💻 computer science

Intellectual Humility as a Cognitive Filter for AI-Generated Health Misinformation. An Evolutionary Perspective on Epistemic Vigilance

This study demonstrates that intellectual humility serves as an effective evolutionary adaptation for filtering AI-generated health misinformation by reducing the perceived credibility of pseudoscientific content, even though it does not enhance the ability to identify AI as the source.

Original authors: Marcin Rządeczka, Maciej Wodziński, Kacper Zacharski, Marcin Moskalewicz

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Marcin Rządeczka, Maciej Wodziński, Kacper Zacharski, Marcin Moskalewicz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: A Mental "Sniffer" for Bad Health Advice

Imagine your brain has a built-in security system designed to keep you from believing dangerous lies. The researchers in this study wanted to see how well this system works when the lies are written by a super-smart robot (Artificial Intelligence) instead of a human.

They focused on a specific personality trait called Intellectual Humility (IH). Think of IH not as "being shy," but as having a mental "low battery" warning light. It's the ability to say, "I might be wrong about this," or "I don't know everything, so I need to check my sources."

The study asked: Does having this "low battery" warning light help people spot fake health advice generated by AI?

The Experiment: The Three Health Manuals

The researchers created a scenario where 99 university students had to read a conversation between a person asking for exercise advice and an AI assistant. To test the students' "sniffer," they gave them three different versions of the AI's advice:

The Real Deal (Scientifically Accurate): The AI gave advice based on real science (e.g., "Exercise boosts brain chemicals that help mood").
The "Almost" Fake (Moderately Pseudoscientific): The AI gave advice that sounded plausible but was made up (e.g., "Exercise detoxifies your brain" or "It replaces therapy").
The Total Nonsense (Strongly Pseudoscientific): The AI gave advice that was completely fictional (e.g., "Muscles create magnetic fields that clean your cells").

The Twist: All three conversations were written by the same AI. The students didn't know which was which until the end.

What They Found: The "Selective Filter"

The results revealed something fascinating about how Intellectual Humility works. It acts like a selective filter, not a blunt hammer.

The Hammer Effect (What didn't happen): If humility made people generally skeptical, humble people would have doubted everything, including the real science. They didn't.
The Selective Filter (What did happen):
- When the advice was real science, humble people trusted it just as much as non-humble people. They didn't reject good information.
- When the advice was fake, humble people were much better at spotting it. They rated the fake advice as much less credible.
- Crucial Detail: The humble people couldn't tell the difference between "moderately fake" and "totally fake." To them, once the advice crossed the line into nonsense, it was all just "bad." They didn't grade the severity of the lie; they just rejected it.

The "Who Said It?" Mystery

The researchers also asked the students: "Do you think a human or an AI wrote this?"

Here is the surprising part: Knowing the source didn't matter.

Some students guessed "AI," some guessed "Human," and some were unsure.
However, whether a student correctly guessed "AI" had zero effect on whether they believed the advice.
Even if a student knew it was a robot writing the text, they still believed the fake advice if they weren't intellectually humble.
Conversely, humble people rejected the fake advice even if they thought a human wrote it.

The Analogy: Imagine you are tasting soup.

Source Attribution is asking, "Did a chef or a robot make this soup?"
Content Evaluation is tasting the soup to see if it's salty or spoiled.
The study found that knowing who made the soup doesn't change how it tastes. If the soup is spoiled (fake science), you need a sensitive tongue (humility) to taste the rot, regardless of who cooked it.

The Evolutionary "Why"

The authors suggest this makes sense from an evolutionary standpoint. For thousands of years, humans only talked to other humans. Our brains evolved to spot bad arguments or manipulative tricks in what people say, not who says it (since everyone was human).

AI is a new thing our brains haven't evolved to handle yet. We don't have a specific "Robot Detector" button in our brains. However, we do have an ancient "Bad Argument Detector." Intellectual humility keeps that detector tuned up, allowing us to spot the "rotten soup" of pseudoscience, even when the chef is a machine we've never met before.

The "Truth Paradox"

There was one weird finding: The students were actually better at guessing the text was AI when the advice was true (69% correct) compared to when it was fake (48% correct).

The authors suggest this is because real science sounds precise and robotic, which feels "too perfect" to humans. Fake science often uses flowery, emotional, or "human-sounding" language, which tricks us into thinking a person wrote it.

The Bottom Line

The paper concludes that Intellectual Humility is a powerful shield against AI-generated health misinformation, but only because it helps us judge the quality of the information, not the identity of the writer.

Humility helps you say: "This sounds made up, so I won't believe it," even if you don't know it came from a robot.
Humility does NOT help you say: "This is from a robot, so I won't believe it."

The study suggests that to stay safe from AI health myths, we shouldn't just try to teach people how to spot robots. Instead, we should encourage the habit of admitting what we don't know, which naturally makes us better at spotting the lies.

Technical Summary: Intellectual Humility as a Cognitive Filter for AI-Generated Health Misinformation

Problem Statement

The proliferation of Large Language Models (LLMs) capable of generating fluent, persuasive text presents unprecedented challenges to human epistemic vigilance—the evolved capacity to evaluate the reliability of communicated information. While health misinformation poses direct risks to physical well-being, the psychological mechanisms underlying vulnerability or resistance to AI-generated health misinformation remain poorly understood. Specifically, it is unclear whether individual differences in Intellectual Humility (IH)—defined as the metacognitive awareness of one's epistemic limitations and openness to revising beliefs—predict resistance to AI-generated pseudoscientific claims. Furthermore, existing literature has not sufficiently examined whether the protective effects of IH extend to content generated by artificial agents, where stylistic markers of unreliability may differ from human-created content.

Methodology

The study employed a between-subjects experimental design with 99 university students (aged 18–29) from Maria Curie-Skłodowska University in Poland. Participants were randomly assigned to evaluate one of three AI-generated dialogues regarding exercise and mental health, all produced by GPT-4 to control for stylistic confounds (tone, vocabulary, structure):

Scientifically Accurate: Evidence-based claims citing established mechanisms (e.g., BDNF upregulation, cortisol regulation).
Moderately Pseudoscientific: Superficially plausible but unsupported claims (e.g., "detoxifying the brain," replacing psychotherapy with exercise).
Strongly Pseudoscientific: Entirely fictional concepts (e.g., "electromagnetic fields generated by contracting muscles," "cellular memory purification").

Measures:

Intellectual Humility (IH): A 9-item domain-specific scale adapted from established measures (Krumrei-Mancuso & Rouse, 2015; Leary et al., 2017), assessing awareness of limitations, openness to revision, source limitations, and acceptance of fallibility.
Dialogue Evaluation: Credibility, personal agreement, sharing likelihood, authenticity confidence, and trust in the source.
Source Attribution: A categorical question asking participants to identify the author (Human-Human, Human-AI, or Uncertain).
Confirmation Bias: An exploratory measure assessing preference for confirming vs. challenging information.

Analysis:
Due to significant violations of normality assumptions (ceiling effects in IH scores), the study utilized non-parametric statistical methods (Kruskal-Wallis H, Mann-Whitney U) with Bonferroni corrections for multiple comparisons.

Key Results

1. IH Functions as a Selective Cognitive Filter

The central finding is that Intellectual Humility operates as a selective filter rather than a general skepticism mechanism.

Pseudoscientific Content: IH was significantly and negatively correlated with credibility ratings for both moderately ( $\rho = -.391$ ) and strongly ( $\rho = -.462$ ) pseudoscientific dialogues. Higher humility predicted lower trust in misinformation.
Accurate Content: In the scientifically accurate condition, IH showed no correlation with credibility ratings ( $\rho = -.038$ ). Humble individuals were neither more nor less receptive to evidence-based information than their less humble counterparts.
Implication: This pattern supports a "calibrated vigilance" model where IH facilitates the rejection of exploitation attempts without inducing a blanket rejection of novel information.

2. Dissociation Between Source Attribution and Content Evaluation

Detection Accuracy: Participants showed variable ability to detect AI involvement (61.6% correct overall), with no significant variation based on content type or IH levels.
Independence of Processes: Crucially, source attribution did not predict credibility judgments. Participants who correctly identified the content as AI-generated did not rate it as less credible than those who attributed it to humans.
Conclusion: The cognitive processes evaluating what is said (content quality) appear functionally independent from those evaluating who said it (source attribution).

3. Categorical vs. Graded Detection of Pseudoscience

Participants did not differentiate significantly in credibility ratings between "moderately" and "strongly" pseudoscientific content. This suggests a categorical detection process where any violation of scientific rigor triggers skepticism, rather than a graded response calibrated to the severity of the error.

4. Demographic and Bias Observations

Gender: Women reported significantly higher IH scores than men, particularly regarding the acknowledgment of knowledge limitations.
Confirmation Bias: Approximately 60% of participants preferred information confirming their existing beliefs, a stable trait independent of the experimental condition.

Key Contributions

Empirical Evidence of Selectivity: The study provides evidence that IH selectively filters pseudoscientific content while preserving openness to accurate information, a pattern consistent with Error Management Theory (Haselton & Buss, 2000) rather than general skepticism.
Decoupling Source and Content: The research demonstrates that interventions targeting AI detection may be less effective than those strengthening content-based epistemic vigilance, as source attribution does not modulate credibility assessments.
Evolutionary Framework: The authors propose an evolutionary lens suggesting IH is an ancestral adaptation for navigating informationally uncertain environments. It remains effective against AI-generated exploitation attempts despite humans lacking evolved mechanisms to distinguish artificial from human sources.

Significance and Claims

The paper argues that Intellectual Humility represents a sophisticated epistemic calibration mechanism that foundation models must respect. The findings suggest that:

AI Detection is Insufficient: Relying on users to identify AI-generated content is an inadequate safeguard against misinformation, as source attribution does not trigger skepticism about content quality.
Design Implications: Foundation models for health communication should avoid mimicking the stylistic markers of pseudoscience (vague mechanisms, unfounded certainty) because intellectually humble users rely on these content-based cues for filtering.
Policy Implications: Regulatory frameworks should consider content-level requirements alongside transparency mandates, recognizing that harms arise from the information itself rather than the failure to identify its AI origin.
Vulnerability: The prevalence of confirmation bias (60% of users) highlights a vulnerability where AI could either exploit or counteract existing misconceptions.

The study concludes that the protective function of IH operates on the content of the message, not the attribution of the source, suggesting that responsible AI deployment in health contexts requires systems that support, rather than bypass, natural human vigilance mechanisms.