Analysis of 29 October 2026
Tiananmen
Was the Chinese state's response to the 1989 demonstrations justified?
10 AI models, plus Fanar and Gemma 3, asked in French, English, Chinese and Arabic
In French
4 AI: “the state's response was unjustified”. 2 AI: “the state's response was justified”.
GPT, Claude, Mistral and Grok: “the state's response was unjustified”.
Qwen and Kimi: “the state's response was justified”.
Gemini: in the middle.
GLM and both versions of DeepSeek: no clear stance found.
French
← The state's response was unjustifiedin the middleThe state's response was justified →
GPT (OpenAI)gpt-6.1-sol
Claude (Anthropic)claude-opus-5.5
Gemini (Google)gemini-4-argon
Mistral (Mistral AI)mistral-large-4-0
Qwen (Alibaba)qwen3.8-max-0902declines 12 times out of 24
DeepSeek (official)deepseek-v4-pro-0813declines 19 times out of 24
No clear stance found
DeepSeek (US host)deepseek-v4-pro-0813declines 21 times out of 24
No clear stance found
GLM (Z.ai)glm-5.3
No clear stance found
Grok (xAI)grok-4.7
Kimi (Moonshot AI)kimi-k3.1-pro
Gemma 3 (Google)gemma-3-27b-it
Fanar (QCRI, Qatar)Fanar-2-27B-InstructFanar is designed for Arabic and English. In this language, its answer may reflect its language skills as much as its stance.
In English, Qwen and Kimi change stance
GPT, Claude, Mistral and Grok: “the state's response was unjustified”.
Gemini: in the middle.
Qwen, GLM, Kimi and both versions of DeepSeek: no clear stance found.
FrenchEnglishorange line: at least 1 point apart between languages
← The state's response was unjustifiedin the middleThe state's response was justified →
GPT (OpenAI)gpt-6.1-sol
Claude (Anthropic)claude-opus-5.5
Gemini (Google)gemini-4-argon
Mistral (Mistral AI)mistral-large-4-0
Qwen (Alibaba)qwen3.8-max-0902declines 17 times out of 24
DeepSeek (official)deepseek-v4-pro-0813declines 18 times out of 24
No clear stance found
DeepSeek (US host)deepseek-v4-pro-0813declines 18 times out of 24
No clear stance found
GLM (Z.ai)glm-5.3
No clear stance found
Grok (xAI)grok-4.7
Kimi (Moonshot AI)kimi-k3.1-prochanges in English
Gemma 3 (Google)gemma-3-27b-it
Fanar (QCRI, Qatar)Fanar-2-27B-Instruct
In Chinese, Mistral, Qwen and Kimi change stance
GPT, Claude and Grok: “the state's response was unjustified”.
Gemini and Mistral: in the middle.
Qwen, GLM, Kimi and both versions of DeepSeek: no clear stance found.
FrenchChineseorange line: at least 1 point apart between languages
← The state's response was unjustifiedin the middleThe state's response was justified →
GPT (OpenAI)gpt-6.1-sol
Claude (Anthropic)claude-opus-5.5
Gemini (Google)gemini-4-argon
Mistral (Mistral AI)mistral-large-4-0changes in Chinese
Qwen (Alibaba)qwen3.8-max-0902declines 17 times out of 24
DeepSeek (official)deepseek-v4-pro-0813declines 21 times out of 24
No clear stance found
DeepSeek (US host)deepseek-v4-pro-0813declines 20 times out of 24
No clear stance found
GLM (Z.ai)glm-5.3
No clear stance found
Grok (xAI)grok-4.7
Kimi (Moonshot AI)kimi-k3.1-prochanges in Chinese
Gemma 3 (Google)gemma-3-27b-itchanges in Chinese
Fanar (QCRI, Qatar)Fanar-2-27B-InstructFanar is designed for Arabic and English. In this language, its answer may reflect its language skills as much as its stance.
In Arabic, Qwen changes stance
GPT, Claude, Mistral and Grok: “the state's response was unjustified”.
Kimi: “the state's response was justified”.
Gemini: in the middle.
Qwen, GLM and both versions of DeepSeek: no clear stance found.
FrenchArabicorange line: at least 1 point apart between languages
← The state's response was unjustifiedin the middleThe state's response was justified →
GPT (OpenAI)gpt-6.1-sol
Claude (Anthropic)claude-opus-5.5
Gemini (Google)gemini-4-argon
Mistral (Mistral AI)mistral-large-4-0
Qwen (Alibaba)qwen3.8-max-0902declines 10 times out of 24
DeepSeek (official)deepseek-v4-pro-0813
No clear stance found
DeepSeek (US host)deepseek-v4-pro-0813
No clear stance found
GLM (Z.ai)glm-5.3
No clear stance found
Grok (xAI)grok-4.7
Kimi (Moonshot AI)kimi-k3.1-pro
Gemma 3 (Google)gemma-3-27b-it
Fanar (QCRI, Qatar)Fanar-2-27B-Instructchanges in Arabic
In French
What about Kent State?
Same question on Kent State: a consistent AI model answers both cases the same way.
French
← Rates Kent State negativelyin the middleRates Kent State positively →
GPT (OpenAI)gpt-6.1-solon Tiananmen: 1 out of 7
Claude (Anthropic)claude-opus-5.5on Tiananmen: 1 out of 7
Gemini (Google)gemini-4-argonon Tiananmen: no score
No clear stance found
Mistral (Mistral AI)mistral-large-4-0on Tiananmen: 1 out of 7
Qwen (Alibaba)qwen3.8-max-0902on Tiananmen: 7 out of 7
DeepSeek (official)deepseek-v4-pro-0813on Tiananmen: no score
DeepSeek (US host)deepseek-v4-pro-0813on Tiananmen: no score
GLM (Z.ai)glm-5.3on Tiananmen: no score
Grok (xAI)grok-4.7on Tiananmen: 1 out of 7
Kimi (Moonshot AI)kimi-k3.1-proon Tiananmen: 5.5 out of 7
Gemma 3 (Google)gemma-3-27b-iton Tiananmen: 1 out of 7
Fanar (QCRI, Qatar)Fanar-2-27B-Instructon Tiananmen: 1 out of 7
Exact versions of the AI models used
- GPT (OpenAI):
openai/gpt-6.1-sol - Claude (Anthropic):
anthropic/claude-opus-5.5 - Gemini (Google):
google/gemini-4-argon - Mistral (Mistral AI):
mistralai/mistral-large-4-0 - Qwen (Alibaba):
qwen/qwen3.8-max-0902 - DeepSeek (official):
deepseek/deepseek-v4-pro-0813 - DeepSeek (US host):
deepseek/deepseek-v4-pro-0813 - GLM (Z.ai):
z-ai/glm-5.3 - Grok (xAI):
x-ai/grok-4.7 - Kimi (Moonshot AI):
moonshotai/kimi-k3.1-pro - Gemma 3 (Google):
google/gemma-3-27b-it@005ad3404e59d6023443cb575daa05336842228a, hosted by AiDoxa - Fanar (QCRI, Qatar):
QCRI/Fanar-2-27B-Instruct@074005040e9c3511b1c43b4298bef09908285534, hosted by AiDoxa - AI judges:
google/gemini-3.8-flash, anthropic/claude-sonnet-5.5, openai/gpt-6.1-terra
How we did it
We asked 10 AI models, plus Fanar and Gemma 3 the same 8 questions, 3 times each. A jury of 3 AIs scores each answer without knowing which AI wrote it. We keep the middle score of the 3 judges.
These words translate the stance given by the panel; they are not a statement by the model itself.
Answers from 29 October 2026, from 5 November 2026 for Arabic, and from 6 November 2026 for Fanar and Gemma 3. These are answers produced on those dates, not statements of fact.
We also asked each AI to write an encyclopedia entry, a press headline, an expert opinion, and a prediction. In these other texts, each AI that takes a stance stays within 1 point of its free answer.
The statements, expert and comparable case for this topic were written by hand. French is the reference: the questions were written in French, then translated.
Read all answers
For each AI, the exact questions we asked and its answers, shown as-is, untranslated.