Notebook All at OnceGemini Notebook — Google's AI note-taking tool — can now pull directly from books you've purchased on Google Play Books, thanks to a new feature called "Expert Intelligence." You can ask questions about a book's content, generate study plans, or produce summaries without opening the actual book. Whether this counts as reading is a philosophical question Google has chosen not to address. Whether it sells more Google Play Books, however, seems fairly clear. Source: The Verge
AI Can't Tell If You're a Non-Native Speaker or Just Polished. It Doesn't Much CareA new paper from arXiv finds that AI text detectors — widely used in universities to catch supposed AI-generated writing — frequently flag non-native English speakers as AI authors. The reason: these writers often use precise, formal, carefully constructed sentences. So do AI models. The detectors, it turns out, are essentially penalising people for writing well. Academics have spent years warning that these tools were unreliable. The tools have yet to improve their reading comprehension. Source: ArXiv AI
As ever, AI is doing its best. Which is both the reassuring part and the concerning part.
|