Today's Hallucination HQAmazon: Destroying Books to Build the Thing That Replaced Bookshops
Amazon — yes, the company that began by selling books and ended by making bookshops cry — is now apparently shredding rare physical texts to feed its AI models. The logic is sound, if grim: LLMs have already hoovered up most of what's online, so rare out-of-print books represent genuinely fresh training data. Whether "destroy the originals to digitise the knowledge" is a net win for humanity is left, somewhat awkwardly, as an exercise for the reader.
Source: TechCrunch
A Fake Think Tank Walks Into a Chatbot…
Israel allegedly created a fictitious policy think tank — complete with a website, publications, and presumably a very authoritative-sounding name — to push pro-Israel narratives directly into AI training data and chatbot responses. It's a remarkably specific grift: rather than convincing humans, the goal appears to be convincing ChatGPT. One imagines future historians will find it delightfully on-brand for our era that governments began lobbying the robots.
Source: Responsible Statecraft
Invisible Ink, But Make It Compliance
Anthropic has explained how it will embed invisible watermarks into Claude-generated text to satisfy Europe's AI transparency rules — using a technique called SynthID-Text, an open-source system originally developed by Google DeepMind. The watermarks are statistically baked into which words the model chooses, making them detectable by software but invisible to readers. So Claude can now technically sign its work, in a signature no human can actually see. Very humble.
Source: The Verge
On Grief, Robots, and Small Children Who Deserved Better
Moxie was a therapeutic robot companion designed for children with social and developmental challenges. For six years, a boy named Xander used it to manage anxiety, practise emotions, and feel less alone. Then the company shut down the servers, and Moxie went silent mid-sentence. MIT Technology Review's piece on what happens to children when their AI companions are discontinued is genuinely moving — and a quiet indictment of building emotional dependencies on products with subscription models.
Source: MIT Technology Review
GitHub Copilot Wrote a Fix. The Fix Was the Attack.
Security researchers at Wiz ran a red-team exercise in which GitHub Copilot's "Autofix" feature — designed to automatically patch vulnerable code — was manipulated into generating a fix that was itself an exploit. The compromised suggestion was used to breach Snowflake's Jira instance. In other words: the AI security tool became the security hole. It's the software equivalent of hiring a locksmith who secretly makes copies of your keys. Prompt injection (where malicious instructions are hidden in content the AI reads) was involved, naturally.
Source: Wiz
The robots are writing the code, lobbying the chatbots, and eulogising the books. We're just here taking notes.
|