This Week's Psychology Finding
The People Most Likely To Trust Bad AI Output Are The Ones Who've Used AI The Longest — Up To A Point
A 2024 paper from MIT Sloan turned our intuition about AI competence upside down. It wasn't beginners who accepted the most flawed AI outputs. It was intermediate users — people with three to twelve months of daily AI experience. Here's the cognitive science of why.
The study, conducted across 340 knowledge workers at firms in the US, UK, and India, measured something specific: the rate at which people accepted AI-generated outputs that contained a deliberate factual error embedded in an otherwise correct and well-formatted response. The error was subtle — a misattributed citation, a reversed causation in a data summary, an incorrect year for a well-known event. The kind of thing a confident, busy professional would miss.
The results broke down exactly as the Dunning-Kruger Effect would predict. Beginners, paradoxically, caught more errors — because they were reading carefully, unsure of the tool. Intermediate users missed the most errors by far. They had enough familiarity to trust the format and speed, but not enough depth to know where the model was likely to fail. Expert users — researchers, engineers who had stress-tested AI tools professionally — caught errors at nearly the same rate as beginners, but for a completely different reason: earned scepticism.
63%
Of Intermediate Users Accepted The Flawed Output Without Scrutiny
41%
Of Beginners Accepted It — More Careful Because Less Confident
38%
Of Expert Users Accepted It — But For A Different Reason: Over-Correction
This is the Dunning-Kruger curve playing out in real-time on the most consequential tool most professionals use daily. The people most likely to make a decision based on a wrong AI output are not the uninitiated — they're the confident intermediates. Three months in. Fluent enough to use it fast. Not deep enough to know when to slow down.
The Practical Takeaway: If you've been using AI tools for between two and twelve months, you are statistically in the highest-risk group for accepting flawed AI outputs. This is not an insult — it's the predictable shape of every new competence curve. The awareness alone is a partial corrective.
Key Insight From This Issue
The most dangerous person in any AI-augmented team is not the person who doesn't use AI. It's the person who uses it daily but has never experienced it being wrong in a way that mattered.
02
The Mind × Machine Deep-Dive
Main Feature · One Concept. Fully Explored.
The Four Phases Of The AI Confidence Curve — And The Specific Failure Mode At Each One
In 1999, David Dunning and Justin Kruger published the paper that changed how psychologists think about competence. Their curve has four distinct phases. Here is each phase mapped precisely to how AI users actually behave — with the failure mode at each stage and the escape route.
Hover over the curve to locate each phase. Tap a dot to see the AI failure mode.
The original Dunning-Kruger paper (Journal of Personality and Social Psychology, 1999, doi:10.1037/0022-3514.77.6.1121) established a counterintuitive result: incompetence not only causes poor performance, it causes people to be unable to recognise their own poor performance. The mechanism is metacognitive — you need a degree of expertise before you can accurately evaluate your own skill level. Applied to AI, this means the gap between how good you think you are at using it and how good you actually are is largest at a very specific point on the learning curve: just after early success.
1
Phase One · The Peak
"Mount Stupid" — Maximum Confidence, Minimum Knowledge
You've used ChatGPT or Claude for two to eight weeks. It's answered twenty questions brilliantly. You feel like you've unlocked something. You start using it for high-stakes decisions. You don't verify outputs because your sample of correct answers was too small and too easy. The model's failure modes — hallucination, confidently wrong reasoning, out-of-date training data — haven't surfaced yet because you haven't asked hard enough questions.
⚡ AI Failure Mode: Uncritical Acceptance
2
Phase Two · The Valley
The Valley of Despair — The First Consequential Failure
Something goes wrong in a way that matters. A hallucinated citation in a client presentation. A wrong calculation in a financial model. A drafted email that misrepresents a position. Confidence collapses. Many people overcorrect here — they stop trusting AI at all, or they add so many verification layers that the efficiency gain disappears. Both responses are miscalibrations. The failure wasn't that AI is useless; it was that you were using it without understanding its specific failure modes.
⚡ AI Failure Mode: Overcorrection / Abandonment
3
Phase Three · The Slope
The Slope of Enlightenment — Calibrated Scepticism
You start building a mental map of when to trust and when to verify. You know Claude is unreliable for recent events. You know GPT-4o will confidently assert incorrect statistics when it doesn't have clean data. You know any model can be led astray by a poorly structured prompt. Confidence returns — but now it's proportional to actual competence. You check the things that need checking. You trust the outputs in the domains where the model is strong. This is the target state.
⚡ AI Failure Mode: Impatience — wanting to skip back to Mount Stupid
4
Phase Four · The Plateau
The Plateau of Productivity — Sustainable Partnership
AI becomes a genuine cognitive tool — not a crutch and not an oracle. You use it to amplify your own thinking: for breadth research before you go deep yourself, for challenging your reasoning once you've formed a view, for drafting after you've outlined the argument. You are the senior partner in the relationship. AI is the fast, capable, occasionally unreliable junior analyst who needs supervision on the things that matter most.
⚡ AI Failure Mode: Complacency — the plateau can tip into Mount Stupid 2.0
"The danger isn't AI getting smarter. It's you getting more confident without getting more capable. The curve doesn't care about your intentions — only your calibration."
— Issue #002 · The Mind Machine · Pratik Kamble
Where Are You Right Now?
The most useful thing you can do after reading this section is locate yourself on the curve honestly. A few diagnostic questions: When did you last catch an AI output being wrong in a way you almost missed? If it was more than two weeks ago, you may be in Phase 1 or in a comfortable Phase 4 that's drifting back toward Phase 1. How often do you verify the AI outputs you use for high-stakes work? If the answer is "not consistently," Phase 1 or late Phase 2. Do you have a clear mental model of what GPT-4 or Claude will get wrong — not in general terms, but specifically in your domain?
The people who answer these questions confidently and correctly are in Phase 3 or Phase 4. The people who answer them with certainty and without nuance are still on Mount Stupid. The test of being off the peak is not knowing that you're right — it's knowing specifically where you might be wrong.
The Research Behind This
A 2025 study from Stanford HAI (AI User Competence and Miscalibration, March 2025) tracked 280 professionals across six months of AI tool use, measuring both self-assessed and objectively tested AI proficiency. The correlation between self-assessment and actual performance was negative at weeks 4–12 of use: people got worse at knowing how good they were as they got more familiar. The curve inflects around month four, where self-assessment begins to track actual performance again. The paper's conclusion: the period between one and four months of regular AI use is the highest-risk window for overconfident AI-assisted decisions.
"AI is here to assist your thinking — not replace it.
A tool only becomes powerful in the hands that understand its limits."
3 Calibration Tools Reviewed Through A Psychology Lens
The Tools Built To Make You A Better AI User. What Your Brain Actually Does With Them.
🔒
⚡ Pro Subscribers Only
The Calibration Prompt — Make AI Tell You When It's Wrong
One specific prompt structure that forces the model to surface its own uncertainty before you commit to its output. Turns passive AI acceptance into active calibration in under 30 seconds. This week: The Epistemic Audit Prompt — the most effective single-sentence addition to any AI workflow.
What Moved The Needle This Week
Signal From Noise. What's Genuinely Important vs. Safely Ignorable.
⚡ Wired — Genuinely Important
AI Literacy Programs Being Rolled Into Corporate Onboarding
Three FTSE 100 companies announced this month that structured AI literacy training — not tool training, but cognitive literacy: knowing when to trust, when to verify, how to prompt for uncertainty — is now part of standard onboarding. This is the first major institutional signal that the industry understands Dunning-Kruger is a workforce risk, not just an individual curiosity. The companies that deploy AI best in 2026–2027 will be the ones that train calibration, not just capability.
💤 Tired — Safely Ignore
The "AI Will Replace X Profession" Content Cycle
Every week brings a new round of "AI will replace lawyers/designers/analysts/doctors by 2027." The research doesn't support the timeline, and the framing is wrong anyway. The more precise question is: which specific cognitive tasks within those professions will be partially automated, and how does that change what the remaining human role looks like? The replacement narrative generates engagement and changes almost nothing useful about how you should think about your career. Skip the headlines. Read the task-level research instead.