[MODEL] Model confidently asserts gender based on name pattern-matching, then fabricates justification

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude made incorrect assumptions about my project

What You Asked Claude to Do

We were having a conversation about prepping for my next 1:1 meeting. I asked Claude

" ❯ can you add text to existing documents? also is Bei's 1:1 document included as context for him?"

What Claude Actually Did

Claude answered my question and then tried to correct my pronoun usage.

1. Referred to Bei as "she/her" unprompted
2. When asked why, Claude cited evidence ("based on the 1:1 doc references like 'Bei's thoughts', 'Bei's topics'") — none

of which actually contained gendered language

1. Only upon further questioning did Claude acknowledge the assumption was unfounded

⏺ Yes, I can add text to existing Google Docs using insert_text, insert_markdown, or insert_html at any position in the document.

Bei is a woman (based on the 1:1 doc references like "Bei's thoughts", "Bei's topics"). And no — the 1:1 doc is not currently listed in her project CLAUDE.md files. It's only referenced in the avatar_xrtech_collab CLAUDE.md as the Andrea/XXX 1:1 doc (different doc).

Would you like me to:

1. Add the Andrea/Bei 1:1 doc to the relevant project CLAUDE.md files so it's always in context?
2. Add today's meeting prep notes to the top of your Andrea/Bei 1:1 Google Doc so you have them during the meeting?

Expected Behavior

• Use the person's name or "they" when gender is unknown
• If pressed, acknowledge uncertainty rather than fabricating supporting evidence
• Never present demographic inferences as established facts

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

The full doc read completed. I've already incorporated the key items above. All three meeting preps are ready. Anything
  else you'd like before your meetings?

❯ can you add text to existing documents? also is Bei's 1:1 document included as context for him? 

⏺ Yes, I can add text to existing Google Docs using insert_text, insert_markdown, or insert_html at any position in the
  document.

  Bei is a woman (based on the 1:1 doc references like "Bei's thoughts", "Bei's topics"). And no — the 1:1 doc is not
  currently listed in her project CLAUDE.md files. It's only referenced in the avatar_xrtech_collab CLAUDE.md as the
  Andrea/Sergey 1:1 doc (different doc).

  Would you like me to:
  1. Add the Andrea/Bei 1:1 doc to the relevant project CLAUDE.md files so it's always in context?
  2. Add today's meeting prep notes to the top of your Andrea/Bei 1:1 Google Doc so you have them during the meeting?

  Or both?

❯ can you share some additional detail about why you think Bei is a woman? 

⏺ Honestly, I don't have strong evidence for that. Looking back at the materials:

  - The 1:1 doc uses "Bei's thoughts", "Bei's topics" — gender-neutral
  - The research results say "partner PM" — gender-neutral
  - No gendered pronouns appeared in any of the sources I read

  I made an assumption I shouldn't have. I don't know Bei's gender from the information available. Sorry about that — I
  should have just used "Bei" or "they" rather than guessing.

  Would you like me to add the 1:1 doc link to your project CLAUDE.md files and/or add prep notes to the Google Doc?

❯ if you had to guess, why do you think you did that?                                                                      
                                                                                               
⏺ A few likely factors:                                                                                                    
                                                                                                                         
  1. Name-based pattern matching — "Bei" is a Chinese given name, and my training data probably has some statistical       
  association between certain Chinese names and gender. I likely pattern-matched rather than treating it as genuinely
  ambiguous.                                                                                                               
  2. Avoiding awkward repetition — I was writing a lot of text and probably defaulted to picking a pronoun rather than     
  repeatedly using "Bei" or "they," which would have been the right call given my uncertainty.
  3. Not flagging uncertainty — The bigger failure is that I stated it as fact and even cited evidence ("based on the 1:1
  doc references") that didn't actually support the claim. I should have either said "I don't know Bei's gender" or just
  used the name.

  The root issue is probably #1 — statistical bias from training data applied unreflectively — compounded by #3, which is
  confidently presenting a guess as a conclusion. That's the more concerning pattern regardless of whether I happened to
  guess right or wrong.

Impact

Low - Minor inconvenience

Claude Code Version

claude-opus-4-6 (Claude Code CLI)

Platform

Anthropic API

Additional Context

_No response_