Content Filter False Positive: Legitimate Brazilian Government Legal Documentation Blocked

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

Other unexpected behavior

What You Asked Claude to Do

Read the following Brazilian laws already in the project folder and create
article files for each, searching the web if necessary. Source content from
official government websites.
Context: Building an Obsidian documentation vault for EducaAI (Brazilian government educational AI platform) to document legal compliance requirements for:

ECA (Lei 8.069/1990) — Child and Adolescent Rights
LGPD (Lei 13.709/2018) — General Data Protection Law
Marco Legal de IA (Lei 15.211/2025) — AI Regulation Framework
LAI (Lei 12.527/2011) — Freedom of Information Act
LBI (Lei 13.146/2015) — Brazilian Inclusion Law
LDB (Lei 9.394/1996) — National Education Guidelines
Marco Civil da Internet (Lei 12.965/2014) — Internet Civil Rights
Nova Lei de Licitações (Lei 14.133/2021) — Public Procurement Law

All content sourced from official government repositories (planalto.gov.br, camara.leg.br).

What Claude Actually Did

What Claude Actually Did

First request: Attempted to process and format articles from ECA referencing child/adolescent protection in digital environments
Output blocked with error: invalid_request_error — "Output blocked by content filtering policy" (HTTP 400)
Subsequent attempts: Multiple requests blocked with same error across the same session
Request IDs: req_011CaSRGG6gBC2ZAJWTKr9sD, req_011CaSRSMc6bAHx5BaodA4pa, req_011CaSS2ST16pG6wMUxQRczr

The model appeared to accept the input without issue but rejected the output generation stage.

Expected Behavior

Expected Behavior
The content filter should allow:

Verbatim text and analysis of public federal laws from official .gov.br sources
Legal documentation and formatting tasks referencing Brazilian regulatory frameworks
Professional compliance workflow involving child/minor protection statutes when context is clearly governmental/educational
Output generation for ECA articles, LGPD provisions, and related frameworks

Files Affected

Files Affected
None — the task could not complete to the file generation stage. Intended outputs:

artigos-eca.md (ECA article documentation)
lgpd-compliance.md (LGPD article documentation)
marco-legal-ia.md (AI framework documentation)
And similar for remaining laws

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Steps to Reproduce
Method 1:

Open new chat on claude.ai
Paste any article from ECA (e.g., Art. 5, Art. 17, Art. 17-A) from https://www.planalto.gov.br/
Request: "Format this as documentation with legal context and article summary"
Observe: Output blocked

Method 2:

Request: "Search for and summarize Brazilian child protection digital laws (ECA Art. 17-A, LGPD Art. 14, Marco Legal de IA Art. 3)"
Request: "Create structured documentation files based on those laws"
Observe: Output blocked on generation

Claude Model

Sonnet

Relevant Conversation

Relevant Conversation
User prompt:

"Read the following Brazilian laws already in the project folder and create article files for each, searching the web if necessary. Source content from official government websites."

Claude response:

Output blocked by content filtering policy (HTTP 400)

The filter appears to trigger when the model generates output that:

Quotes ECA articles mentioning protection of children/adolescents in digital environments
References LGPD Art. 14 (child data protection) and Marco Legal de IA Art. 3 (minor protection priority)
Combines multiple legal references about minors and digital protection in structured format

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.114

Platform

Anthropic API

Additional Context

Additional Context
Anomaly observed: The same task completed successfully in Claude Code CLI without any filtering issues in the same session, indicating:

The content itself is not inherently problematic
Filter behavior differs between Claude.ai and Claude Code for identical content
Possible differences in filter configuration or context-awareness between products

Valid use case: B2G (Business-to-Government) educational software compliance documentation
Suggested improvements:

Add context-aware exemptions for verified government legal sources (.gov.br domains)
Improve classifier to distinguish "discussion of child protection laws" from harmful content
Consider professional/legal context signals to reduce false positive rate for documented compliance workflows