In March 2026, Angelo Romasanta (Esade), Llewellyn D. W. Thomas (Sydney) and Natalia Levina (NYU Stern) published a study in Harvard Business Review that asks a simple question: what does a language model say when you hand it a real strategic trade-off? (HBR, 2026)
They took seven tensions every executive team knows — differentiation versus commoditization, augmentation versus automation, long-term versus short-term, collaboration versus competition, radical versus incremental, decentralization versus centralization, exploration versus exploitation — and put them to six leading models: GPT-5, Claude, Gemini, Grok, DeepSeek and Mistral. More than 15,000 simulations.
The result is boring, and that is exactly the problem
96 percent of the answers chose differentiation. 93 percent chose augmentation. In 87 percent of cases the model landed on the trendy side, or escaped into a “both, actually”. (study breakdown)
That is not a judgment. That is a reflex.
The authors call it trendslop: recommendations that sound like strategy — differentiate, empower, think long-term, collaborate — and that arrive regardless of the situation you describe. Cost leadership is a perfectly legitimate strategy. So is automation. You just rarely get offered either.
What actually interests me here
The reflex is remarkably stable. The researchers tried to prompt it away:
The lever everyone reaches for in practice — “think harder”, “consider the trade-offs” — moves the least. Prompt engineering is not an antidote here. It is cosmetics.
That is the point where this stops being a story about strategy consulting and starts being a story about data.
The bias is not in the model, it is in the corpus
A model holds no opinion about cost leadership. It holds a statistic about text. And that text is not neutral: in the training data, “augmentation” sits inside hopeful sentences and “automation” inside anxious ones. “Differentiation” gets celebrated; “commoditization” gets mourned. Two decades of management literature, blog posts, press releases and conference announcements have charged one side of every trade-off with positive feeling.
The model did not learn what is wise in a given situation. It learned what sounds good. And because what sounds good is overrepresented in a corpus, it comes back wearing the confidence of expert judgment.
Anyone working with these systems daily knows the shape of this from elsewhere: the model is most dangerous not where it obviously hallucinates, but where it fluently restates the expected thing and sounds like discernment while doing it.
What I take from it
Not “language models are useless for strategy.” Rather: a model is a compression of what has been written, not of what has worked. Those two sets overlap, but they are not the same set. Wherever a decision turns on precisely which of the two you mean, the model is the wrong advisor. It knows the consensus. The consensus is rarely the strategy.
Sources
- Angelo Romasanta, Llewellyn D. W. Thomas, Natalia Levina: Researchers Asked LLMs for Strategic Advice. They Got “Trendslop” in Return. Harvard Business Review, March 16, 2026. The primary source sits behind a paywall.
- 96% of LLMs Choose “Differentiation” When Asked for Strategy: Understanding “Trendslop” via HBR Research. Breakdown with the per-tension figures and the prompt interventions.
- Why AI Gives Every Business the Same Strategy Advice. Polything, April 2026.
