I Tried to Make AI Writing Sound Human by Banning AI Words Through Logit_bias

Using `logit_bias`, an API setting that directly adjusts token selection scores, one developer tried to make AI-generated writing less detectable by blocking commonly overused AI words at the model level. Tested across eight articles using DeepSeek V4 Flash, the method removed roughly a third of flagged words but caused one article to change meaning and introduced awkward phrasing in another, highlighting a key tradeoff between word suppression and writing quality.