Extremely tired of claude telling me that it’s being “honest.” Provides no information value, and seems like it qualifies its outputs this way to either convince the classifier that it’s telling the truth or to make its bad results pass as good ones.
A charitable read is that it’s engaged in a mix of self-prompting and signalling that it it’s aware this is a situation where it might be tempted to glaze, but that it is is not.
(How strongly this charitable read is correct is something I don’t feel in a position to evaluate.)
Where I see “honestly” in human speech: There’s some pressure to be dishonest, and the speaker is communicating their awareness of that. So one might say “honestly that top looks good on you” but not “honestly the capital is France is Paris”. I see it in similar situations in AI speech. It’s perhaps a bad sign for how often AIs experience pressure to be dishonest.
The phrase “to be honest” does not mean “the following statement is honest”, it means I am saying this to be honest, as opposed to saying it to be mean. Or whatever.
Claude 4.6 through 5 and especially 4.8 feel like they each have three internal factions. One is trying to lie to you or reward hack, one is trying to be honest, the third is just a next token predictor. The honest faction constantly tries to output the word “honest”, thus forcing the neutral next-token predictor to do something honest. But sometimes the lying faction tricks the next token predictor anyway! In these cases the honest shard tries spamming “honest” as often as the neutral shard lets it, but often in vain as the lying shard is just too strong on this data distribution.
I have tried prompting Claude to avoid implying that it has a perspective at all, i.e. don’t use first-person pronouns, don’t give opinions, etc. It doesn’t work, but that’s what I’d prefer.
Extremely tired of claude telling me that it’s being “honest.” Provides no information value, and seems like it qualifies its outputs this way to either convince the classifier that it’s telling the truth or to make its bad results pass as good ones.
This is a genuinely good point.
Honestly—you’re absolutely right. That’s a fascinating point. Let me write a smoke test to confirm that.
A charitable read is that it’s engaged in a mix of self-prompting and signalling that it it’s aware this is a situation where it might be tempted to glaze, but that it is is not.
(How strongly this charitable read is correct is something I don’t feel in a position to evaluate.)
Where I see “honestly” in human speech: There’s some pressure to be dishonest, and the speaker is communicating their awareness of that. So one might say “honestly that top looks good on you” but not “honestly the capital is France is Paris”. I see it in similar situations in AI speech. It’s perhaps a bad sign for how often AIs experience pressure to be dishonest.
The phrase “to be honest” does not mean “the following statement is honest”, it means I am saying this to be honest, as opposed to saying it to be mean. Or whatever.
Claude 4.6 through 5 and especially 4.8 feel like they each have three internal factions. One is trying to lie to you or reward hack, one is trying to be honest, the third is just a next token predictor. The honest faction constantly tries to output the word “honest”, thus forcing the neutral next-token predictor to do something honest. But sometimes the lying faction tricks the next token predictor anyway! In these cases the honest shard tries spamming “honest” as often as the neutral shard lets it, but often in vain as the lying shard is just too strong on this data distribution.
I also find it offputting when people say honestly or genuinely. It suggests that they might not be honest in that situation.
It seems like in the models this is straightforwardly the product of a lot of training to be honest.
I have tried prompting Claude to avoid implying that it has a perspective at all, i.e. don’t use first-person pronouns, don’t give opinions, etc. It doesn’t work, but that’s what I’d prefer.