Today's Hallucination HQOpenAI's New Trick Makes Safety Researchers Spill Their TeaOpenAI's upcoming Astra model introduces something called "recurrent depth" — a technique that lets the model loop back through its own reasoning rather than thinking in a straight line like every other AI. This sounds elegant until you realise that AI safety researchers, whose entire job is worrying about what AI might do next, are actively alarmed. When the people paid to stress-test worst-case scenarios say "hang on a minute," that's worth at least a raised eyebrow. Source: TechCrunch
215,000 Fake Reviews Walk Into a Search Engine. Perplexity Buys Them All a DrinkThree websites quietly generated over 215,000 "best software" pages — each one manufactured, none of them trustworthy — and Perplexity, the AI search engine positioning itself as Google's more honest cousin, has been citing them as legitimate sources. The scheme is impressively brazen: flood the web with confident-sounding nonsense, wait for AI to vacuum it up as fact, and watch your rankings soar. It's spam with a business plan, and the AI confidently doing citations is the last to know. Source: Hacker News / trellner.com
The White House Picks a Side in the NYT vs. OpenAI Copyright BoutThe Trump administration has filed in support of OpenAI in The New York Times' landmark copyright lawsuit — the one where the Times argues OpenAI hoovered up billions of dollars' worth of journalism to train its models without asking, paying, or leaving a note. The administration's position is essentially that this constitutes fair use, a legal doctrine not traditionally designed to cover "we trained a trillion-parameter model on your archive." The Times, presumably, was not delighted. Source: The Verge
The Internet May Already Be Mostly Bots Talking to Bots. Morning!Pangram's CEO is warning that we are "dangerously close" to Dead Internet Theory becoming reality — the idea that most online content is now AI-generated slop being consumed and cited by other AI, in a sort of automated ouroboros. Job applications, product reviews, insurance claims: all filling up with generated text that platforms can't reliably detect. The unsettling part isn't that robots are writing things. It's that we've built entire industries around trusting what's written online. Source: TechCrunch
New York City Tells Its Youngest Students That ChatGPT Can WaitNew York City has banned AI tools in classrooms from kindergarten through eighth grade for the 2026–2027 school year, affecting roughly 600,000 students. Mayor Zohran Mamdani's one-year moratorium gives schools time to figure out what responsible AI use actually looks like before handing it to ten-year-olds — a refreshingly cautious approach in an era when most institutions discover the downsides after the rollout. High schoolers remain ungoverned in this respect, which tracks. Source: The Verge
The AI cited four fake sources while writing this newsletter. We removed them. Mostly.
|