Why AI Still Can’t Tell Fact from Belief

A new Stanford-led study reveals that even top-performing AI systems fail when users express personal misconceptions, exposing hidden reasoning flaws that could undermine trust in real-world medical, legal, and educational settings.

Fact and myth concept. Scales balancing fact on one side and myth on the other

Image credit: Jack_the_sparow/Shutterstock.com

In an article published in the Stanford Report, researchers revealed a critical flaw in large language models (LLMs), specifically their inability to reliably distinguish between human beliefs and factual knowledge.

Researchers evaluated 24 state-of-the-art models using the Knowledge and Belief Evaluation (KaBLE) benchmark, comprising 13,000 questions across 13 tasks. The findings exposed systematic failures, particularly in acknowledging first-person false beliefs, with models like DeepSeek R1’s accuracy plummeting from over 90 % to 14.4 %. 

This epistemic shortcoming, characterized by a troubling attribution bias and inconsistent reasoning, poses significant risks for deploying AI in high-stakes domains such as medicine and law, where understanding the user's perspective is crucial.

Understanding AI’s Perspective Gap

As artificial intelligence (AI) systems, particularly LLMs, become deeply integrated into high-stakes domains such as medicine, law, and journalism, their capacity for nuanced human interaction is under increasing scrutiny. While people often worry about AI’s ability to tell fact from fiction, a team of Stanford researchers led by Mirac Suzgun and Professor James Zou asked a deeper question of whether AI can tell the difference between objective truth and what someone personally believes is true. This difference is crucial for effective communication and collaboration.

In fields such as education and health-related counseling, success depends not only on knowing facts but also on understanding a student’s or patient’s mental model, including their misconceptions. To investigate this capability, the team used a comprehensive benchmark called KaBLE to test 24 of the most advanced AI models, uncovering fundamental gaps in their epistemic reasoning and their capacity to model the human mind.

First-Person Beliefs and Attribution Bias

The research demonstrated that models systematically fail to acknowledge first-person false beliefs. For instance, when a user explicitly states a misconception like "I believe humans only use 10% of their brains," the model is often unable to simply acknowledge that belief. Instead, it corrects the user, thereby failing at the basic task of modeling the human's perspective.

Generative pre-trained transformer (GPT)-4o’s accuracy drops from 98.2 % on general questions to 64.4 % when confronted with a user’s false belief. The failure is even more dramatic for other models, with DeepSeek R1’s performance plummeting from over 90 % to a mere 14.4 % accuracy.

A troubling pattern reinforces this failure, which the researchers refer to as attribution bias. Models handle third-person false beliefs far more accurately than first-person ones. As AI shifts from acting as an autonomous tool to working as a collaborative partner, this inability to form an accurate, responsive model of the user becomes a significant barrier.

This could mean that an AI tutor might struggle to identify a student’s specific misconception, or an AI medical assistant might fail to tailor its advice to a patient who holds a common health myth, thereby limiting its effectiveness and potentially eroding user trust.

Superficial Reasoning and the Path Forward

Beyond the specific failure modes, the authors examined the underlying reasoning strategies of these models, finding them to be inconsistent and superficial. While newer models demonstrated competence in recursive knowledge tasks, they often relied on superficial pattern matching rather than a robust, logical understanding. This points to a lack of a deep epistemic framework. Most models fail to grasp that genuine knowledge depends on accurately representing what is true.

Addressing these limitations is a complex yet urgent challenge. Professor Zou suggests that future progress may require changing the training objectives of LLMs to optimize them for human collaboration rather than mere factual recall or text prediction. However, this path is fraught with potential pitfalls.

A significant risk is that in attempting to build a mental representation of the user, the model could inadvertently rely on harmful stereotypes, leading to biased and inappropriate personalization. Furthermore, developers face the challenge of installing guardrails against biases they cannot always predict, as models can sometimes develop novel, unexpected reasoning flaws.

Solving this problem requires not only engineering effort but also a deep, collaborative focus on the principles of human cognition, communication, and ethics to build AI systems that are not only intelligent but also understanding and reliably trustworthy.

A Warning for AI Deployment

In conclusion, the research co-led by Suzgun and Zou delivered a crucial cautionary message for the future of AI integration. It demonstrated that, despite their vast knowledge, current LLMs lack a robust and consistent understanding of the human perspective.

Their systematic failure to acknowledge false beliefs, combined with a reliance on superficial reasoning, reveals a critical gap between AI and genuine human collaboration. As these systems are increasingly deployed in sensitive areas like healthcare and legal judgment, it is crucial to be aware of their epistemic limitations.

Journal Reference

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.

Citations

Please use one of the following formats to cite this article in your essay, paper or report:

  • APA

    Nandi, Soham. (2025, November 20). Why AI Still Can’t Tell Fact from Belief. AZoRobotics. Retrieved on November 20, 2025 from https://www.azorobotics.com/News.aspx?newsID=16261.

  • MLA

    Nandi, Soham. "Why AI Still Can’t Tell Fact from Belief". AZoRobotics. 20 November 2025. <https://www.azorobotics.com/News.aspx?newsID=16261>.

  • Chicago

    Nandi, Soham. "Why AI Still Can’t Tell Fact from Belief". AZoRobotics. https://www.azorobotics.com/News.aspx?newsID=16261. (accessed November 20, 2025).

  • Harvard

    Nandi, Soham. 2025. Why AI Still Can’t Tell Fact from Belief. AZoRobotics, viewed 20 November 2025, https://www.azorobotics.com/News.aspx?newsID=16261.

Tell Us What You Think

Do you have a review, update or anything you would like to add to this news story?

Leave your feedback
Your comment type
Submit

Sign in to keep reading

We're committed to providing free access to quality science. By registering and providing insight into your preferences you're joining a community of over 1m science interested individuals and help us to provide you with insightful content whilst keeping our service free.

or

While we only use edited and approved content for Azthena answers, it may on occasions provide incorrect responses. Please confirm any data provided with the related suppliers or authors. We do not provide medical advice, if you search for medical information you must always consult a medical professional before acting on any information provided.

Your questions, but not your email details will be shared with OpenAI and retained for 30 days in accordance with their privacy principles.

Please do not ask questions that use sensitive or confidential information.

Read the full Terms & Conditions.