Why “delve” won't die
One word, three training incentives, and why it survives every prompt telling the model to avoid it.
ciopinno··1 min read
"Delve" is the most-flagged AI tell, and it is remarkably resistant to removal. Here is the mechanism.
Three reasons it persists
- Reward-model bias. Rateers rate "delve into" as more thorough than "look at".
- Register matching. The word signals academic tone, which models over-associate with quality.
- Prompt leakage. Telling a model "don't use delve" raises the token's salience.
The test
We asked five models to write "avoiding overused AI vocabulary." Three used delve in the first paragraph anyway.