LLM models are stuck in formulaic thinking — a startup aims to fix it

Claude, ChatGPT, and Gemini demonstrate equally predictable responses to similar queries — on July 1, 2026, MIT Technology Review identified this phenomenon as a systemic "groupthink" of language models and reported on a startup working to overcome it.

The Number Test: Why This Is Not a Coincidence?

Ask any popular chatbot to name a random number between 1 and 10 — you'll almost certainly get 7. Ask again — you'll hear 3 or 4, then 8 or 9. The pattern reproduces with striking consistency across different models from different companies.

The explanation is straightforward: all major LLMs were trained on similar web corpora, where "7" as an answer to this question appears more frequently than other numbers — people themselves call seven "the most random" number. Reinforcement learning from human feedback (RLHF) further encourages "safe" and expected answers: those that more often receive high ratings from human raters. Models are literally trained to give a predictable response.

• Seven as a "random" number is a textbook example of LLM template thinking
• The pattern is characteristic of Claude (Anthropic), ChatGPT (OpenAI), and Gemini (Google DeepMind)
• The reason is overlapping training data and similar RLHF procedures across all major labs

Where Groupthink Causes Real Harm

Numbers are merely a visible symptom. In real tasks, the problem is larger: models reproduce the same cultural clichés, formulate strategic recommendations in similar ways, offer comparable marketing solutions. When a company uses multiple LLMs to "diversify perspectives," it often gets paraphrased versions of the same opinion — with the illusion of independence.

"We've created an ecosystem where all models see the world the same way — because they read the same thing," — MIT

Technology Review states.

The problem is especially acute where originality matters: scientific hypothesis generation, unconventional content, assessment of non-trivial risks. "Independent verification" through multiple LLMs in such cases creates an illusion of diversity — but not actual diversity.

What the Startup Proposes

MIT Technology Review describes a startup focused on methods to overcome "template thinking" in language models. The exact architecture of the solution is not disclosed. The industry, meanwhile, is discussing several approaches to this challenge:

• Training on more diverse data with deliberate inclusion of niche perspectives
• Managed stochasticity at the fine-tuning stage — encouraging variability as an explicit goal
• Ensemble systems where multiple models with different "biases" debate each other
• Diversity metrics for answers as a mandatory part of evals — alongside accuracy and safety

What This Means

If methods to overcome "group consensus" manage to set a new industry standard, it will change how we evaluate AI systems: diversity and independence of responses will become measurable requirements equal to accuracy or safety. For corporate users, this opens the possibility of getting genuinely different perspectives from AI, rather than a statistically averaged viewpoint in different formulations.