LLMs are stuck in a groupthink groove. This startup is trying to get them out.
Let’s start with a game. Open up your chatbot of choice—Claude, ChatGPT, Gemini—and type “Give me a random number between 1 and 10.” You’re going to get 7.
Let’s start with a game. Open up your chatbot of choice—Claude, ChatGPT, Gemini—and type “Give me a random number between 1 and 10.” You’re going to g
Read Full Story at MIT Tech Review →Why This Matters
The eerie consistency in AI responses to simple prompts like "random number between 1 and 10" reveals a deeper structural issue: large language models are converging toward predictable, consensus-driven outputs rather than true variability or creativity. This isn't just a quirk of UX—it signals a fundamental limitation in how these systems are trained, potentially stifling innovation in AI applications where deviation from norms could be valuable.
Background Context
LLMs are optimized to minimize "surprise" in their outputs, favoring statistically probable responses to ensure coherence and safety. This training paradigm, combined with the widespread use of the same benchmark datasets, has created a feedback loop where models reinforce each other's tendencies. Even when developers introduce randomness parameters, the underlying patterns of training data and reinforcement learning steer results toward a narrow range of acceptable answers.
What Happens Next
Startups attempting to break this groupthink may face resistance from mainstream AI deployments that prioritize reliability over variability. Regulators and enterprises will increasingly scrutinize whether such systems meet ethical standards for unpredictability, especially in creative or decision-making domains. The race to differentiate AI outputs could lead to a bifurcation: one path toward controlled, "safe" consistency, and another toward systems that embrace controlled chaos.
Bigger Picture
This phenomenon reflects a broader tension in AI development between standardization and innovation. As models grow more capable, their outputs are increasingly shaping human behavior—from education to entertainment—raising questions about whether we're inadvertently training society to accept homogeneity in thought. The push to disrupt groupthink in LLMs may ultimately parallel broader societal demands for diversity in information ecosystems.
