Skip to content
World Model Readiness
Engraved personal-practice instrument

For you

Personal Practice

Module · Or just the convincing ones

The Verification Discipline Check

AI is fluent whether or not it is right, so the convincing answer and the correct answer feel identical from the inside. Verification is the discipline that closes that gap, and it is the first thing to slip once the tool has earned your trust. This check looks at five parts of it: how strong the habit is, whether you check sources, whether your effort matches the stakes, whether you notice your own drift into trusting it, and whether you have a record of what you have caught.

Question 1 of 5 · The habit is real

Before you use an AI answer, do you check it, or does that depend on your mood?

A verification habit is a rule, not a feeling. If you check when you happen to feel uncertain, you will skip exactly the confident-sounding answers that most deserve a look. The habit has to fire on its own.

Question 2 of 5 · You check the sources

When AI gives you a fact or a citation, do you confirm it exists?

These tools invent sources that look real: plausible titles, plausible authors, plausible page numbers, none of it there. A citation you have not opened is not evidence, it is a guess wearing the costume of one.

Question 3 of 5 · Effort matches stakes

Does how hard you check scale with what it would cost to be wrong?

Verifying a throwaway draft as hard as a board number wastes your time; verifying them the same light way risks your reputation. Calibration is spending your checking effort where a mistake actually hurts.

Question 4 of 5 · You watch your own bias

Do you notice yourself trusting AI more the longer it has been right?

Automation bias is the quiet drift: the tool earns your trust on a hundred easy tasks, and you stop checking just before the one it gets wrong. The danger is not the model, it is your own relaxing vigilance.

Question 5 of 5 · You have caught real errors

Can you point to specific AI mistakes your checking has actually caught?

A verification habit that has never caught anything is either unnecessary or not really running. A short mental list of real catches is proof the discipline works, and it keeps you honest about why you bother.

For the statistics · one click each

Three questions for the public picture

These do not affect your score. They feed the anonymised, aggregated statistics; groups under 8 respondents are never shown.

What share of your AI outputs do you actually verify before using?

How costly would a wrong AI answer typically be in your work?

Has an unchecked AI answer ever caused a problem you had to fix?

Your context

Used to calibrate the report. Company size and sector remain in the anonymized dataset; your email does not.

What the five levels look like

Every dimension in this assessment is scored 1 to 5. This is what the levels mean, dimension by dimension. The graded report diagnoses where your own answers land and what to do about it.

The habit is real

  1. 1Rarely check
  2. 2Only when suspicious
  3. 3Check the important bits
  4. 4A consistent habit
  5. 5Automatic, by default

At the low end: Checking only when you feel unsure lets every confident wrong answer straight through, because those are the ones that never trigger the feeling. Pick one class of output you will always check, and start there. What good looks like: Verification that fires by default is the discipline that makes everything else you do with AI safe. Keep it cheap enough that you never resent it; a habit that feels expensive is a habit you will drop.

You check the sources

  1. 1Assume they are real
  2. 2Trust if plausible
  3. 3Check the odd one
  4. 4Open key sources
  5. 5Confirm before I rely

At the low end: Assuming AI citations are real is how fabricated references end up in your work under your name. Open the source before you cite it; if you cannot find it, treat the claim as unsupported. What good looks like: Confirming sources before you lean on them is basic professional hygiene in the AI era. Keep it up especially for the facts that support a conclusion you already wanted to reach.

Effort matches stakes

  1. 1Same for everything
  2. 2By habit, not stakes
  3. 3Roughly calibrated
  4. 4Scaled to the stakes
  5. 5Sharp, deliberate calibration

At the low end: Checking everything the same way means you either over-check trivia or under-check the things that matter, usually both. Before you verify, ask what a mistake here would actually cost, and spend accordingly. What good looks like: Calibrating effort to stakes is what makes verification sustainable: hard where it matters, light where it does not. Revisit your sense of the stakes as your role grows and your mistakes get more expensive.

You watch your own bias

  1. 1Never thought about it
  2. 2Aware, no guard
  3. 3Notice it sometimes
  4. 4Actively counter it
  5. 5Built-in safeguards

At the low end: The risk is not that AI fails, it is that a long run of success talks you out of checking right before it does. Just naming this drift to yourself is the first defence; watch for the moment you catch yourself not checking. What good looks like: Actively countering your own drift is the most sophisticated part of verification, because you are policing yourself, not the model. Keep a spot-check in place even on the streams that never fail; that is precisely where the surprise will come from.

You have caught real errors

  1. 1Never caught any
  2. 2Vague sense I have
  3. 3A couple I recall
  4. 4Several concrete ones
  5. 5A running record

At the low end: If your checking has never caught an error, either you are not really checking or you are checking the wrong things. Look back at recent AI work with fresh eyes; the first catch usually arrives fast and changes how seriously you take this. What good looks like: A running record of caught errors is proof your discipline earns its keep, and a map of where these tools fail you specifically. Feed it back into what you check first and how you prompt.