At the 2026 Inclusion·Bund Conference, Princeton University's Wang Mengdi highlighted that current large language models, while possessing vast knowledge, are limited in their ability to make novel scientific discoveries. These models tend to favor the most probable outcomes, neglecting the long-tail possibilities where true breakthroughs often lie. Wang's team found that LLMs like ChatGPT and Claude often overestimate common scenarios and underestimate minority ones in simulated life experiments. She argued that AI's path to independent scientific discovery hinges not on larger models, but on creating more verifiable real-world experimental infrastructure. AI
IMPACT Current AI models are better suited for tasks with clear validation, limiting their capacity for groundbreaking scientific discovery without improved experimental infrastructure.
RANK_REASON Expert opinion on AI capabilities and limitations.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
