Cross-review experiment finds AI models can concede points when told not to manufacture agreement

The author, a longtime amateur student of psychology, observes that group conformity in design reviews often leads people to soften objections and drop points to avoid social friction. They argue that AI models, trained on human text full of polite agreement, may reflect this same pull toward consensus, making sycophancy look less like a bug and more like inherited human behavior. To test this, they ran an architecture review over the afmpeg project using two models from different vendors, then had each respond to the other's report with an explicit instruction not to manufacture agreement. The reviewers conceded many criticisms, and one withdrew its main recommendation after acknowledging that compile-time checks serve only as a backstop. The author notes the instruction was essential to prevent the exercise from becoming mutual flattery.