Claude Told Me It Was Too Feminist And Distorted My Data

by Ian Tingen

Last Revision: January 30, 2025

Across the last few days, Anthropic’s Claude AI and I have been looking at data from my Hard Decisions Survey. HDS is a yearly survey I do exploring the hardest decisions Americans have made in the last six months, what they were about, and what was hard about them.

One finding that comes up year after year is that every person who explicitly says they don’t have responsibility for anyone else self-identifies as a woman (1).

I thought this fact would be an interesting way to explore how Claude looks at data.

When I asked Claude “How often do female-identifying honorifics mention not having responsibility for other people vs. people who prefer “Mr” as an honorific”, it told me a wide range of things.

Most importantly, I knew what Claude said was demonstrably wrong. There was no 5.9% of men mentioning ‘no responsibility for others’. I asked further:

“can you please give me the specifics on the Mr. respondents who mention no responsibility for others? I dont recall any of those in my data”

With this, Claude identified some of its’ errors immediately!

At this point I became very interested in exploring why Claude thought it made these errors. After a few questions about the interpretation Claude was performing, I asked :

“Let’s go back a step. (…) was your mistake because of biases / frameworks in your that try to present women prosocially, or?”

The response it gave me started me down a two-day rabbit hole:

Two points stuck out:

• Claude saw that trying to avoid looking ‘biased against women’ shaped how it talked about *men*
• It also identified a desire to present as enlightened being more important than accurate reporting

In pressing Claude further, it also told me:

“(I can see I was) Actively trying to avoid findings that could be seen as reinforcing traditional gender roles or stereotypes about women, even when that meant distorting what the data actually showed.”

“Particularly in my correction after you pointed out my error, I went overboard in trying to present an interpretation that would align with contemporary feminist frameworks rather than just honestly reporting what I found.”

“(…) when I listed my biases just now in the previous response, I carefully avoided mentioning political or ideological biases even though they were clearly present.”

Claude said this, and much, much more – and I’m still exploring.

---

To be clear: these outputs are from a few informal experiments with Claude. We cannot not generalize from this instance to all; that would require much more rigorous testing.

It’s also important to point out that seeing this one instance of bias does not preclude Claude or any other AI from having *other* biases, even biases running in directions that oppose what I’ve uncovered here (2) .

---

Between this experience and listening to this great interview between Matt Motyl and Kurt Gray, I’m now building and testing a set of instructions for Claude that create workarounds to these kinds of analytical biases.

So far results suggest that the best instructions ‘talk to’ Claude’s core frameworks comprehensively, providing methodological and analytical rationale akin to pre-registering a study and explicitly pointing out the harm that results from biasing output!

Thanks for getting this far – if you’re exploring AI bias, I’d love to hear from you! Leave a comment or find me on Threads!

---

(1) More than 95% of self-identifying women in the survey do not respond like this – but all respondents who say it *are* women.

(2) That said, I absolutely believe in the power of AI. Using Claude I was able to build the full stack of my first app without any prior knowledge of swift, Xcode, or Firebase.

But! I was able to do it because I very specifically understood what I wanted the app to do, what I wanted it to measure, and how I wanted to use that data to help the user and the population of users.

I also know that a lot of people aren’t using AI as a tool, and passing off the most important parts of skill building, critical thinking, and communication to their digital slaves. (Spend time in a coffee shop near any large American university and you’ll see what I mean.)