To verify an AI answer, try recording your provisional judgment and the evidence you need before reading it. If the answer changes your mind, record what caused the change. This is an exercise proposed in this essay, not a tested training intervention or a report of an experiment we conducted.
A fluent answer can feel like completed work. Its orderly paragraphs may make the remaining review seem small. Yet understanding an explanation does not establish a reason to adopt it. The question here is where to keep your reasons while an AI system helps produce the answer. A short note made before the answer arrives gives you something to compare with the explanation you later accept.
The sources were checked on October 14, 2026. The starting point is the CHI 2025 paper by Hao-Ping Lee and colleagues. Microsoft Research's publication page describes a survey of 319 knowledge workers and 936 examples of using generative AI at work. These are respondents and reported examples, not two groups of experimental participants to add together.
What did the study establish?
The researchers investigated reported critical thinking and the confidence people brought to AI-assisted tasks. Their summary associates higher confidence in generative AI with less critical thinking, and higher confidence in one's own ability on a task with more critical thinking. Those relationships do not, on their own, establish that AI use caused a loss of thinking ability or that an individual's skills deteriorated over time.
Self-reported effort also differs from accuracy. Someone might spend less time while completing the necessary checks; someone else might think at length and still reach a wrong conclusion. I would not turn a reported reduction in effort into a measured loss of ability. Nor would I treat this survey as a result established for every Korean classroom or organization. Its findings are a starting point for asking how people examine AI output.
The researchers describe a shift toward information verification, integrating responses into work, and overseeing the task. Drafting less material can leave substantial work for the person deciding whether to use it. It helps to ask when that review begins. A process that begins only after the answer arrives may preserve the corrections but lose the original question and the evidence the reader intended to check.
A closed laptop, an empty notebook and a pencil on a wooden desk beside a windowView original
Leave a note before reading
Imagine writing a weekly report about an increase in customer inquiries. The cause is still unknown, but an AI draft attributes it to a discount campaign. After reading that account, you look up the campaign dates. Shipping delays and a change in how inquiries were classified might receive less attention. This is a hypothetical example about the order of review, not a customer incident we investigated.
Suppose your earlier note said that the classification rules needed checking. The AI explanation would then sit beside an unresolved question. To remove that question, you would need evidence that the rules were unchanged, or a note that you could not yet verify them. Your original judgment does not have to be correct. It needs to leave a recoverable account of where you started.
Keep the exercise small enough to use. One sentence can state a provisional conclusion; another can identify evidence that would change it. If the subject is unfamiliar, the conclusion can simply be unknown. Name what you would need to learn before deciding. An answer that fills a gap in your knowledge still leaves the separate job of finding evidence for its claims.
Do not put sensitive information into this note or into an external AI service without the required permission. Use approved material or a fictional case for practice. The exercise grants no consent to disclose customer documents and no authority to perform a business action. Existing organizational security rules continue to apply.
How should a change of mind be read?
Changing your mind after examining evidence is an ordinary result of review. The reason for changing deserves more attention than the fact of change. Did you discover an omitted condition, correct a calculation, or simply find the explanation more convincing? These reasons require different follow-up. Persuasive wording alone leaves the evidence check unfinished.
The following table is a proposed recording aid. It is not a scoring system validated by the study or an approval rule for every kind of work. Its purpose is to help a reader explain where a judgment changed. High-risk decisions must still follow the organization's separate review requirements.
Scroll horizontally to view a wide table.
| Point in the review | What to record | Next check |
|---|---|---|
| Before reading the AI answer | Provisional conclusion or an unknown | Evidence that could change it |
| After reading | Claim you intend to adopt | Source text and applicable conditions |
| Where the judgment changed | New evidence or a persuasive impression | Distinguish the evidence from the impression |
| What remains unresolved | Missing information and reason to defer | Who will check it and when |
Specify the kind of evidence when that helps. A condition in an official document, a value in an approved dataset, and a schedule confirmed by its owner need different checks. A working link is not verification of the claim beside it. Read whether the source contains that claim and whether an exception changes its application. Missing evidence should remain visible as missing evidence.
The same distinction applies when you ask the AI whether its first answer was correct. A follow-up can provide more explanation, but two matching answers do not automatically provide independent confirmation. Reading an original source or calculating separately is a different action from extending the conversation. Record what you actually checked so the gaps remain visible.
Two separated blank record cards and a slim document folder on a wooden worktableView original
How does this differ from DMS's task boundary map?
The DMS AI training task boundary map is a practical guide to deciding which work to delegate. This essay proposes recording the reader's reasons for accepting an answer. The scope of delegation and the grounds for adopting a conclusion require separate attention. Permission to draft a document may not include permission to finalize its conclusions.
If the one-page job sheet already sets acceptance criteria, those criteria can supply a starting point for the exercise. Decide how you will read the output before it arrives. Afterwards, separate the criteria it meets from the points you have not verified. This is a proposed extension of the judgment structure in the published DMS guides, not a claim that we have measured its effect in a class.
Keeping an initial judgment must not become a reason to defend it. Better evidence from another person, or an overlooked condition surfaced by AI, can justify a revision. The notebook preserves the reason for the revision. It does not certify the first answer. An explanation of why the original judgment was wrong can become useful material for the next task.
Must I solve the task without AI first?
You do not have to complete every task unaided. In a new field, identify what you do not know and a route for checking it. Requiring a beginner to produce an expert answer before using assistance would make this exercise burdensome. Revising a conclusion after reading outside material differs from knowing the correct answer at the start.
The length of the record should fit the stakes. A reason for preferring one headline can be brief. Interpreting a contract requires attention to the original wording and appropriate review. Important safety, medical, legal, and financial decisions cannot be settled by this exercise alone. A notebook entry is not a substitute for a responsible professional or an established organizational process.
For a short training exercise, I propose reading the starting assumption and the reason for revision alongside the final answer. Ask whether the learner properly deferred a condition they could not check. An unchanged answer deserves a reason too. The note cannot fully reveal whether the learner checked the evidence or merely liked the first explanation, so hearing the learner's account remains useful.
Where to go next
To connect the exercise with a team's work, establish the acceptance criteria and the points a person must verify. The English DMS guide to designing AI training explains the published task-boundary approach. The Korean edition also links to the Korean practical AI education portfolio; a separate English version of that portfolio page is not currently available. This essay does not promise a new course format or an improvement in training outcomes.
The note you can make today is modest: what remains unknown, and what evidence would change the conclusion? After reading the AI answer, place the evidence you actually checked beside that note. Any gap between the two can stay in the record until someone resolves it.
