MIT Tech Review: LLMs Stuck in 'Groupthink' — A Startup Aims to Break the Pattern

MIT Technology Review published an in-depth feature on July 1 exposing a surprising limitation of today's large language models: their overwhelming tendency toward predictable, homogenous responses. Staff writer Michelle Kim demonstrates the phenomenon with a simple test — asking ChatGPT, Claude, and Gemini for a random number between 1 and 10 almost always returns 7, followed by 3 or 4 on the next attempt, then 8 or 9.

While this predictability may go unnoticed in everyday chat, the article argues it becomes a genuine limitation for creative or exploratory use cases where users expect fresh perspectives. The underlying issue stems from how training data distributions, optimization objectives, and model architectures collectively push toward the most statistically likely answer at the expense of diversity.

The report notes that for closed-ended tasks like research and code generation, this uniformity is rarely a problem. But for brainstorming, content strategy, or ideation — where serendipity and unexpected connections add real value — the groupthink effect undermines one of the key promises of AI assistance: the ability to think beyond human patterns.

According to the article, a startup is now working on breaking this uniformity. The company believes the core problem lies in training methodologies that overemphasize the single most probable response. By adjusting training strategies or introducing architectural innovations, it aims to preserve answer quality while significantly expanding the range of acceptable outputs.

Although the startup's identity and technical specifics are not disclosed in the report, its mission signals a growing industry recognition that model 'quality' should encompass not just accuracy but also diversity and creative range. The piece positions this work as part of a broader conversation about what AI creativity means and how to measure it.

The implication for the broader AI industry is significant: as knowledge workers increasingly rely on LLMs for strategy, content creation, and brainstorming, model output homogeneity could subtly constrain human thinking itself. If every AI assistant offers broadly similar answers, the diversity of human innovation may suffer in ways that are hard to detect but material in effect.

Looking ahead, the key signals to watch include whether the startup's approach will be open-sourced, and whether major model providers like OpenAI, Anthropic, and Google will introduce dedicated diversity enhancements to their offerings. In an age of ubiquitous AI, ensuring model diversity may prove as critical as ensuring model accuracy.