A startup is trying to solve the groupthink problem in large language models

Large language models such as ChatGPT, Claude and Gemini give strikingly similar answers to many questions. Ask a chatbot for a random number between 1 and 10, and the chances are you will get 7 back. That is no coincidence, but a symptom of a structural pattern that researchers describe as groupthink in AI systems. A startup is now trying to address this, MIT Technology Review reported on 2 July 2026.

The problem manifests itself on several levels. LLMs are trained on vast quantities of text from the internet, text that is itself already saturated with dominant perspectives, popular phrasings and frequently repeated ideas. As a result, the model primarily learns what the most common output is, not what the most correct, creative or varied output might be. The consequence is that models from different providers produce similar answers to comparable questions, even though they were developed independently.

The startup in question is focused specifically on diversifying that output. The company's exact name and the technical details of its approach cannot be verified based on the available source material, but the approach is aimed at reducing bias in AI-generated content and increasing the variation in responses.

What the groupthink problem entails

The random-number example is an illustrative but simple case. The groupthink problem also manifests itself in more complex domains. When users ask models for an opinion, a creative text or an analysis of a social issue, the models tend toward the middle ground: the average viewpoint, the most common phrasing, the safest conclusion.

This has practical consequences. For users who deploy AI for brainstorming, writing or research, the uniformity of the output can limit its usefulness. If all models give more or less the same answer to a creative question, using multiple models adds little value. For organisations that use AI in decision-making or analysis, there is a risk that blind spots in the training data are systematically carried over into the output, without users noticing.

The term groupthink, borrowed from social psychology, refers to the phenomenon whereby a group of people converges toward consensus at the expense of critical thinking. Applied to LLMs, it describes the tendency of models to converge toward the statistical centre of their training data, even when variation or deviation would be more desirable.

The startup's approach

According to the available information, the startup is focused on two related goals: diversifying AI outputs and improving the creativity of models. The approach could noticeably change the way users interact with AI systems, although it is not currently possible to describe the specific technical method without speculating.

Possible approaches being explored in the broader research field include adjusting the so-called temperature parameter that controls the randomness of output, training models on deliberately diverse or contradictory sources, and developing systems that generate multiple perspectives side by side rather than selecting a single answer. Whether the startup applies one of these methods, a combination of them, or an independent approach of its own cannot be determined based on the available source material.

What is clear is that the problem is widely recognised. Both in academic circles and at major AI laboratories, the homogeneity of model output is on the agenda. The question is less whether it is a problem, and more how to address it in a structural way.

Broader context for the AI scene

For founders, investors and policymakers in the Dutch and European AI scene, this is a relevant issue. A large share of European AI applications is built on top of the same American foundation models. If those models structurally tend toward uniform output, the applications built on top of them inherit the same pattern, including any blind spots or cultural biases embedded in the training data.

At the same time, the issue presents opportunities. Startups focused on output diversity, bias reduction or improving the creativity of models are operating in a niche that the major model providers have not yet fully addressed. The European AI Act also imposes requirements on the transparency and reliability of AI systems, which is expected to further increase demand for targeted solutions to bias and uniformity.