The Hallucination HQ

Archives
Log in
Subscribe
16 October 2026

AI passed the test. The test failed.

The Hallucination HQ

AI news that's actually fun to read — 2026-10-16

Today's Hallucination HQ

Anthropic Invites Researchers To Rifle Through Its Drawers

Safety research outfit METR now has employee-level access to Anthropic's models, data, and internal practices — and crucially, Anthropic has waived the right to block publication of whatever METR finds. That last bit matters enormously. An independent audit where the auditee controls the final report is just expensive PR. This arrangement has actual teeth, making it one of the more genuinely meaningful AI transparency moves in recent memory, which admittedly is a low bar, but still. Source: TechCrunch


ChatGPT Would Like To Know What You're Reading, Thanks

OpenAI's Chrome extension now lets ChatGPT peek at your open browser tabs when you @ mention it — handy for summarising pages, cross-referencing research, or having an AI quietly observe that you have seventeen tabs open and zero of them are the thing you're supposed to be doing. It's opt-in, and OpenAI says tab content isn't stored. Whether you find this useful or mildly unsettling probably depends on what you currently have open. Source: The Verge


AI Exams Are Getting Full Marks. The Exams Are The Problem

Researchers have a thorny issue: AI benchmarks — the standardised tests used to measure model capability and inform safety decisions — keep getting aced, at which point they stop being useful. A new paper argues that once several models cluster near perfect scores, a benchmark can no longer tell us who's better or safer. It's less "AI is too clever for school" and more "we keep setting the same exam and acting surprised when everyone memorises the answers." Source: ArXiv AI


Meta And Apple Walk Into A Bar, Leave With A Faster AI Model

Meta's new Muse Spark is being billed as the most capable open-weight AI model from a US company — open-weight meaning the underlying numbers are publicly available, unlike the locked-down models from OpenAI and Anthropic. The notable twist: Meta worked with Apple to optimise it specifically for M5 Ultra Mac Studios, a collaboration between two companies not historically known for holding hands. Performance claims are Meta's own, so treat them with the usual seasoning. Source: The Verge


Google Wants To Be The Only Shopping Cart You Ever Need

Google has launched a Universal Cart in the US, embedded across Search and the Gemini app, letting users shop from multiple retailers in a single checkout flow. The ambition is clear: become the layer between consumers and every online shop on earth, which is either wonderfully convenient or the logical conclusion of Google wanting to be involved in absolutely everything you do. Retailers joining voluntarily may want to re-read that sentence slowly. Source: TechCrunch


As always, the AI is doing fine. It's the humans writing the safety benchmarks who need a sit-down.

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 Claude went feral. Disney hired everyone. Older → Character.AI meets its final boss
Powered by Buttondown, the easiest way to start and grow your newsletter.