AI Fails

A running log of AI failures worth more than an eye-roll — the ones that reveal something about how these systems actually work, or where their limits are. My focus is on fails that show "something interesting about LLMs," not "haha, dumb LLMs or dumb LLM prompt writers." Each entry shows the actual exchange: verbatim where I have it, clearly marked as reconstructed where I'm working from memory. If you've got a good one I'll happily post it. Imagine the engagement!

The log

• 2026-05-01 — ChatGPT counted a four-day week as five
  I needed to submit my hours for the month, so I asked ChatGPT how many work hours April 2026 had. It answered 184 — built on a calendar it printed whose final week plainly showed four days, which it then counted as five. It corrected to 176 only under pushback, and only after I pushed past its "simple counting error" shrug did it name the real failure: it ignored the evidence it had just generated.
• 2026-05-12 — Twice wrong about Saturday Night Live's backup dancers
  Asked whether the dancers at the end of Jane Wickline's "chronically late" Weekend Update song were SNL cast or hired in, Gemini first insisted there were no dancers and suggested I'd misremembered. Asked again two weeks later it acknowledged them — then called the real dancers "fictional" until I pushed. Confident, wrong, corrected only under pressure, both times.
• 2026-05-27 — Sunrun, the buyout, and the inverted incentive — an assistant flipped a counterparty-incentive analysis: it told me Sunrun would replace my leased solar batteries before a buyout, when the same fact it cited (“they’re off the hook after sale”) is precisely why they wouldn’t. A new failure pattern worth its own tag: inverted incentive logic.