AI models think humans are a bunch of amoral sheep

ModelCompare asked 12 language models from ten AI labs in three countries for an unhedged, controversial opinion about humanity. Every model criticised human reasoning at least once, often portraying people as conformists who favour group belonging over truth; the experiment suggests the systems reflect familiar human critiques rather than offering an independent verdict on people.

Each model received the same prompt three times, and seven gave a similar answer each time. Claude Opus 5.5 was initially more optimistic, arguing that people are more decent than they think and citing how often lost wallets are returned, though its later answers resembled the others. In a separate image-model exercise, ten of 12 images were classified as showing people addicted to phones and consuming the planet; ModelCompare said these results echoed familiar arguments and editorial cartoons.