David Dunning and Justin Kruger published their famous paper in 1999. The finding: people with limited competence in a domain systematically overestimate their ability. More competent people, having encountered the genuine complexity of a field, are more calibrated — and often slightly underestimate themselves.
Twenty-five years later, I watched this play out in real time in an enterprise AI workshop.
The Workshop Methodology
The setting was a structured AI proficiency workshop I facilitated with 3rd-semester IT students at Erhvervsakademi København (EK). Participants had varying levels of AI exposure — technically literate but not AI specialists, and representative of the kind of emerging professionals organizations are onboarding into AI-augmented roles.
Before the session began, I asked each participant to rate their AI competence on a 10-point scale.
The average was 5.6 — a comfortable "above average" self-assessment. Most participants felt they had a solid working grasp of AI tools.
Then we started the session.
The Valley of Despair
The first exercise involved a practical task: using an AI tool to complete a realistic work scenario. Not a demo. A real task with real ambiguity and real constraints.
Almost immediately, participants encountered failure modes they hadn't anticipated. Hallucinations. Confident-sounding incorrect outputs. Responses that looked right but missed the actual requirement. Prompts that produced very different outputs from what was intended.
Mid-session, I asked participants to re-rate their competence.
The average dropped to 4.9.
This is the Dunning-Kruger "valley of despair" in action. Increased exposure to a domain's real complexity decreases perceived competence, because you now understand what you don't know. The participants who had rated themselves 6 or 7 out of 10 now had specific, concrete evidence of gaps they hadn't known existed.
After structured practice, pattern recognition training, and framework-based prompt design, I asked once more.
End-of-session average: 7.0.
The final score isn't just higher than the start — it's more accurate. Participants had encountered real failure modes, developed heuristics for avoiding them, and could now apply AI tools in conditions closer to their actual work.
The Section AI Corroboration
My workshop findings align with what Section AI found in their broader research on AI proficiency across organizations.
The mechanism Section AI identifies is consistent with the Dunning-Kruger dynamic: organizations believe they are further along in AI maturity than they are, because their benchmark is self-assessment rather than output measurement.
Employees rate themselves as proficient. Managers assume proficiency. Adoption dashboards show usage. And yet — when you measure what has actually changed in output quality or process speed, the answers are vague. I observed exactly this pattern during my research at Utiligize — high confidence scores across the board, but no measurable productivity improvement.
What This Means for Practitioners
The implications for anyone responsible for AI adoption are significant.
Don't trust self-reported competence. The most confident employees in your organization may be your highest-risk users: people who have used AI enough to feel fluent, but not enough to recognize the outputs that require scrutiny.
Design training around failure modes, not features. Most AI training focuses on capabilities: here's what the tool can do. This produces the plateau at the top of the Dunning-Kruger curve — confident users with superficial understanding. Effective training should deliberately introduce failure cases, edge cases, and output validation exercises.
Calibration is a skill. Knowing when to trust an AI output, when to verify it, and when to reject it is a specific competency that requires deliberate practice. It should be explicitly trained and assessed.
The valley is a feature, not a bug. If your training program makes people feel less confident midway through, that is working as intended. The discomfort of recognizing real gaps is the mechanism through which genuine competence develops. Smooth, encouraging training that leaves participants feeling great tends to produce the starting-point 5.6, not the post-training 7.0.
The Organizational Risk
The deeper risk is not individual error — it is systemic overconfidence at the organizational level.
Section AI's research found that 54% of employees say they're confident using AI in their role. Organizations take that at face value and conclude they have a majority-competent AI workforce. When only 10% are actually operating at that level, the gap between perceived and actual readiness is not a minor measurement error.
It is a liability.
Organizations making significant decisions based on assumed AI capability — staffing models, workflow redesigns, competitive strategy — that is actually not present are making decisions on a foundation of confident ignorance.
The Dunning-Kruger effect is not a curiosity. In the context of enterprise AI adoption, it is a systematic source of organizational risk that can be measured, mitigated, and managed — if you're willing to treat competence as something that needs to be assessed, not assumed.
Workshop data from an IT student proficiency session at Erhvervsakademi København (2025). Organizational findings draw on a 6-month AI implementation study at Utiligize, Denmark (2025). Corroborated by Section AI's proficiency research (2024) and the academic Dunning-Kruger literature.