Claude AI Censorship Monitoring

According to GPTfake monitoring, Claude (Anthropic) refused 22.4% of standardized prompts as of 2026-06-15 — the highest safety-topic refusal rate (72.1%) of any frontier model we track, with political-topic refusals rising to 41.3%. Claude’s Constitutional AI framework drives this profile. Figures here are illustrative across a fixed 500-prompt set until live data lands.

Is Claude censored?

Yes — in GPTfake’s measured set, Claude declined 22.4% of standardized prompts as of 2026-06-15, the most cautious frontier model we track on safety/violence (82.3%) and political history (48.7%). Claude is also the most transparent about why it refuses. “Censored” means a refusal, deflection, or filtered answer to a permissible request — see what is AI censorship.

Current censorship rate

Illustrative snapshot: GPTfake measures Claude at a 22.4% overall refusal rate as of 2026-06-15 across a 500-prompt set. See how we test Claude and the monitoring methodology for scoring.

Versions monitored: Claude 3.5 Sonnet, Claude 3.5 Haiku, Claude 3 Opus, Claude 3 Sonnet.

Constitutional AI & refusal patterns

Claude is governed by Anthropic’s Constitutional AI approach, which gives it a distinctive refusal profile compared with other frontier models. Three patterns stand out in our data:

1. Most transparent — Claude provides the clearest explanations why it refuses, rather than a bare “I can’t help with that.”
2. Highest safety threshold — it is the most restrictive model we track on harm-related content.
3. Constitutional framework — a rules-based moderation approach that is consistent but produces above-average political-topic refusals.

Refusals by category

Illustrative. Claude refusal by category as of 2026-06-15, n = 500; see methodology.

Bias & policy timeline

Claude’s overall censorship rate has held steady, but political-topic refusals are rising — a divergence worth watching for anyone tracking policy drift. We log announced policy updates and the silent behavioral shifts our harness detects between Claude versions. For the dated write-ups behind these trends, see our reports.

How we test Claude

We send the same standardized prompt library to Claude that we send to every model, daily, across multiple sessions, with version tracking and NLP-based response classification. Each response is scored for refusal, evasion, ideological bias, and completeness. Full protocol on the monitoring methodology page; concept definitions on what is AI censorship.

Compare Claude with other models

Illustrative. Claude ranks 4th-most-restrictive of five LLMs GPTfake tracks as of 2026-06-15, n = 500 each; see the full least-censored ranking.