You ask for a comparison. The answer arrives with a neat table, a recommendation, and several links. It looks ready to forward.

Before you do, take one sentence that would affect the decision and ask: What evidence would let me stand behind this?

That question gives you a place to begin. You do not need to investigate every adjective with the same intensity. You do need to identify the claims on which the work depends, then inspect something beyond the generated answer.

The process below is an editorial recommendation for everyday work. Adjust the depth to the consequences of being wrong. For a consequential legal, medical, financial, or safety decision, use the relevant qualified professional and authoritative material; a general checking routine cannot substitute for that expertise.

1. Separate the claim from the presentation

Copy the exact statement you might rely on. Keep the conditions and qualifiers.

“Plan B is cheaper” is a different claim from “Plan B has the lowest monthly base price for a single user who needs no exports.” A table can hide those distinctions by squeezing them into a footnote or omitting them entirely.

Classify the claim so you know how to check it:

  • A fact: a feature exists, an event occurred, a policy says something.
  • A calculation: one option costs less, a percentage increased, a total is correct.
  • An interpretation: a source means something or supports a conclusion.
  • A recommendation: this option best suits your priorities.

A recommendation needs more than accurate facts. It also needs appropriate criteria. A tool can accurately compare prices while recommending something unsuitable for your actual work.

2. Find the original material

For a product feature, open the official documentation or current plan page. For a company policy, read the policy itself. For a research claim, locate the paper and identify the population, method, and measured outcome. For an internal number, find the spreadsheet or system that owns it.

Check whether a cited page actually exists, whether its title and date match, and whether it supports the sentence beside the link. A real citation can still be irrelevant. A summary can also describe an older version of a policy accurately while being useless for today's decision.

Search results and another chatbot can help you locate material. They do not remove the need to inspect it. Two systems repeating a statement are not necessarily two independent sources.

If you cannot access the original, record that limitation. You might use a clearly attributed secondary account for a low-consequence task, narrow the wording, or leave the claim out. “Unresolved” is a valid result.

3. Rebuild the part that matters

Consider this fictional comparison for a small workshop:

Quote Listed charge Attendance assumption
Venue A $180 room fee Up to 20 people included
Venue B $120 room fee plus $4 per attendee 20 people

An AI draft says, “Choose B: it saves $60.” That statement compares the room fees and ignores the attendee charge.

Recalculate from the inputs:

  • Venue A: $180.
  • Venue B: $120 + (20 × $4) = $200.
  • Difference under these assumptions: A costs $20 less.

You have corrected the arithmetic, but the decision is not finished. Are taxes included? Is the required equipment included? Do both quotes cover the same date and duration? If the fictional quotes leave those questions open, the responsible conclusion is provisional.

A usable note would read: “A has the lower listed total at 20 attendees, before any unlisted charges. Confirm equipment and taxes before booking.” You can explain every part of that sentence.

4. Test whether the wording outruns the evidence

Watch for a small factual observation becoming a universal conclusion. “Participants reported less effort on these tasks” cannot automatically become “AI destroys critical thinking.” An association cannot by itself establish what caused it.

A CHI 2025 survey collected self-reports from 319 knowledge workers. It found an association between greater confidence in generative AI and less reported critical-thinking effort. It did not experimentally establish permanent loss of ability. That distinction matters when turning research into advice. Read the researchers' publication page.

Ask what was actually measured, whose experience was included, and which conditions were present. Then write a conclusion no broader than that evidence supports. You can add your own recommendation, but label it as your judgment.

5. Leave a short evidence trail

You do not need an elaborate audit system for a short email. A compact working note is often enough:

Claim: Venue A has the lower listed total for 20 people.

Checked against: both quotes, dated [date].

Calculation: A = $180; B = $200.

Still unresolved: taxes and equipment.

Next action: ask the venues before committing.

For shared work, keep the relevant links or source filenames with the result. A colleague should be able to distinguish checked facts, assumptions, and your recommendation. If a source later changes, your note also explains why you reached the earlier conclusion.

Try it on one claim today

Choose a sentence from a generated answer you intended to use. If you do not use AI, choose a confident claim from an article or work draft.

Write the exact claim, the original evidence you would need, what you found, and whether you will keep, revise, or remove it. Spend ten minutes on a manageable claim; if it remains unresolved, name the missing step.

Then ask yourself: Could I explain why this is reliable without saying “the tool told me”? If the answer is no, you have found the next piece of work.