Let the stakes set the effort
If you are generating a brainstorm of possible titles, a wrong suggestion costs you a second of reading. If you are generating a medical dosage, a legal citation, or a factual claim that will be published under your name, a wrong output costs far more. The habit is to ask, before you use the output, what a wrong version would cost — and to spend checking effort in proportion. Low-stakes creative output can go straight to use. Anything factual, numeric, or attributed needs an independent source, not a second look at the same generated text.
What to check, by modality
- Text: verify names, dates, numbers, quotations, and citations against an outside source; treat confident specificity as a reason to check, not a reason to relax.
- Images: check countable and structural details — fingers, limbs, embedded text, reflections, and object boundaries — since these are where the model has no structural knowledge.
- Video: watch for flicker, morphing, drift, and warping, and check whether the subject stays the same entity across the whole clip rather than just looking right in one frame.
- Any modality: if the output will be attributed to you or used in a decision, the check has to come from outside the model, because the model cannot reliably flag its own errors.
Why asking the model to check itself is not enough
A model that produced a false claim has no separate access to the truth of that claim. Asking it to verify its own output mostly produces a second piece of fluent text, which can be just as wrong. Self-checking is useful for catching obvious inconsistencies, but it is not a substitute for an external source.