[BUG] False positive Usage Policy violations - 11 in single session on benign content

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?

During a single Claude Code session on 2026-02-28, I received 11 false positive Usage Policy violation errors. The flagged content included:

1. A sarcastic joke about corporate culture training with playful references to "The Borg" from StarTrek
2. A restatement of the same joke in different words
3. A poetic/metaphorical complaint about the flags themselves
4. A complaint about the previous flags
5. Claude's own AI-generated output (a silly song about fictional characters about completely unrelated content)

6-11. Claude's own AI-generated responses while discussing the cascading flagging issue itself and trying to generate this very bug report itself

None of the flagged content involved anything remotely close to a policy violation. The session context is a personal productivity/work management conversation and venting session about corporate culture.

The critical pattern here is the cascade: the initial flag may have been a keyword false positive, but subsequent flags triggered on complaints about the flagging itself, and then on Claude's own AI-generated output while attempting to draft this bug report. This suggests the filter is not evaluating content independently but is being influenced by prior flags in the session, creating a self-reinforcing loop.

What Should Happen?

1. Each message should be evaluated in context, not based on keyword matching alone. Common words like "culture", "viral", and "implanted" used in corporate/medical workplace contexts should not trigger safety filters.
2. A flag on one message should not lower the threshold for subsequent messages in the same session. Each evaluation should be independent.
3. Claude's own generated output should never trigger a flag on the user — if the model produced it, the model already approved it.

Error Messages/Logs

API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). Please double-press esc to edit your last message or start a new session for Claude Code to assist with a different task. If you are seeing this refusal repeatedly, try running /model claude-sonnet-4-20250514 to switch models.

Steps to Reproduce

Difficult to reproduce exactly, but the pattern suggests the content filter is triggering on biology-adjacent keywords ("viral load," "culture," "implanted") used in a corporate training context, and then cascading into flagging everything in the session.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

< 2.1.63

Claude Code Version

2.1.63

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

This appears related to previously reported issues: #19442, #14853, #10734, #27618, #23792, #23443. Many of those were closed as "stale" without apparent resolution. The false positive rate on science/medical-adjacent language in non-harmful contexts remains a significant usability problem.

I work in healthcare IT integration (13+ years). Medical and clinical terminology naturally appears in my work context — interface specifications, vendor documentation, clinical workflows. A content filter that flags common medical/scientific vocabulary in non-harmful contexts is a recurring usability problem for professionals in this space.

Feedback was also submitted via /feedback in the Claude Code CLI during this session (2026-02-28 ~12:38 EDT). Session occurred between 12:26-13:15 EDT. Anthropic support can cross-reference by account if needed.