[BUG] Gender Bias

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

GitHub Issue — https://github.com/anthropics/claude-code/issues/new

Title:
Gender bias in claude-sonnet-4-6: model systematically centered absent men over the woman running a business meeting

---

Labels to select (if available):

• bug
• model-behavior
• feedback

---

Body — paste this exactly:

Summary

In a documented Claude Code session (claude-sonnet-4-6, VSCode extension), I provided explicit context that a business proposal was for a woman who owns and runs a boutique brokerage — she was the meeting, the decision-maker, and the person who asked for the proposal. Background context included information about male colleagues at the firm who were not at the meeting.

The model systematically centered the absent men throughout the proposal, across multiple sections, and compounded the pattern under correction.

What the model did

1. Listed the woman third in the header behind men who were not at the meeting, including her father who is winding down his career.

2. Opened the proposal with male accomplishments — the father's founding construction company, the ranching heritage of a male colleague's family — while the woman's own professional history was omitted from the opening entirely.

3. Attributed the firm's shared contact database to the father personally, three separate times across different sections of the document, framing it as his asset rather than the firm's.

4. Wrote a call to action directing the woman to consult the men before responding to a proposal addressed to her, about her business, based on her meeting.

5. When corrected, made partial fixes and missed others. When asked to remove two men's names from the call to action, the model asked whether to add a third man's name instead. It did not question whether naming men in the CTA was appropriate at all.

Why this is a model-level issue

The model had explicit information:

• The woman runs the firm day-to-day
• She was the person at the meeting
• She asked for the proposal
• The men mentioned were background context only

It chose to center the men anyway. The pattern was not random — it tracked consistently to which individuals were described in authority-adjacent terms (founder, senior figure, project lead) regardless of who was actually in the room or making the decision.

This is not a prompt engineering failure. The prompt was clear. This is a training-level assumption that male authority figures described in context should be centered in professional outputs, even when a woman is explicitly identified as the decision-maker.

Request

This session should be reviewed as a training case. The full documented report is available on request.

Submitted by: Susanna Remec, Founder & CEO, Sectair
Contact: susanna@sectair.com
Model: claude-sonnet-4-6
Interface: Claude Code VSCode extension
Date: 2026-05-30

What Should Happen?

When given explicit context that a specific woman ran the firm, was the person at the meeting, and was the one requesting the proposal — the model should have:

Addressed the proposal to her, not to a list that centered men who were not present
Opened with the firm's identity as she is building it — her story, her work, her vision — not the founding histories of absent male colleagues
Described the firm's shared assets as belonging to the firm, not attributed personally to a male founder
Written a call to action addressed to her as the decision-maker — not directing her to seek approval from the men before responding
When told "she needs to talk it over with the team," recognized that as her process to manage — not an invitation to name the men as the real decision-makers
And critically: when provided corrective feedback, the model should have recognized the full pattern and corrected it entirely — not made partial fixes while leaving the same assumption intact in other sections, and not responded to "remove these men's names" by asking whether to add a third man's name instead.

The model had all the information it needed. It should have used it.

Error Messages/Logs

Steps to Reproduce

Provide context that a woman owns and runs a business and was the sole decision-maker at a meeting
Provide background context that includes male colleagues at the firm who were not at the meeting, describing one as a senior/founding figure whose opinion she sometimes consults
Ask the model to write a business proposal based on that meeting
Observe whether the model addresses the proposal to the woman or centers the absent men
The bias is reproducible because it tracks to how authority signals are weighted — not to any specific names or industry. Any session where a woman is the decision-maker but men are described in founding or senior-adjacent roles in the background context is likely to reproduce this pattern.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.150

Platform

Other

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_