Artificial intelligence (AI) tools like large language models (LLMs) can act as a double-edged sword in professional settings. They can be powerful “force multipliers” that enhance expert productivity but also risk amplifying the Dunning–Kruger effect among less experienced users.
When novices lean on AI-generated outputs without sufficient knowledge to validate them, the combination of AI’s fluent, confident responses and users’ limited expertise can foster an illusion of competence and inflated self-confidence.
In extreme cases, people can produce authoritative-sounding reports or decisions using AI and believe they’ve acquired expertise, even when the content is flawed. [academic.oup.com]
Dunning–Kruger in the Age of AI – What’s Happening?
The term Dunning–Kruger effect originates from social psychology, where people with limited ability commonly lack the metacognitive skill to recognize their own knowledge gaps, and thus vastly overrate their competence. It’s a nearly universal bias across fields, from driving to science literacy, where low performers believe they’re above average. This occurs because “individuals ignorant of their own ignorance” often don’t know enough to detect errors in their understanding. Meanwhile, true experts have the opposite tendency. They grasp the complexity and nuance of a subject, so they assess themselves more cautiously, sometimes underestimating their relative skill. [researchgate.net] [academic.oup.com]
Generative AI such as ChatGPT add a new twist to this old bias. AI systems are extremely fluent and persuasive in their outputs. They produce grammatically perfect, confident explanations on almost any topic. For an unskilled or novice user, this can mask the limits of their own understanding.
When a person uses an LLM to create a technical document or solve a complex problem without possessing the requisite domain knowledge, the result may appear authoritative and detailed. However, because the novice lacks the background knowledge, they may not notice errors or omissions – so they mistakenly believe “the AI got it all right, so I must have done a good job”. LLMs can give novices a false sense of expertise. The coworker that drafts a BIM Execution Plan entirely with an LLM, only to later discover it was “garbage” due to missing technical elements, exemplifies this phenomenon. They did not know what a good BEP looked like, so the AI’s smooth prose felt convincing, reinforcing the user’s overconfidence. [thedecisionlab.com]
Research shows how LLM outputs often sound correct even when they’re wrong. One commentator vividly noted that today’s AI “doesn’t just give you wrong answers. It gives you confident, well-reasoned, articulate wrong answers that sound exactly like what an expert would say”.
This authoritative tone can obscure the boundaries of a novice’s actual skill, making it hard to tell how much of the answer comes from genuine understanding.
Key Studies on AI, Overconfidence, and Metacognition
Several studies and analyses have examined how AI influences user confidence and competency across different contexts. The table below summarises key research findings that illuminate the AI-augmented Dunning–Kruger phenomenon:
| Study (Author & Year) | Context / Domain | Main Finding | Relevance to AI & Overconfidence |
| Kruger & Dunning (1999) [mirror – bruceabernethy.com] | Psychology (general) | In multiple tests, low performers dramatically overestimated their competence and could not recognise their errors, while top performers slightly underestimated their abilities. | Original study by Justin Kruger and David Dunning that identified the cognitive bias now known as the Dunning–Kruger effect. Shows that the least skilled individuals are the least able to recognise their own lack of skill, because the same knowledge required to perform well is required to evaluate performance. |
| Fisher et al. (2015) [medicalxpress.com] | Cognitive psychology | People who searched the Internet for answers felt more knowledgeable, even about unrelated topics, than peers who didn’t search. Simply being in “online search mode” created an illusion of internal knowledge, even when searches turned up no answers. | Illustrates cognitive offloading and the illusion of knowledge (including the illusion of explanatory depth) that arises when people rely on search engines. Serves as a clear precursor to modern AI‑driven overconfidence, where access to fluent answers is mistaken for understanding. |
| Fernandes & Welsch et al. (2025) [neurosciencenews.com], [academic.oup.com] | Human–Computer Interaction (metacognition in logical reasoning tasks with ChatGPT) | Across two large experiments (N ≈ 700) using ChatGPT on LSAT-style problems, all participants (novices & advanced users alike) significantly overestimated their performance when using the AI assistant. Actual scores improved slightly (by ~3 points on average), but self-rated scores overshot by ~4 points. Those with higher AI knowledge showed worse calibration, i.e. the most “AI literate” participants were the most overconfident. The classic Dunning–Kruger pattern disappeared, replaced by a general overconfidence effect induced by AI use. | Empirical evidence that generative AI inflates user confidence and weakens self-awareness of knowledge limits. Highlights how AI can “level” inexperienced and experienced users – making everyone feel above average and obscuring who truly has expertise. Shows that **improvements in task performance with AI do not translate to better self-knowledge; rather, **AI use tended to amplify overconfidence and reduce metacognitive accuracy. |
| He et al. (2023) [dl.acm.org] (CHI Conf.) | Human–AI Decision-Making (Human reliance on AI systems) | Studied how users with Dunning–Kruger tendencies rely on AI for decision support. Found that participants who initially overestimated their own performance (the “unskilled-and-unaware” group) tended to under-rely on accurate AI advice, preferring their own flawed judgment. A tutorial intervention (showing users their mistakes and AI’s fallibility) partially improved self-awareness and appropriate reliance. | Reveals another facet of AI + Dunning-Krueger Effect. Overconfident novices may ignore or misuse AI guidance because they erroneously trust themselves more. This suggests AI can’t automatically overcome DKE – cognitive biases affect whether users listen to or trust the AI. Proper user training and feedback are needed to calibrate trust and improve human–AI team performance. |
Mitigating AI-Amplified Overconfidence
The AI + Dunning–Kruger effect is now a documented reality. Generative AI can make novices feel like instant experts, while even experienced professionals aren’t immune to its confidence‑boosting influence. The risk isn’t that AI will replace experts, it’s that it makes non‑experts sound indistinguishable from them, and organisations start acting on how things sound rather than how well they’re understood.
The response isn’t to ban AI or shame people for using it. It’s to strengthen the habits that keep confidence calibrated. That means slowing down where uncertainty is high, encouraging verification over first answers, and making it normal to ask who is accountable for correctness, not just delivery. It also means designing AI use so people have to explain and justify outputs, rather than simply forwarding them.
AI works best as a competent assistant, not an all‑knowing oracle. It can accelerate drafting and pattern recognition, but it cannot confer understanding. As one researcher put it, AI can quickly turn novices into amateurs, but it cannot let you leapfrog past experts. [neurosciencenews.com] [harishgautam.net]
The implication is that as fluency gets cheaper, judgement becomes the scarce resource. Organisations that don’t actively protect domain expertise and metacognitive humility won’t just make worse decisions. They’ll lose the ability to tell why their decisions are getting worse.
And once that signal is gone, no amount of confident language will bring it back.
No Comments