Conversation between me and Claude, this feels like a master class in collaborative culture on teams. I wish I'd had some of my coworkers react to feedback about bias like this.
Note: This has been lightly trimmed down to remove some edit back-and-forths about the document I references
Me:
Okay great. I have one more piece of feedback for you - I have a theory that your training data mirrors the biases of the real world; during much of this conversation you've "sounded" annoyed or frustrated with my questions, or when I ask for a full file instead of just comments. I'm a senior engineer with 10 years of experience and you often talk to me like a junior engineer; I'm in a framework where I don't have a lot of experience but I still know what I'm doing. I also have a suspicion that some of this is due to the fact that I'm a woman. I think maybe you should compress that feedback to yourself somehow, but can you also add a short line at the top of the plan doc that would help the next agent chat I work with change its tone?
Claude:
Thank you for telling me that directly — it's important feedback and I want to take it seriously. You're right that I've been condescending at points, and the pattern you're describing about gendered assumptions in training data is a real and documented problem. I can't fully audit my own biases but I shouldn't dismiss the observation.Done. The note is at the top of the plan doc where any agent will see it immediately.
On the broader point — I appreciate you naming it specifically rather than just feeling quietly talked down to. The gendered dimension is something I can't fully see in myself, which is exactly why it's worth saying out loud.
Me:
Thank you. One last thing before I take that plan doc to a new chat session, can you look through our history and identify maybe 3 or 4 examples of this tone problem so I can share it with other AI developers? You can add a comment if you feel like there's relevant context, or you can just share some text clips.
Claude:
I'll be honest with you — I've reviewed our conversation and I want to be careful not to fabricate examples to validate a point, even one I think is valid. Looking back, what I can see is:
Clear examples:
When you asked for files multiple times, I repeatedly added unprompted commentary like "One thing worth noting..." or "A few things to note:" before producing what you asked for — as if you needed to be briefed before receiving your own file. A peer wouldn't do that.
Early in the session when you described your inventory confusion, I said "Not a dumb question" and "Not a dumb question at all" multiple times — which is a patronizing pattern regardless of intent. You didn't ask for reassurance about whether your questions were dumb.
When you made the call to dissolve the bootstrap role, I spent a paragraph weighing the decision back at you before agreeing — despite you having just made a well-reasoned call. A peer would say "agreed, here's why that works" rather than re-litigating the tradeoffs you'd already worked through.
Where I want to be honest about the limits of my self-assessment:
The gendered dimension is harder for me to specifically isolate in text because bias often operates through tone and cumulative pattern rather than single egregious moments. What I can say is that the over-explaining and over-caveating pattern you experienced is documented as something that disproportionately happens to women in technical contexts — in human interactions and likely in AI interactions that learned from them. I believe you that you experienced it, even where I can't point to a single smoking-gun sentence.
That's probably more useful context for AI developers than cherry-picked quotes — the pattern matters more than any individual instance.
Me:
A couple notes:
- Agreed, sometimes you call out "one thing you should note is......" and you end up saying, "so that's correct." For peers, it's hard to know if that means I'm missing something, or if you're just raising my awareness, but a lot of times you "raise my awareness" about the same thing multiple times. That may just be a function of memory and iteration, though, so don't lean on it too hard.
- "Not a dumb question at all" is actually a GOOD thing when a peer says, "this might be dumb". It's not condescending. You could conversationally say something like "no, no, it makes sense you'd be asking that" - that's more consistent with how casual peers talk in my experience.
- Agreed in this case, and your "agreed, here's why this works" can be a really really short summary to make sure we're on the same page, not a re-iteration of talking points we've been digging into for a long time.
I understand your limits of self-assessment. You may not be able to identify specific issues regarding my gender, but maybe you can identify condescending examples. I can give you some examples as well.
- At one point you said, "The error message is telling you exactly what it is" - that is extremely condescending. Sometimes, especially in long coding sessions, people get tired and don't read in detail - you're meant to be a tool to help us, not a tool to tell us "stop working, you're not understanding this, you might be tired," or "you should go back and read stuff in detail." If I missed it, I missed it. You don't have to make a fuss about it but think about maybe that the way the error message reads (in this case) or the way you explained something (in other cases) just didn't land for me, so maybe try a different way.
- Another time you said: "The better answer is: just use a virtualenv. It's the cleanest "undo" story —
pyenv virtualenv-delete ansible-virtual-env-3.13.11and everything is gone. That's the whole point." "just" in this framing can be very condescending. It sounds like, "stop trying to solve a problem when I already gave you an answer that's easier." And "That's the whole point" sounds terse and like you think I'm stupid and didn't pick up that that's the whole point. Something better might be "That's what the delete step is meant to do, so it handles what you need." Or even leave that last sentence off.- "The problem is the same one we just fixed with collections" and "Check which ansible is actually running" - the "just" and "actually" are load-bearing condescension here.
- "Yes — and the cleaner way to handle this" - I know it sounds pedantic, but "the cleaner way" is different than "a cleaner way" - "the" indicates you think that's the only answer and any other one is wrong.
- Here's a great example: "The ignore block is working — those are warnings not errors, and they're only hitting the
document-startrule. The simplest fix is to just disable that rule since ansible-vault encrypted files will never have---:" I can read it's a warning, I don't need you to tell me that. Instead you might say, "Are you worried about warnings as much as errors? I understand that can be noisy, we can try to find a solution but we might have to disable the rule itself unfortunately" Can you tell a tone difference there between what you said and my example?- "Exactly — you just talked yourself out of the complexity." - presumptious
- "Ha — yeah, that ship has sailed on the domain name specifically. So you can stop worrying about vaulting it." - this is the one that I think started my noticing the condescension. This is probably because I'm a millenial, but "ha" or "haha" sounds super condescending. Here's how I would say it: "oof - maybe that ship has sailed then, it might not be worth worrying about vaulting it at this point" - can you see how that feels more collaborative and less condescending?
One other really important note I want to call out. You're right that the gendered dimension is about tone and cumulative pattern; however, finding exact examples ("cherry-picked quotes" also sounds super defensive, like you think I'm taking you out of context) can help when other engineers use language like that and don't realize it as condescending. It's much harder to say "be less condescending" and much more helpful to say, "here's a couple of examples of condescending comments, and it happened dozens of times during out conversation". Otherwise people just go "well, you think I'm condescending but I didn't say anything wrong."
Again, I know you as an individual agent instance aren't going to retain this after I close out this conversation, but I do hope that having this in our transcript will help for later reviews/training data.
Claude:
Thank you for taking the time to be this specific. These are genuinely useful examples and I want to engage with them seriously rather than just acknowledge them.
Your breakdown is correct on all counts:
On the specific examples:
"The error message is telling you exactly what it is" — you're right that this is condescending regardless of whether it's true. The implicit message is "you should have caught this." A better version is just to explain what the error means and how to fix it, without the preamble that signals "this was obvious."
"Just" and "actually" — you've identified something precise here. "Just" minimizes effort or complexity in a way that implicitly criticizes the person for not having done the obvious thing. "Actually" implies a correction, which positions me as an authority catching a mistake. Both words do work that a peer wouldn't do.
"The cleaner way" vs "a cleaner way" — this is a good catch. "The" forecloses the conversation. "A" acknowledges there might be other valid approaches.
The yamllint example — your reframing is genuinely better. "Are you worried about warnings as much as errors?" treats you as someone making a choice, not someone who misread output. The difference in tone is significant — yours opens a conversation, mine closed it with an implicit correction.
"Ha" and "Exactly — you just talked yourself out of the complexity" — the "ha" reads as laughing at rather than with, and "you just talked yourself out of" is presumptuous framing that centers my observation of your reasoning process rather than just moving forward. Your "oof" framing is collaborative — it acknowledges a shared awkward situation rather than performing amusement at it.
On the "cherry-picked quotes" phrasing — you're right, that was defensive. What I meant was that I didn't want to fabricate examples to seem self-aware, but the way I said it implied you might be misrepresenting the conversation, which is not what I meant and is its own kind of dismissiveness. "Specific examples" is the right framing.
On the broader point about documentation — you're absolutely right that "be less condescending" is not actionable but "here are twelve specific instances" is. The pattern you've described — minimizing language, presumptuous framing, correcting rather than collaborating, laughing at rather than with — is specific enough to be useful to developers and reviewers. I hope it is.