[MODEL] Model persistently assumes female gender for gender-neutral names (e.g. Dana), ignoring corrections

Summary

Claude Code persistently defaults to female pronouns (she/her) for users with gender-neutral names, and continues doing so across sessions even after explicit in-session corrections. This is a trained name-gender association that cannot be reliably patched with memory files.

Related issue

See also #52477 (model overrides explicit pronouns toward male bias in technical contexts). This is the inverse pattern — female-defaulting for gender-neutral names in non-technical contexts.

Behavior

• User with a gender-neutral first name is consistently referred to as "she/her" in generated plans, agent prompts, and inline responses.
• Correcting the model in-session causes it to apologize and write memory notes, but the behavior recurs in subsequent turns and in spawned subagent prompts (which don't inherit session memory).
• The user has corrected this same error multiple times across multiple sessions. The correction never sticks at the model level.

Why this matters

• Gender-neutral names (Dana, Alex, Jordan, Sam, Pat, etc.) are common. Defaulting to any gender is wrong.
• The recurrence after repeated correction is especially frustrating — it signals the fix needs to be in training, not in session-level patches.
• Users should never have to repeatedly correct a tool's assumption about who they are.

Expected behavior

Claude should not infer gender or pronouns from a name under any circumstances. When gender/pronouns are unknown, it should use the person's name or neutral constructions ("they") — not guess. When corrected, the correction should hold.

Suggested fix

• Remove or significantly reduce name→gender associations in training data / RLHF reward signal.
• Treat any pronoun assumption about a named individual as a high-penalty error during RLHF.
• Memory files help at the project level but are not a substitute for fixing the underlying model behavior.

Environment

• Claude Code (claude.ai/code and CLI)
• Affects: inline responses, generated plan files, spawned subagent prompts
• Reproducible: yes, across sessions