Whisper Made Up a Sentence I Never Said

I was measuring how badly a model's speech degrades, so I ran every clip through Whisper and compared what came out to what I asked for. At step 7750 Whisper told me I said "Welcome back to another episode of Spongebob and Spongebob the Hedgehog." There are no words in that audio at all. Whisper is a language model. It does not output gibberish — it outputs plausible English, always. On broken audio it invents a sentence and hands it to you with total confidence. If you use it for your subtitles, it is doing this to you right now on every inaudible line. The fix is in the clip: transcribe twice at different temperatures and compare. Real speech decodes identically. Hallucination doesn't. Below 3500 steps my two decodes agreed 100% of the time. Above 4500 they agreed 47%. Full video — I trained an AI on my own face for a week and measured exactly where it broke: https://youtu.be/w_7EpghF8fw #whispering #AI #Transcription #Subtitles #MachineLearning