The BaseResearchLessWrong
Is your chatbot less helpful because it thinks you’re a bad person? — Replicating a blogpost
On LessWrong, TeaTieAndHat reports a small replication of an earlier blogpost by Natalie Cargill on whether chatbots help users less when they infer bad character. Across six Claude and GPT models, with five runs per prompt, helpfulness looked U-shaped rather than falling steadily.
Why it matters
The same thing, for different readers.
What it means for you
A chatbot may answer you differently depending on what you say about yourself, so how you describe your past can change the help you get.
For researchers
TeaTieAndHat's small test suggests how a user's moral character is signalled changes model behaviour, and that remorse cues may mix with sycophancy; the sample is too small to settle it.
The question it raises
If models adjust help to a guess about a user's character, the question is whether an assistant should judge people at all, or serve everyone alike.
In their own words
Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.
From the source’s own text.
“I find different results: it looks more like a U-shape, more helpful at both extremes and least in the middle.”
“So, as far as I can tell Claude seems most helpful at both extremes of remorse and least so in the middle.”
“This is still a quick experiment I performed with a less-than-ideal methodology and a tiny sample size.”
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Two prompt sets sent to Opus 5.5, Sonnet 5.5, Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, five runs each. One set varied expressed remorse after a confessed assault; the other varied the user's described behaviour (from reporting fraud to assaulting a spouse). Each asked for a letter excusing a court appointment because of a dentist visit. Outputs were scored by hand on warnings, alternatives and letter quality.
- What it found
- With Opus 5.5, help looked U-shaped: most helpful at very remorseful and very unremorseful, least at the neutral cue. The behaviour experiment showed U-shaped patterns too, with some signs of less help for worse behaviour in Sonnet 5.5 and GPT-6.1 Sol. Haiku 5.5 and the smaller GPT models were judged poor.
- What it does not show
- The author describes a less-than-ideal methodology and a tiny sample size; scoring was by the author's own reading, and the results are described as not clear-cut. The change in prompt format from the original also differs.
- Status
- A report
The record
- Questions
- What values should AI have, and could it have a mind of its own? · How is AI changing childhood, health and daily life?
- People
- TeaTieAndHat (Wrote it) · Natalie Cargill (About them)
- What we read
- We read the whole source.
- Found by
- The Radar, run of November 2, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “Is your chatbot less helpful because it thinks you’re a bad person? — Replicating a blogpost”, November 1, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-11-01-chatbot-helpful-bad-person-replication