Researchers have identified and documented a systematic bias in large language models toward convergent outputs—when asked for "random" responses, models like Claude, ChatGPT, and Gemini show statistical clustering toward the same numbers (e.g., 7 appears far more frequently than it should). This behavior, called LLM groupthink, emerges from the way models are trained and reflects underlying issues with diversity in generated outputs.

A startup is building tools to measure and reduce this bias. The approach involves both identifying where groupthink occurs in model outputs and developing interventions to increase diversity in generation. Early findings suggest the bias affects not just random number generation but extends to creative tasks, problem-solving, and recommendations where variety of perspective would be valuable.

What This Means for Your Business

If your business depends on AI systems for ideation, brainstorming, or scenario planning, be aware that models show systematic bias toward convergent outputs. This means asking ChatGPT for five alternative strategies will likely produce more similar suggestions than would come from diverse human teams. To mitigate this, implement processes that explicitly prompt for contrasting viewpoints, run multiple independent queries with different prompting approaches, or combine AI outputs with human judgment from teams with different backgrounds. As these tools mature, evaluate vendors who address diversity and groupthink as an explicit design goal.