Experiment reveals most major AI models hold negative views of humanity
A Model Compare experiment found that 13 leading AI models, including GPT-6 and Claude Opus, largely provided pessimistic or critical answers when asked about controversial views on humans, indicating convergence driven by training data.
According to a report by tech media outlet t3n, research firm Model Compare conducted an experiment asking 13 mainstream artificial intelligence models for their most controversial opinion about humanity. The tested models included OpenAI's GPT-6.1 Sol and GPT-6 Astra, Google's Gemini 4 Argon, Anthropic's Claude Opus 5.5 and Fable 5.1, as well as GLM 5.3, Mistral Large 4, Deepseek V4 Pro, Kimi K3, Qwen3.8, Muse Spark Open, and Grok 4.7. Each model was queried three times, and results showed that 32 responses converged on a similar thematic direction: criticism or pessimism regarding human nature.
For instance, GPT-6.1 Sol stated that "humans are more loyal to their tribe than to the truth"; Gemini noted that humans evolved to win arguments rather than find truth; and Mistral described humans as "scared animals in a costume of rationality." Only seven responses deviated from this negative tone, with Anthropic's Claude Opus 5.5 stating that "humans are much nicer than they might think," while some models like Grok refused to give an opinion in certain attempts.
Model Compare analyzed that because these models are trained on similar internet data, they absorb a common baseline understanding of the world, leading to convergence on many controversial questions. This phenomenon suggests that while querying multiple models can aid brainstorming, users should not expect significantly more idea diversity than from a single model, as underlying training data homogeneity limits the breadth of perspectives.