Why checking is part of the job
AI assistants write fluent, confident text whether or not the content is right. Fluency is not evidence. The tool is generating likely wording, not looking up verified facts, so it can produce a wrong date, a misquoted figure or a source that does not exist, all in the same calm tone as correct material. The reasons behind these fabricated answers are covered elsewhere on this site. What matters here is a routine that catches them.
You do not have to check everything with the same intensity. A brainstorming list needs little verification. A paragraph with figures going to a client needs a lot.
Step one: decide how much checking the output deserves
Ask two questions:
- What happens if this is wrong? A typo in an internal note is minor. A wrong number in a budget, a wrong claim about a regulation or a wrong statement in a customer message can cause real harm.
- Who will rely on it? Text only you will read needs less care than text going to many people.
Match your effort to the answers. For health, legal, financial or safety topics, treat AI output as a starting point only and confirm it with qualified sources or professionals.
Step two: separate the claims
Read the output and mark each statement that could be true or false: names, dates, quantities, quotations, cause-and-effect statements and anything described as a rule or requirement. Opinions and phrasing can be judged on their merits; factual claims need support.
A useful trick is to ask the assistant to list the factual claims it made, in a numbered list. This turns a smooth paragraph into a checklist.
Step three: verify against sources you trust
For each claim that matters, find it in a source that is independent of the AI answer:
- the original document, if you supplied one;
- official or primary sources such as laws, regulators, company filings or standards bodies;
- reputable reference works or published research;
- internal records for company-specific facts.
If you cannot find support, do not use the claim. “I could not confirm it” is a legitimate outcome, and it is better than passing on something unverified.
Handle numbers and calculations with extra care
Language models are not calculators. They can set out arithmetic that looks right and is not. Recompute any totals, percentages or conversions yourself with a calculator or spreadsheet. Check units and time periods too: monthly versus yearly mix-ups are common. If the figures came from a document you provided, compare them with the original, since a summary may have changed a number or dropped a qualifier.
Be skeptical of citations
Requests for sources are risky. An assistant may produce a title, author and journal that look real and are not, or attach a real author to a paper they never wrote. If you see a citation:
- Search for the exact title in a library catalog or search engine.
- Confirm the authors, date and publication.
- Open the source and make sure it actually says what the answer claims.
A source that exists but does not support the claim is still a failed check. If you cannot locate it, remove it.
Also be cautious with quotations. Treat any quote attributed to a named person as unverified until you find it in the original.
Look for what is missing
Errors are not only false statements. An answer can be accurate but one-sided, outdated or silent about important exceptions. Ask yourself what a knowledgeable colleague would add. You can also prompt the assistant to list caveats, assumptions or situations where its advice would not apply, then check those yourself.
Check for staleness too. Many tools do not know about recent events or changes in rules, and may not tell you so. For anything time-sensitive, confirm the current position from a current source.
Ask the tool to help you check, not to certify itself
You can ask an assistant to flag its own uncertainty or to explain its reasoning, which sometimes reveals gaps. But a tool agreeing with its previous answer is not verification. Asking “Are you sure?” can produce either a repeated mistake or an unnecessary reversal. Use the answer as a prompt for your own checks.
Build a record
For work that matters, keep brief notes: what the AI produced, what you checked, which sources you used and what you changed. If questions arise later, that trail helps. It also supports workplace expectations around human review and clear accountability, since the person who sends the work remains responsible for it.
A quick checklist
- Have I rated the stakes and chosen how deeply to check?
- Have I listed the factual claims?
- Is each important claim supported by an independent source?
- Have I recomputed the numbers?
- Do the cited sources exist and say what is claimed?
- Is anything outdated or missing?
- Am I comfortable putting my name on this?
If the last answer is no, keep checking or leave the claim out.