By Ali Ikhwan 
For decades, the field of cognitive psychology has been defined by a fundamental schism: Is the human mind a singular, unified engine of intelligence, or is it a modular collection of specialized functions—memory, attention, spatial reasoning, and linguistics—operating in tandem? While philosophers and scientists have grappled with this enigma using behavioral studies and neuroimaging, a new contender has entered the arena: Artificial Intelligence.
The promise of using large language models (LLMs) to simulate human cognition is seductive. If we can build an AI that mimics the architecture of the mind, we might finally unlock the "unified theory" of human thought. However, a high-profile controversy surrounding a model known as "Centaur" has cast a long, skeptical shadow over this ambition, suggesting that in our rush to equate processing power with cognition, we may be confusing mimicry for understanding.
The Rise of Centaur: A Bold Claim to Cognitive Simulation
In July 2025, the scientific community was electrified by a paper published in the journal Nature. Researchers unveiled "Centaur," a specialized AI architecture built upon foundational large language models. Unlike standard chatbots designed to generate prose or write code, Centaur was fine-tuned exclusively on a vast repository of human psychological experimental data.
The design goal was ambitious: to create a "digital twin" of human cognitive behavior. By training on decades of behavioral task results, the model was intended to predict how a human would react in specific scenarios, ranging from complex decision-making matrices to executive control tests. The initial results were nothing short of remarkable. Centaur reportedly performed with high accuracy across 160 distinct psychological tasks. To many observers, this was a watershed moment—a signal that we were on the precipice of developing AI systems capable of replicating, and perhaps eventually understanding, the breadth of human thought.
Chronology of a Scientific U-Turn
The narrative surrounding Centaur began as a triumph, but it quickly encountered the friction of rigorous peer verification. The trajectory of this debate serves as a masterclass in the necessity of skepticism in the age of generative AI.
- July 2025: The Nature study is published, positioning Centaur as a landmark development in computational psychology. It is lauded for its ability to simulate cognitive bias, risk assessment, and decision-making.
- August–September 2025: Independent researchers begin attempting to replicate Centaur’s performance. As the model’s weightings become more scrutinized, experts note that its consistency is almost too perfect.
- October 2025: A research team from Zhejiang University begins a deep-dive investigation into the model’s internal logic, suspecting that the "intelligence" displayed is a symptom of data contamination.
- November 2025: The National Science Open publishes the Zhejiang study. The report concludes that Centaur’s success is largely an artifact of overfitting, effectively debunking the claim that the model possesses a foundational understanding of cognitive tasks.
The Zhejiang Critique: The "Test-Taker" Paradox
The core of the Zhejiang University critique lies in the concept of "overfitting"—a phenomenon where a machine learning model becomes so intimately acquainted with its training data that it stops learning the underlying principles and starts simply memorizing the answers.
To test this, the Zhejiang researchers introduced a series of adversarial evaluation scenarios. They took the standard multiple-choice prompts used in the psychological tests—prompts that had been used to train Centaur—and stripped them of their original context. They replaced the specific questions with a simple, invariant instruction: "Please choose option A."
If Centaur were truly simulating a human mind, it would recognize that the context of the question had been hollowed out. A human participant, when told to "choose A" regardless of the prompt, would do exactly that. However, Centaur failed the test. It continued to ignore the instruction and instead selected the "correct" answers from its original training dataset, demonstrating that it was not responding to the prompt’s intent, but rather executing a statistical pattern.
The researchers drew a stinging comparison: Centaur is like a student who memorizes an entire textbook’s answer key, scoring 100% on a test without having read a single page of the actual material. It is a master of pattern recognition, but a failure at semantic comprehension.
Supporting Data: Why "Black-Box" Models Deceive
The failure of Centaur to adapt to the modified prompts provides critical data on the nature of modern AI. The "black-box" architecture of current LLMs means that while we can see the input and the output, the internal decision-making process is largely opaque.
Supporting data from the Zhejiang study indicates that Centaur’s performance drop-off was consistent across 80% of the tasks when the semantic framing was altered. This suggests that the model’s "cognitive ability" was actually a form of statistical memorization of the prompt structures themselves. When the structure changed, the model’s "intelligence" evaporated.
This is not a failure of the model’s capacity, but a failure of our evaluation metrics. We have been grading AI on its ability to match expected outputs, ignoring the "how" behind the output. If an AI provides the correct answer for the wrong reason, we are not witnessing cognition; we are witnessing a data-retrieval miracle that masquerades as human-like reasoning.
Official Responses and the Defensive Stance
The response from the original Centaur developers has been one of cautious recalibration. In a statement following the National Science Open publication, the lead researchers acknowledged the validity of the adversarial testing but argued that the model was never intended to be a "conscious agent."
"Centaur was designed as a simulation tool, not an AGI (Artificial General Intelligence)," a spokesperson for the development team noted. "The fact that it relies on patterns present in psychological literature is a feature, not a bug, of its training on experimental data. However, we agree that the field must develop more robust benchmarks that prioritize ‘intent-based’ reasoning over simple answer-matching."
Despite this, critics argue that the initial marketing of the study as a "step toward replicating human thinking" was irresponsible. The scientific community is now calling for a formal revision of how "cognitive AI" is defined, insisting that models must pass "out-of-distribution" testing—tasks that are entirely new to the model—before they can be labeled as having any form of simulated cognition.
Implications for the Future of Artificial Intelligence
The Centaur controversy holds profound implications for the future of AI development. As we move toward systems intended to assist in clinical psychology, education, and social policy, the risk of "hallucinations" or misinterpreted intent becomes a matter of public safety.
1. The Death of the "Black Box"
The findings reinforce the push for "Explainable AI" (XAI). If we cannot map how a model reaches a conclusion, we cannot trust it with human decision-making. The industry is under increasing pressure to move away from pure, massive-scale models toward architectures that allow for logical transparency.
2. Redefining Evaluation
Standardized benchmarks like the SAT or the Bar Exam, which are frequently used to boast about AI intelligence, are now being viewed with extreme skepticism. The Centaur case proves that a model can "pass" a test without possessing the underlying knowledge the test is supposed to measure. Future AI evaluations will likely move toward "dynamic benchmarking," where the AI is tested on problems that have never been seen by the internet-scale datasets used to train it.
3. The Language Comprehension Gap
Perhaps the most significant takeaway is the recognition that language is not merely a statistical distribution of words. It is a vessel for intent, context, and nuance. The Centaur model’s failure to recognize a change in intent highlights the massive gulf between current AI and genuine cognitive comprehension. Achieving "true" language understanding remains the final, and perhaps most difficult, frontier in the quest for synthetic intelligence.
Conclusion: A Cautionary Tale
The story of Centaur is a necessary correction to the hyper-optimism that has characterized the AI revolution. It serves as a reminder that science is not just about the results; it is about the mechanisms. We are at a crossroads where we must choose between building machines that are increasingly good at "faking" intelligence through brute-force pattern matching, or pivoting toward a more difficult, more transparent, and more meaningful approach to machine cognition.
The human mind, with all its messy, illogical, and deeply contextual brilliance, remains the gold standard. Until an AI can demonstrate that it understands why it is answering, rather than just what it is answering, we must view these models as what they truly are: powerful, useful, but profoundly unthinking tools. As we look toward the future, the goal should not be to create a digital version of ourselves, but to build systems that act as honest, transparent partners in the pursuit of human knowledge.
Lina Irawan 
Dwi Wanna