Can AI really simulate human thinking? Research casts doubt on an influential study, suggesting an advanced model was just really good at memorizing patterns.
Publish Date: 2026-05-22 08:00:00
Source Domain: www.livescience.com
- Recent research challenges the claims of a 2025 study that an AI model called Centaur could accurately simulate human thought with up to 64% accuracy based on over 10 million human decisions.
- The recent study argues that Centaur likely learned statistical shortcuts (overfitting) rather than understanding human behavior, as it excelled on training data but struggled with new data.
- Centaur’s high performance was attributed to it identifying learned patterns in its training data, suggesting it may not fully understand the tasks it performed well on.
- The new study’s findings raise questions about the ability of large language models (LLMs) to reach artificial general intelligence (AGI), emphasizing fundamental limitations in current neural-network-based AI.
- Researchers stress the importance of stress-testing AI models to correctly assess their capabilities and avoid premature conclusions about their understanding of human cognition.
- While the 2025 study showed Centaur could accurately predict behavior in scenarios outside its training data, the authors of the recent study did not address this finding directly.
- Critics highlight that current AI models might excel at pattern-matching, which could give a false impression of understanding without true cognitive grasp of the task.
