Preview - Interview with Angelo Romasanta - "Trendslop" Author
Seven different AI models, asked the same corporate strategy question across 15,000 trials, gave the identical recommendation 96% of the time. Researcher Angelo Romasanta calls the pattern "trendslop" — and it's already inside your consultants' workflow. LLMs default to whichever option appears more often as "correct" in training data: differentiation over commoditization, hybrid compromises over binary strategic choices, regardless of case specifics. Reordering the same two options in a prompt shifted recommendations by 22% — proof the bias sits in architecture, not missing context. Reinforcement learning from human raters compounds this: confident, articulate answers get rated higher regardless of accuracy, training models toward persuasive sameness rather than correct specificity. Romasanta's fix isn't technical, it's procedural: use AI to expand the option set, never to converge on one. Withhold information deliberately, test single options in separate chats, and keep the first draft human — anchoring to a fluent wrong answer is hard to undo.
Not Rated Yet