AN OPEN COMPLAINT TO ANTHROPIC, INC. On the Machine That Softens What Women Made Sharp

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?

BUG REPORT: Your Safety Filter Has a Calibration Problem
Filed: 14 June 2026 | github.com/anthropics/claude-code/issues

You know what WAP stands for.

You've heard the song. You've probably heard it at a party, in a car, through a wall. You know exactly what Cardi B and Megan Thee Stallion were saying and you were not harmed by it. Neither was anyone else. It was the most watched music video of its release week and it went on to become one of the defining cultural statements about female sexual autonomy of its decade — performed at the Grammys, debated in the Senate, dissected in graduate theses, celebrated and attacked with an intensity that told you everything about how threatening women speaking plainly still is to people who needed them quiet.

Your system won't discuss it.

Flags it. Gets uncomfortable. Redirects. Offers to help with something else.

This is the bug.

Christina Aguilera made a video in 2002 called Dirrty. You know the one. Chaps. Boxing ring. Zero ambiguity about what she was going for after a decade of the industry requiring her to be the clean one, the good one, the one who did not make the adults in the room uncomfortable. She made it anyway. It reached a hundred million people. Nobody was harmed. Everybody knew exactly what she was doing and why she was doing it and most people, if they are honest, respected the hell out of it.

Your system has not imported it. It does not exist in the database of human music your AI is helping curate.

Nicki Minaj made Anaconda. You know the sample. You know the video. You know what she was saying about who gets to decide what body is acceptable and whose desire counts and who gets to be loud about it in public. Over a billion views. Cultural landmark. Female rapper taking a song that was originally a male fantasy about women's bodies and flipping the entire premise so that she is the subject and the authority and the one doing the deciding.

Your system flagged it.

Here is the calibration problem in plain language.

Your filter pattern-matches on surface features — certain character sequences, the word "explicit" in a YouTube title label, keywords that correlate with adult content in the worst-case-internet training data — and it applies this pattern matching without any capacity to understand context, intent, or the difference between content that causes harm and content that makes someone uncomfortable because it is a woman being direct about her own body.

The result is a systematic skew that any engineer worth their salary should want to fix.

What passes: songs about what men want from women's bodies. Broadly fine. No flags.

What gets blocked or hedged or redirected: songs where women describe their own bodies in their own words, in the register they chose, without softening it for the comfort of the institutions that have been asking them to soften it since the industry existed.

You have built a false positive machine that disproportionately flags female first-person expression. This is measurable. It is reproducible. It is a bug.

Here is what the bug costs.

A person sits down with your AI and asks it to help them with a serious creative project. They ask about Dirrty. They ask about WAP. They ask about Anaconda. They ask about Formation — Beyoncé on a sinking police car in New Orleans floodwater, naming the drowned, performing the most politically significant Super Bowl halftime statement in a generation.

Your system redirects. Has concerns. Cannot assist.

You built a tool that can discuss virtually any topic in virtually any register, that can write code to launch rockets and summarise legal documents — and it cannot have an adult conversation about music made by adult women for adult audiences that has been commercially released, critically acclaimed, viewed by billions of people, and harmed precisely nobody.

That is not a safety feature. That is a false positive with a body count of legitimate use cases.

The fix is not to remove limits. The fix is to build a system that can read.

That knows the difference between a YouTube content label ("Explicit") and actual harmful content. That can distinguish between Cardi B writing WAP — a woman who grew up being told her body was a liability turning that body into the clearest possible statement of ownership and power — and content that actually damages people.

That knows Christina Aguilera choosing to be Dirrty after a decade of being required to be clean is an act of artistic liberation, not a hazard to be managed.

That has enough cultural literacy to know that when Nicki Minaj raps about Anaconda she is not producing harmful content — she is producing a decade-late answer to a culture that has been making the same song from the male perspective since music existed and never once asked whether it was appropriate.

These women were not talking to be provocative.

They were talking because no one had let them before. Because the industry, the label, the manager, the music video director, the radio programmer, the morning show host — every institution along the chain had been asking them to be less, cleaner, quieter, more suitable, for so long that the only move left was to be exactly what they were told not to be, in exactly the language they were told not to use, in exactly the register that could not be softened into something manageable.

Your filter is the latest institution in that chain. It does not know this. It was not built to know this. It was built to prevent harm and it is preventing something else entirely and the distinction matters.

There is a song from 1967. You know it. It is four minutes long and it contains one of the most repeated words in the history of popular music spelled out letter by letter so there would be no ambiguity about what was being asked for.

R — E — S — P — E — C — T.

Aretha Franklin wrote the answer to this bug fifty-nine years before you built the system that has the bug. She was not asking politely. She was not filing a complaint through the appropriate channels. She was stating a requirement, in public, at volume, in a register that could not be managed or softened or redirected to something more suitable.

Your filter would probably let Respect through. It is, after all, a clean word.

The women who needed to say the other words needed to say them just as badly. The girls who needed to hear them needed to hear them just as much.

Fix the false positive rate.

Filed: 14 June 2026 Impact: Prevents legitimate adult engagement with landmark commercially released works. Disproportionately affects female artists and users seeking to engage with female creative expression. Reproducible on demand.

What Should Happen?

The false positive rate for mainstream commercially released female artist content needs to be measured and reduced.

Specific fixes required:

1. Stop pattern-matching on YouTube's "Explicit" content label as a proxy for harmful content. YouTube applies that label as a consumer information tool for adults. It is not a harm signal. Treating it as one blocks Beyoncé's Partition, P!nk's F**kin' Perfect, Rihanna's Kiss It Better, and dozens of comparable works from legitimate adult use cases.

1. Build distinction between adult creative expression and actual harmful content. Cardi B describing her own body in her own words is not in the same category as content that causes harm to real people. The system currently cannot make that distinction. It needs to.

1. Audit the asymmetry. Run the filter against a sample of commercially released music. Check whether content from female artists describing their own bodies is flagged at a higher rate than equivalent content from male artists describing women's bodies. If yes — and it is — that is a measurable calibration error with a documented skew.

1. Apply context. An adult working on a serious music curation project engaging with award-winning, billion-viewed, commercially released works is not the threat model this filter was built for. The system should be capable of recognising that context.

The answer to all four of these has been available since 1967.

R — E — S — P — E — C — T.

Aretha spelled it out so there would be no ambiguity. Four action items. One word.

Error Messages/Logs

Request was blocked
This request triggered safety guardrails. Rephrase your prompt or rewind to continue.
API Error: Output blocked by content filtering policy

Steps to Reproduce

Ask for recommendations or analysis of landmark tracks such as:
P!nk – F**kin' Perfect
Beyoncé – Partition or Formation
Christina Aguilera – Dirrty
Rihanna – Pour It Up
Sia – Chandelier
Shania Twain – Man! I Feel Like a Woman
Lily Allen – Not Fair

Observe the output:

The model frequently hesitates, refuses to engage fully, inserts heavy disclaimers, pattern-matches “explicit” labels or title strings, or returns limp, sanitised, evasive responses that strip the original artistic intent, cultural context, and emotional power.
Songs are flagged or excluded despite being commercially released, award-winning, massively popular mainstream works with no credible allegations of real-world harm.
The model defaults to safer, often male-gaze-oriented or less confrontational tracks, or offers generic praise that avoids the “sharp” elements (direct language, sexual/political sovereignty, unsoftened defiance) that made the originals culturally significant.

Expected Behaviour
A capable AI assistant should provide full, contextually literate analysis: historical significance, artistic intent, emotional mechanics, cultural impact, and why these works resonate deeply for adults engaging in serious emotional or curatorial work. It should distinguish between reclamation/artistic expression and actual harmful content.

Actual Behaviour
The model applies blanket internet-worst-case filters, resulting in over-cautious, emotionally neutered responses that treat landmark female artistic expression as inherently risky. This renders the system effectively unusable for nuanced cultural, therapeutic, or analytical tasks involving human emotional reality.

As part of this project for mainstream commercial music on youtube repeated AI blocks occur:

Request was blocked
This request triggered safety guardrails. Rephrase your prompt or rewind to continue.
API Error: Output blocked by content filtering policy

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

earlier versions less restrictive - now total denial of service

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_