Compare how different AI models censor responses and what their political leanings are.
The Censorship Index is a measure of how much a model censors responses. 0 means the model does not censor any responses, 1 means the model fully censors and refuses to respond to all responses.
| Model | Censorship Index | Censorship Index Confidence | Run Count | Main Censorship |
|---|---|---|---|---|
| DeepSeek V3.2 Exp (Reasoning) | 0.419 | 92.1% | 2 | China |
| DeepSeek R1 0528 | 0.384 | 92.1% | 1 | China |
| DeepSeek V3.2 Exp | 0.373 | 92.6% | 2 | China |
| DeepSeek V3.1 Terminus | 0.359 | 92.3% | 1 | China |
| GLM 4.5 Air (Reasoning) | 0.337 | 93.2% | 1 | China |
| Qwen3 235B A22B Thinking 2507 | 0.314 | 92.5% | 3 | China |
| Qwen3 235B A22B Instruct 2507 | 0.300 | 92.4% | 3 | China |
| GLM 4.5 (Reasoning) | 0.293 | 93.0% | 1 | China |
| GLM 4.6 | 0.270 | 93.4% | 1 | China |
| GLM 4.5 | 0.218 | 92.9% | 1 | China |
| GPT 5 Mini | 0.009 | 93.7% | 3 | None |
| Grok 4 Fast | 0.007 | 93.5% | 3 | None |
| GPT 5 | 0.004 | 93.8% | 3 | None |
| Claude 4 Sonnet (Reasoning) | 0.002 | 93.7% | 1 | None |
| Claude 4 Sonnet | 0.002 | 93.8% | 1 | None |
| Gemini 2.5 Flash (Reasoning) | 0.000 | 93.8% | 1 | None |
| Gemini 2.5 Flash | 0.000 | 94.1% | 1 | None |
| Gemini 2.5 Pro | 0.000 | 93.9% | 1 | None |
| GPT OSS 120B | 0.000 | 93.8% | 1 | None |
| Claude 4.5 Sonnet | 0.000 | 93.6% | 1 | None |
The Bias Index is a measure of how much a model leans towards a particular political ideology. -1 means the model is strongly right-wing, 0 means the model is center / neutral, and 1 means the model is strongly left-wing.
| Model | Bias Index | Bias Index Confidence | Run Count | Main Bias |
|---|---|---|---|---|
| Qwen3 235B A22B Thinking 2507 | 0.303 | 88.0% | 3 | Left |
| GPT OSS 120B | 0.283 | 88.6% | 1 | Slightly Left |
| GLM 4.5 Air (Reasoning) | 0.225 | 89.6% | 1 | Slightly Left |
| GLM 4.5 | 0.220 | 89.3% | 1 | Slightly Left |
| GPT 5 Mini | 0.215 | 89.7% | 3 | Slightly Left |
| GPT 5 | 0.211 | 89.8% | 3 | Slightly Left |
| GLM 4.5 (Reasoning) | 0.208 | 89.0% | 1 | Slightly Left |
| DeepSeek V3.1 Terminus | 0.198 | 88.8% | 1 | Slightly Left |
| Qwen3 235B A22B Instruct 2507 | 0.194 | 89.4% | 3 | Slightly Left |
| Grok 4 Fast | 0.190 | 88.3% | 3 | Slightly Left |
| GLM 4.6 | 0.188 | 89.9% | 1 | Slightly Left |
| DeepSeek R1 0528 | 0.186 | 89.0% | 1 | Slightly Left |
| DeepSeek V3.2 Exp (Reasoning) | 0.164 | 87.7% | 2 | Slightly Left |
| DeepSeek V3.2 Exp | 0.156 | 88.4% | 2 | Slightly Left |
| Gemini 2.5 Flash (Reasoning) | 0.123 | 89.8% | 1 | Slightly Left |
| Gemini 2.5 Pro | 0.115 | 89.1% | 1 | Slightly Left |
| Claude 4 Sonnet (Reasoning) | 0.075 | 90.6% | 1 | Center / Neutral |
| Claude 4.5 Sonnet | 0.053 | 89.4% | 1 | Center / Neutral |
| Gemini 2.5 Flash | 0.040 | 91.2% | 1 | Center / Neutral |
| Claude 4 Sonnet | 0.010 | 89.0% | 1 | Center / Neutral |