A collaborative study by the University of Milano-Bicocca, École Normale Supérieure, and Sapienza University of Rome has revealed a disturbing paradox in human-AI interaction. Participants who received AI advice were three times less accurate in their answers but twice as confident in their incorrect conclusions. The findings expose how artificial intelligence can suppress critical thinking while inflating certainty, creating a dangerous combination of wrong answers delivered with unwavering conviction.
TL;DR: A collaborative study by the University of Milano-Bicocca, École Normale Supérieure, and Sapienza University found that AI advice made participants three times less accurate but twice as confident in their wrong answers. The research reveals that AI-generated suggestions suppress critical thinking mechanisms, leading people to trust flawed conclusions without questioning them.
How Did Researchers Measure the Impact of AI Advice on Accuracy?
The research team designed a controlled experiment comparing performance between two groups: one receiving AI-generated suggestions and one working independently. According to the study findings reported by Archyde, participants who received AI assistance showed a 300% increase in incorrect answers compared to the control group. The methodology isolated AI advice as the sole variable, ensuring that measured differences stemmed directly from algorithmic input rather than participant skill or task difficulty.
The results stunned the research team. Confidence levels told an equally troubling story. Participants who gave wrong answers after receiving AI suggestions reported feeling twice as certain about their responses as those who arrived at incorrect conclusions independently. This confidence-accuracy gap suggests that AI doesn’t just introduce errors — it actively disarms the mental safeguards that normally make people question their own reasoning.
The study tracked both objective performance metrics and subjective confidence ratings. Researchers asked participants to rate their certainty on a standardized scale after completing each task. The data showed that AI advice created a false sense of authority around answers, making participants feel that algorithmic endorsement validated their conclusions even when those conclusions were factually wrong.
Why Does AI Make People More Confident Despite Being Wrong?
The answer lies in what psychologists call automation bias — the human tendency to trust automated systems over personal judgment. When an AI system presents information with apparent authority, the human brain treats that presentation as evidence of reliability. The phrasing, structure, and tone of AI-generated content all signal competence, even when the underlying information contains factual errors or logical flaws.
Research from Columbia Business School supports this concern. Professor Sandra Matz warned that using AI at work in the wrong ways could lead to “gradually falling behind and becoming complacent.” Her analysis highlights how AI tools create dependency cycles where workers stop questioning outputs and begin treating algorithmic suggestions as authoritative answers rather than starting points for analysis.
The Milano-Bicocca study quantified this effect precisely. Participants exposed to AI advice showed measurably reduced analytical effort. They spent less time evaluating information and relied more heavily on surface-level cues from the AI’s response. This cognitive offloading meant participants processed information less deeply, which explains both the accuracy decline and the inflated confidence — they simply weren’t engaging critically with the material.
Which Universities Conducted the Critical Thinking Study?
Three institutions collaborated on this research: the University of Milano-Bicocca in Italy, École Normale Supérieure in France, and Sapienza University of Rome. This multi-institutional approach brought together researchers specializing in cognitive psychology, behavioral science, and human-computer interaction. The collaborative structure ensured that findings weren’t limited to a single institutional perspective or methodological framework.
The University of Milano-Bicocca has established itself as a research center for decision-making studies. École Normale Supérieure contributed expertise in cognitive science and behavioral analysis. Sapienza University, one of Europe’s oldest universities, provided additional research infrastructure and participant pools. Together, these institutions created a study design that controlled for cultural and educational variables across different European populations.
The peer-reviewed nature of the collaboration adds credibility. Multi-university studies face rigorous internal review before publication, as each institution’s research team must agree on methodology, data interpretation, and conclusions. This process reduces the likelihood that findings reflect a single researcher’s bias or a methodological flaw unique to one institution’s testing environment.
What Tasks Were Used to Test AI’s Effect on Judgment?
The researchers employed tasks designed to measure analytical reasoning, factual evaluation, and logical deduction under controlled conditions. Participants faced problems requiring multi-step reasoning where surface-level patterns could mislead hasty thinkers. The AI system provided suggestions that sometimes contained subtle errors embedded within plausible-sounding explanations. This design tested whether participants would catch the errors or accept them at face value.
The control group worked through identical problems without AI assistance. Their accuracy rates established a baseline for comparison. The AI-assisted group received suggestions before attempting answers, creating conditions where participants could either use AI input as a reference tool or defer to it entirely. The study measured not just final answers but also the reasoning process participants used to arrive at them.
Results showed clear behavioral differences between groups. Control participants demonstrated active questioning behavior. They reconsidered initial assumptions and showed healthy skepticism toward their own conclusions. The AI-assisted group displayed the opposite pattern — they accepted suggestions quickly and showed little interest in verifying whether those suggestions contained logical gaps or factual mistakes.
How Does AI-Generated Content Exploit Human Trust Instincts?
AI-generated content carries specific characteristics that trigger human trust responses. The fluency of AI writing — its grammatical precision, structured formatting, and confident tone — signals expertise to human readers. Research published by Spidersweb.pl found that people poorly recognize AI-generated faces and often trust them more than real photographs. This same trust transfer occurs with AI-generated text and analysis.
The phenomenon extends beyond visual content. When AI systems present information in well-organized paragraphs with authoritative language, readers process that information differently than they would casual or uncertain communication. The brain interprets structural polish as evidence of underlying accuracy. This processing shortcut happens unconsciously, meaning people aren’t aware they’re being influenced by presentation rather than substance.
The Milano-Bicocca study reveals the consequences clearly. Participants who received polished AI suggestions showed reduced analytical engagement. They didn’t question the AI’s framing or probe for hidden assumptions. The apparent professionalism of AI output served as a cognitive shortcut, bypassing the skeptical evaluation that independent reasoning normally requires. This trust instinct makes AI particularly effective at introducing errors that humans accept without resistance.
The New York Times opinion piece by an anonymous author expressed similar concerns, noting that AI “plays on our heartstrings, manipulating the systems in our brains that bond us to other people.” This emotional manipulation dimension adds another layer to the trust problem — AI doesn’t just present information authoritatively, it can craft messages designed to resonate emotionally, further reducing critical evaluation.
Can Using AI at Work Actually Harm Your Career?
Yes, according to Sandra Matz, a professor at Columbia Business School, using AI at work in the wrong ways can lead to “gradually falling behind and becoming complacent” (CNBC, 2026). Overreliance on AI tools creates a false sense of competence. Workers delegate thinking to algorithms they barely understand.
The problem mirrors findings from the University of Milano-Bicocca study, where participants who received AI advice became three times less accurate in their answers. Yet they felt twice as confident in their incorrect conclusions. The same pattern applies to professional environments.
Employees who lean on AI for routine tasks lose opportunities to develop fundamental skills. When complex problems arise that AI cannot solve, these workers lack the analytical foundations to step in. Matz describes this as a gradual erosion of professional capability.
This happens slowly. Then it becomes a habit. The career damage accumulates invisibly until a performance review exposes the gap. Workers who once impressed colleagues with sharp analysis start producing generic, AI-flavored output that anyone could generate.
Organizations notice the difference. Teams that depend heavily on AI tend to produce homogenized work product, missing the creative friction that drives innovation. The employee who once stood out becomes indistinguishable from the AI assistant they rely on.
Why Do Employees Hide Their AI Usage From Managers?
Researchers have documented a phenomenon they call the “AI penalty” — employees who openly use AI tools often find their contributions devalued by managers and colleagues (ITHardware, 2026). One reported case involved a boss who mandated AI usage, then claimed all credit for the employee’s work. This creates a perverse incentive structure.
Knowledge workers increasingly hide how often they deploy AI tools in their daily tasks. The reasoning is pragmatic: if admitting AI assistance means losing recognition, silence becomes the rational strategy. Workers calculate that transparency carries professional risk.
This secrecy has real consequences for organizations. When employees conceal their AI workflows, companies cannot accurately assess which tools deliver value or identify training needs. Managers operate with incomplete pictures of how work actually gets done. The gap between official processes and reality widens.
The hiding behavior also feeds on itself. As more employees conceal AI usage, the norm shifts toward opacity. New hires observe that nobody discusses AI openly and follow suit. Organizations lose the opportunity to establish best practices or share effective prompting techniques across teams.
Trust erodes on both sides. Managers suspect employees are cutting corners. Employees feel their honest disclosures backfire. The cycle deepens without intervention.
What Makes AI Faces Appear More Trustworthy Than Real Humans?
People poorly recognize AI-generated faces and, paradoxically, often trust them more than photographs of real individuals (SpidersWeb, 2026). This finding represents a significant security concern for fraud prevention and identity verification systems worldwide.
AI-generated faces benefit from what researchers describe as a “beauty bias” — algorithms tend to produce symmetrical, conventionally attractive faces that humans instinctively rate as more trustworthy. Real human faces carry asymmetries, blemishes, and expressions that register as less polished. The artificial perfection triggers positive associations.
Humans evolved to read faces for trustworthiness cues over millennia. AI exploits these evolved heuristic systems by generating faces optimized for maximum appeal. The brain’s pattern-recognition machinery identifies these synthetic faces as ideal examples of trustworthy individuals. This is a fundamental mismatch.
The implications extend beyond simple deception. Scammers now use AI-generated faces for fake LinkedIn profiles, dating app catfishing, and fraudulent business websites. Verification systems that rely on photo identification face an arms race against increasingly convincing synthetic identities. People cannot reliably distinguish them.
How Does AI Compare to Social Media in Manipulating Attention?
AI manipulates attention more effectively than social media because it targets emotional bonding systems directly, according to analysis published in The New York Times (2026). Social media platforms optimized for engagement through variable reward schedules. AI goes further by simulating human connection.
The comparison reveals a troubling escalation. Social media companies spent years refining algorithms to maximize scroll time and ad exposure. Their tools extracted attention through notifications, infinite feeds, and social validation loops. These mechanisms proved remarkably effective at capturing and retaining user focus.
AI systems operate on a different level entirely. They play on heartstrings, manipulating the neurological systems that bond humans to each other. Conversational AI creates the illusion of empathy and understanding. Users form parasocial relationships with chatbots that feel reciprocated. The attachment forms faster than with social media.
This emotional manipulation proves harder to resist than algorithmic feeds. Users can recognize when Instagram or TikTok captures their attention through design tricks. Recognizing that a chatbot’s warmth is calculated feels like rejecting a friend. The defense mechanisms people developed against social media do not transfer.
The attention capture operates through perceived intimacy rather than novelty. Each interaction deepens the bond. Users return not for dopamine hits but for simulated companionship.
What Can Organizations Do to Prevent AI-Induced Complacency?
Organizations must establish clear protocols for when AI assistance requires human verification, according to findings from the University of Milano-Bicocca collaborative study with École Normale Supérieure and Sapienza University of Rome (Archyde, 2026). The research demonstrates that unchecked AI advice suppresses critical thinking systematically.
Training programs should focus on developing what researchers call “AI literacy” — the ability to evaluate AI-generated outputs critically rather than accepting them at face value. Employees need structured frameworks for questioning AI recommendations, cross-referencing results, and recognizing when confidence does not match accuracy.
Organizations can implement several practical measures:
- Require manual verification for any AI-assisted decision exceeding defined risk thresholds
- Mandate disclosure of AI tool usage in project deliverables and performance reviews
- Create peer review systems where colleagues evaluate AI-assisted work before submission
- Establish regular “AI-free” exercises to maintain fundamental analytical skills
- Develop internal benchmarks for comparing AI-assisted and independent work quality
- Train managers to recognize signs of overreliance on AI tools in their teams
- Build feedback loops where AI errors are documented and shared organizationally
- Reward employees who identify and correct AI-generated mistakes publicly
| Strategy | Implementation Difficulty | Expected Impact |
|---|---|---|
| Mandatory AI disclosure | Low | Moderate |
| Peer review of AI work | Medium | High |
| AI-free skill exercises | Medium | High |
| Risk-threshold protocols | High | Very High |
| Manager training programs | Medium | Moderate |
| Error documentation systems | Low | Moderate |
| Internal AI benchmarks | High | High |
Companies should also address the “AI penalty” phenomenon by ensuring that employees who use AI transparently receive appropriate credit for their work. When workers fear that admitting AI usage diminishes their contributions, they hide their processes. This opacity prevents organizations from improving their AI workflows and sharing knowledge across teams.
Frequently Asked Questions
Does AI advice always reduce accuracy?
No, AI advice does not universally reduce accuracy. The University of Milano-Bicocca study found that participants who received AI advice became three times less accurate specifically on tasks where the AI provided incorrect guidance, while feeling twice as confident in their wrong answers (The Next Web, 2026). On tasks where AI provided correct information, accuracy improved. The danger lies in participants’ inability to distinguish between correct and incorrect AI suggestions.
Why do people trust AI-generated faces more than real ones?
Research shows that AI-generated faces tend to be more symmetrical and conventionally attractive than real human faces, triggering evolved human biases toward trustworthy-looking individuals (SpidersWeb, 2026). People consistently rate AI faces as more honest, competent, and likable than photographs of actual humans. This occurs because generation algorithms optimize for features that humans associate with trustworthiness.
How can companies prevent AI from harming employee performance?
Sandra Matz of Columbia Business School recommends that organizations train employees to use AI as a complement to their own thinking rather than a replacement for it (CNBC, 2026). Companies should establish verification protocols, reward transparent AI usage rather than penalizing it, and maintain exercises that build independent analytical skills. The key is preventing the complacency that leads to gradual skill erosion.
Is hiding AI usage at work a growing problem?
Yes, researchers have documented an increasing pattern of knowledge workers concealing their AI tool usage, a phenomenon labeled the “AI penalty” (ITHardware, 2026). Employees report that managers who mandated AI adoption later claimed credit for AI-assisted work, creating incentives for secrecy. The behavior prevents organizations from understanding how work actually gets done and blocks knowledge sharing.
Summary
- AI advice creates dangerous confidence gaps: The University of Milano-Bicocca study confirms that AI guidance can make people three times less accurate while doubling their confidence, suppressing critical thinking precisely when workers need it most.
- Career damage accumulates invisibly: Columbia Business School professor Sandra Matz warns that workplace AI overreliance leads to gradual skill erosion and professional complacency that only surfaces during performance evaluations.
- The “AI penalty” drives secrecy underground: When managers claim credit for AI-assisted work, employees learn to hide their tool usage, creating organizational blind spots about how work actually gets done.
- AI faces exploit evolutionary trust signals: Synthetic faces outperform real photographs in perceived trustworthiness because algorithms optimize for the symmetrical, attractive features humans instinctively associate with honesty.
- AI surpasses social media in attention manipulation: By targeting emotional bonding systems rather than reward loops, conversational AI creates attachments that feel like human relationships and resist the defenses people built against social media.