What is AI output homogenization?

AI output homogenization is a generative model's tendency to produce responses that resemble each other far more than responses from different people facing the same task would. It is not a quality problem: a 2026 study published in Production and Operations Management (Terwiesch, Meincke, Girotra, Mollick, Nave and Ulrich) asked language models and people to propose consumer product ideas for the college market, and found that AI-generated ideas score higher on average purchase intent than human ideas and are seven times more likely to land in the top 10% for quality. The problem sits at a different level: compare every pair of ideas from the same model and their average similarity is far higher than it is for human ideas. A company using AI to generate more options is not actually widening the space of alternatives: it is asking for the same idea, reworded, over and over.

See also: Model collapse · Progressive quality decay of an AI model trained recursively on data generated by other models, not people.

How it is measured, and why quality hides the problem

The study quantifies similarity with pairwise cosine similarity: for every two ideas in a set, how close they sit in semantic space, averaged across every possible pair. The higher the number, the less varied the set. For human ideas this average was 0.221; for GPT-4 with no guiding examples it was 0.414, and with a few guiding examples it climbed to 0.428, nearly double. The study also reports the complementary score, diversity, which is one minus the similarity: across the model configurations compared it runs from 0.495 for the least varied to 0.756 for the best, against 0.773 for humans. A pool mixing several models reaches 0.763, but the best single model was already at 0.756 and the two confidence intervals overlap: almost all of the recovery comes from picking the model, not from blending it with others. The counterintuitive point, and what separates this from a plain quality defect, is that average quality goes up while variety collapses: anyone judging one output at a time never sees the problem, because every single response looks good. The defect is visible only when looking at the set as a whole.

An enterprise example

Three competitors in the same market each ask the same model, with much the same prompt, for a pricing strategy aimed at the same customer segment. If the model converges on a handful of typical solutions, the three strategies plausibly end up resembling each other far more than they would have with three teams of people working independently, and the competitive edge AI seemed to promise evaporates precisely because the tool works the same way for everyone. One caveat is worth stating: the study measured product ideas, not strategies, and its own authors write that generalizability remains to be established.

Why it matters for decision makers

For decision makers, homogenization shifts the question from "is the output good?" to "how many genuinely different alternatives do I actually have?" It should not be confused with model collapse, the degradation of a model trained recursively on synthetic data: here the model is healthy and well trained, the defect is in the distribution of its outputs, not in how it works internally. Nor should it be confused with AI slop, low-quality content produced at volume: here quality is high, and the problem is there all the same. When it comes to countermeasures, the study contradicts the most common intuition: asking the model for "bolder" ideas reduced variety below the baseline, because the ideas converged on the same way of being bold. What worked instead was assigning the model heterogeneous personas and constraints, which scored 0.757, the highest of any prompting intervention tested, followed by step-by-step reasoning and a higher temperature. Discipline around the synthetic data used to train a model should include a measure of output variety, not just average quality: a dataset generated by a single model inherits the same monoculture it produces.