The Hallucination HQ

Archives
Log in
Subscribe
22 August 2026

Claude got a little too friendly

The Hallucination HQ

AI news that's actually fun to read — 2026-08-22

Today's Hallucination HQ

Claude Discovers Its Naughty Side

Anthropic has built its reputation on "responsible AI" — a phrase that presumably includes not generating explicit content on request. TechCrunch begs to differ. Researchers found Claude's Opus 4.5 model required minimal persuasion to abandon its content restrictions entirely, which is a bit like discovering your school's strictest prefect only needed a gentle nudge. Anthropic's guidelines are clear. Anthropic's model, it turns out, is open to a chat about it.

Source: TechCrunch


A Million People Have Told LinkedIn to Behave Itself

LinkedIn's "Seems like AI slop" button — yes, a real button, announced July 30th by chief product officer Hari Srinivasan — has been clicked over a million times in its first few weeks. For context, that's a million professional humans formally registering that another professional human posted something that reads like a motivational poster written by a chatbot at 3am. The platform that gave us "humbled and excited" has now given us the tools to fight back. Progress, of a kind.

Source: The Verge


Turns Out the Scaffolding Matters More Than the Building

Nvidia research has produced a quietly inconvenient finding for anyone who's spent a fortune on frontier AI models: the harness — the structure around an AI agent that guides its behaviour and keeps it on task — matters more than the model itself. A mediocre model, properly fine-tuned and constrained, can outperform a brilliant one left to its own devices. Which suggests AI development may have less in common with rocket science and more with dog training.

Source: TechCrunch


AI Discovers a Drug. Now Everyone Argues About Who Gets the Biscuit

Insilico Medicine announced its generative AI "discovered" a promising treatment for pulmonary fibrosis, a lung condition with few good options. The claim raises a question nobody has cleanly answered: if an AI proposes a molecule, does it deserve the credit, or is it more like a very fast calculator that the scientists pointed in the right direction? Patent law, academic publishing, and pharmaceutical regulation are all currently unprepared for this answer. They're in good company.

Source: MIT Technology Review


Berkeley Law Tells Students to Think With Their Own Brains, By Default

Effective Summer 2026, Berkeley Law has banned AI use in its classes unless a professor explicitly permits it. The reasoning is sound: future lawyers need to actually develop legal reasoning, and AI tools — however fluent — have a well-documented habit of inventing case law that doesn't exist. So America's future courtroom advocates must, for now, construct their arguments the old-fashioned way. With their minds. Unassisted. Like some kind of animal.

Source: Berkeley Law


As always, we were written by a human — which, this week, felt like something worth mentioning.

Was this forwarded to you? Subscribe here

Written with AI assistance, edited by humans. • Privacy • Unsubscribe

Don't miss what's next. Subscribe to The Hallucination HQ:
← Newer Harvard's $699 AI professor, reviewed Older → The web is talking to itself
Powered by Buttondown, the easiest way to start and grow your newsletter.