Different AI models, asked the same strategy question, tend to agree — not because they reasoned it out, but because they share the same biases.

This explores whether different AI models, asked for business strategy advice, tend to land on the same answers, and whether that sameness comes from shared built-in leanings rather than independent reasoning. The short answer from the corpus is yes, and the result is starker than you might expect. In roughly 15,000 simulated strategy dilemmas, six different LLMs recommended the same side of every strategic tension they were tested on Do LLMs consistently favor the same strategic choices regardless of context?. The telling detail is what moved their answers. Changing the industry context shifted recommendations by only 11%. Simply swapping the order the options were listed in shifted them by 19%. The models responded more to how the question was laid out than to the business situation itself. The authors read this as models recombining fashionable strategy vocabulary rather than analyzing the case.

The shared lean also points in a particular direction. A strategy simulation benchmark found that mid-to-late 2025 frontier models scored *below* earlier models and below MBA students, because they consistently chose immediate profit over uncertain bets on future growth Do newer frontier LLMs actually make better strategic decisions?. So the models aren't just agreeing with each other. They're agreeing on a specific, short-term-leaning view of what good strategy looks like, and newer models aren't escaping it.

Where does that come from? Work outside strategy suggests these leanings are inherited from pretraining, not learned from the task. When LLMs act as recommender systems, they show position, popularity, and fairness biases that come from the language-model training objective and the makeup of the training text Where do recommendation biases come from in language models?. That's the same order-sensitivity seen in the strategy studies. Models also reproduce human-like belief biases, judging an argument's logic partly by whether its conclusion sounds plausible Do language models show the same content effects humans do?. If every major model learns from overlapping slices of the internet, including the same business-press clichés, matching recommendations are what you'd predict.

There's a twist on whether this sameness can be fixed. One study found that individual LLM rankings of business models are inconsistent and biased, but pooling many evaluations produces rankings that resemble human-expert judgments Can aggregated LLM rankings match human expert judgments?. The gains came mostly from *more* evaluations, not from *different* models or prompts. Read next to the shared-bias finding, that makes sense: if the models share leanings, adding a different model doesn't bring a genuinely independent view. Structure helps more than variety. In game-theory settings, LLMs drift away from rational play as games get more complex, but guided step-by-step workflows bring them back to near-optimal choices Do language models make rational strategic decisions in games?.

What you might not have thought to worry about is the human side. In a 348-person experiment, LLM help widened the range of factors people considered but did not make their predictions more accurate, and it reduced their sense of owning the decision Does using LLMs actually improve strategic decision making?. Put that together with the studies above. If many firms consult different AI tools that quietly share the same strategic lean, the outcome is not a range of perspectives. It's a homogenized one: many companies nudged toward the same short-term moves, with each decision-maker feeling less ownership of the choice.