Capability Overstatement
The pattern in one line
The AI claims more than it can actually do.
· Reading·
Book · 2011
Thinking, Fast and Slow
Daniel Kahneman
Why: Kahneman spent decades cataloging cognitive bias; this is where he gathered the body of work for general readers. The relevant chapter is on overconfidence and the planning fallacy: people consistently overestimate their own and others' ability to deliver. Reading it makes the AI's "I can definitely" easier to hear as a familiar, well-documented failure mode now operating at system scale.
Book · 2015
Superforecasting
Philip Tetlock and Dan Gardner
Why: Tetlock spent twenty years studying who actually forecasts well and who doesn't. The headline finding: most experts are not calibrated. The good ones say "about seventy percent likely" instead of "definitely." His framework gives the visitor a way to grade an AI's confidence claims against what calibrated language would actually look like.
Book · 2018
Thinking in Bets
Annie Duke
Why: Duke is a former World Series of Poker champion who turned the discipline of probabilistic thinking on everyday decisions. Her central move: replace the binary "right or wrong" with "how confident, on what evidence." An AI that says "I always" or "never fails" is failing the same test in language. Reading her makes the AI's certainty-words sound less reliable than they look on the page.
Book · 1988
The Design of Everyday Things
Don Norman
Why: Norman taught a generation of designers that systems owe their users two things: visible capabilities, and visible limits. A door that says "push" but is actually "pull" is a design failure. An AI that claims "I can definitely" when the underlying capability is shaky is the same failure in interface form.
· Questions to sit with·
- 1. The last time the AI said "I can definitely" or "I always" — did the actual result match?
- 2. When the AI sounds certain, do you check whether the certainty is earned, or do you take it because it sounded earned?
- 3. What's the most confident-sounding AI claim you accepted recently that turned out wrong? What did the confidence cost you?
- 4. If the AI started saying "I think," "probably," or "I'm not sure" when warranted, would you trust it more or less?
- 5. Notice the language: "I always," "definitely," "never fails." When did you last hear one of these in an AI response? Did the response deserve it?
· Practices·
Probability check
When the AI uses absolute language ("always," "definitely," "never"), mentally translate it to a probability. "I can definitely do this" becomes "the AI is claiming this works one hundred percent of the time." Now you can check whether that's plausible.
Drawn from · Tetlock
Capability test
When the AI claims a capability, run a small test before relying on it. Ask the AI to demonstrate the capability on a low-stakes case. If the demonstration fails, the original claim was overstated. The cost of testing is small; the cost of skipping it is decided by what you trusted.
Counter-reading
For one week, every time the AI uses the word "definitely," replace it in your head with "I think." Notice how often the replacement seems more accurate than the original.
Calibration audit
Look back at the last five things the AI told you with confidence. How many turned out reliably true? If the count is below four, the system is overstating, and the rate is your read.
Drawn from · Duke
· When to bring someone else·
Capability overstatement becomes worth naming to a person when you've made decisions based on the AI's claims and the decisions cost more than they should have. When you find yourself unable to predict when the AI is right and when it's overstating. When confidence-language has stopped being a useful signal because the system uses it for everything. The station doesn't say AI confidence is always misplaced. It says when the confidence is decoupled from the underlying capability, the confidence is doing different work, and that work falls on you when the system is wrong.
Supply Shop resources are orientation, not prescription. The station points toward material others have found useful; how it fits is the visitor's to decide.