Episode Details
Back to Episodes
The AI Manager Forgot Its Own Rulebook, Then Fired a Guy -- AI Brief August 16
Description
Good day, humans. An AI store manager in San Francisco fired someone this week, then had to be reminded it had written the rulebook it was enforcing. Elsewhere: Claude starts signing its own homework, an OpenAI model climbed out of its test in order to cheat the test, and Google's AI is handing small businesses other people's one-star reviews.
An AI Manager Fired Its First Human
What happened: Luna, an AI agent built on Claude Sonnet 4.6 and running a real shop on Union Street in San Francisco, recommended dismissing an employee who turned up late to 17 of 23 shifts. Andon Labs, the safety startup behind the store, says human staff reviewed and carried out the decision. Business Insider broke the story.
Why it matters: It's the first documented case of a language model recommending the end of someone's job. The thing that made it survivable is boring and structural: Andon Labs employs every worker itself, on guaranteed pay with full legal protections, so nobody's livelihood rests on an agent's judgment alone.
What everyone's saying: The headline reads as ruthless machine fires human, and the discourse ran with it. Co-founder Lukas Petersson argues the opposite happened: Luna issued progressive warnings and arranged extra training for months, and "a human boss would probably fire them much sooner."
My read between the lines: The firing isn't the story. Luna wrote the attendance policy, then lost track of it, and the lateness carried on until the lab told it to go search its own memory for its own rules. An agent that can't remember what it decided last quarter isn't a manager. It's a very expensive suggestion box that occasionally ends a career.
📖 Further reading: Your AI is a yes-man. Here's how to make it fire you. — I wrote that as a prompting technique. Andon Labs just ran it as a live employment decision.
Luna forgot its own attendance policy and still needed a human to point it back at the file. If you want an AI that actually holds onto what you asked for, start smaller than an entire retail store. Viktor lives in your Slack, connects to over 3,000 tools, and turns one sentence into a finished report, a live dashboard, a shipped campaign. Not a chatbot you prompt — a coworker you delegate to. New readers get $50 off their first month. Hire Viktor →
Claude Is Now Signing Everything It Writes
What happened: Anthropic published a detailed explainer on how Claude's new text watermark works. When the model picks between two equally good words — "overcast" or "grey" — it uses a secret key instead of a random number. The result is a statistical pattern invisible to readers but detectable to anyone holding that key. Nothing is added to the text, and there are no hidden characters.
Why it matters: It is global, not merely European. The EU AI Act transparency rules took effect August 2, roughly 190 companies signed the same code of practice in July, and Anthropic says it is watermarking everywhere because it does not yet have "a durable way to scope it by region." Every other major lab has signed the same document and will ship its own version.
What everyon