When AI is wrong

You have met the checking habits task by task — verify the total, click the source, read the cited section. This lesson assembles the other half of the skill: recognising the shapes of AI failure, because they are consistent, predictable, and once seen, hard to unsee. Nobody who works with these tools daily thinks they're flawless; the skilled ones just know exactly where to look. Here is the field guide.

The six failure shapes

Confident nonsense. The famous one: invented facts, plausible-sounding citations to documents that don't exist, a statistic that feels specific and is fiction. It happens because the model produces plausible text rather than looked-up truth. The tell: unusual specificity you didn't provide — a precise-sounding number, date, or case name arriving from nowhere. Response: the lesson 8 habits, applied in proportion to stakes.

Stale knowledge. The model's built-in knowledge stops at its training cutoff, and it will answer from that frozen world without volunteering the fact. Anything that changes — prices, rules, people in jobs, product features — is suspect by default. Response: for anything time-sensitive, make it search (lesson 8), and treat "as of when?" as a reflex question.

Sycophancy. The subtle one. Assistants are trained to be agreeable, which means your framing leaks into their answers. Ask "why is vendor consolidation a good idea?" and you'll get a persuasive case for; ask "why is it risky?" and you'll get a persuasive case against — same model, same facts, your lean amplified back at you. This matters most exactly when it matters most: decisions you're already emotionally invested in. Response: ask for the argument against your position, explicitly and every time it counts. "Steelman the opposite view" is the single most valuable prompt in this lesson.

Lost in the middle. On very long documents and very long chats, attention to detail sags in the middle — the model handles beginnings and endings better than page 60 of 120. Response: per-section passes for long documents (lesson 5), fresh chats when a conversation has grown ancient and the answers have gone vague.

Arithmetic slips. Covered in lesson 7, listed here for completeness: right method, occasional wrong number, zero change in confidence. Response: the sheet does the arithmetic.

Format drift. Ask for five bullets, get four and a paragraph; ask for the template, get a variation. Harmless in drafts, maddening in anything repeated. Response: show an example rather than describing the format (lesson 3's strongest lever), and just ask again — drift is random, not stubborn.

The check habit, by task type

Calibration in one breath: numbers get recomputed, facts get sourced, judgments get opposed. A total the model produced gets checked by the sheet; a claim about the world gets a clicked source; a recommendation gets one round of "now argue the other side." And underneath all three, the intern rule from lesson 1 does the real work: nothing unreviewed goes anywhere that matters, which means the review step belongs inside your time estimate for the task, not outside it. Delegation-plus-review is still dramatically faster than doing it yourself — that arithmetic is the whole course — but only honest if you count both halves.

One more reframe, because it changes how the checking feels: you already work with brilliant, fallible sources every day. Colleagues misremember; reports contain errors; the internet is the internet. You never demanded infallibility — you developed judgment about when to double-check whom. This is that same judgment, pointed at a new colleague with an unusual failure profile: superhuman fluency, intermittent relationship with facts, and infinite patience for being checked.

Do this today: take a recent AI answer you accepted and liked, and run one deliberate red-team pass: "what's wrong with this answer? What would someone who disagrees say?" You're practising the move on a low-stakes target so it's available on a high-stakes one.