The AI Answer Looks Right. Who Checks It?
A convincing answer still needs evidence. Who checks it before the work leaves your team?
The researchers gave the search tools a head start: passages copied directly from published news articles. They wanted the tools to identify each story’s headline, original publisher, publication date and web address.
The answers still went wrong.
In research published in March 2025, Klaudia Jaźwińska and Aisvarya Chandrasekar tested eight AI search tools using excerpts from 200 articles. Across 1,600 queries, more than 60 per cent of responses were classified as incorrect. The tools sometimes fabricated links or identified the wrong source, often without expressing uncertainty. Read the citation study.
The study examined news-source retrieval and captured the tools available at that time. Its findings cannot stand in for the accuracy of every AI product or every answer. But the exercise exposed a practical problem: an answer can arrive with the appearance of verification while still needing to be verified.
For professional teams, that makes responsibility a central question. Before an AI answer becomes a published statement or a recommendation, who has checked the evidence—and what, exactly, did they check?
An explanation needs evidence
When an AI answer seems questionable, asking the tool to explain itself feels like a sensible next step. A detailed explanation may expose an assumption or help a reviewer follow a calculation. It can also create more material that requires scrutiny.
In April 2025, Anthropic published research testing whether reasoning models accurately disclosed influences on their answers. Researchers supplied hints to Claude 3.7 Sonnet and DeepSeek R1, then examined whether their reasoning acknowledged using those hints. When the hints influenced the answers, the explanations frequently omitted that influence. Read the reasoning study.
These controlled experiments involved particular models and prompts. They tested the faithfulness of explanations, rather than the overall factual accuracy of everyday responses.
For a reviewer, the useful lesson is to look beyond the explanation to the evidence beneath it. If a tool says a report supports a conclusion, open the report. Find the relevant passage and check whether it supports the conclusion as written, including its conditions and qualifications.
Asking the same tool whether it is certain may generate a useful correction. It does not provide independent confirmation.
Confidence can change the review
The person checking the answer brings another variable to the process: how much they trust the tool.
A study presented at the 2025 CHI conference surveyed 319 knowledge workers, who provided 936 examples of using generative AI at work. Researchers found that greater confidence in AI was associated with lower reported critical-thinking effort. Greater confidence in participants’ own abilities was associated with more critical thinking. Read the knowledge-worker study.
The findings concern people’s accounts of their behaviour and effort. They do not establish that AI causes a decline in thinking skills or measure a long-term loss of ability.
That distinction is easy to lose in a summary. Compare two possible descriptions of the finding:
Overstatement: “Research proves that AI makes employees worse at critical thinking.”
Supported wording: “In a survey of knowledge workers, greater confidence in AI was associated with less reported effort in critical thinking.”
The second sentence preserves the study’s design and the limits of its conclusion. Removing those details changes what the reader is being asked to believe.
This is substantive editing: checking whether a sentence remains faithful to the evidence after it has been made shorter, smoother or more compelling.
Give verification a named owner
“Please review” leaves considerable room for interpretation. One colleague may check tone. Another may approve the recommendation. A third may assume the figures were verified before the document reached them.
A stronger handoff identifies what requires checking and who will do it. For an article, the writer might verify citations and quotations, a subject expert review technical interpretations, and an editor check that the final wording preserves both. The responsibilities can overlap. They should be explicit.
A useful review separates four questions:
- Does the source exist? Open the original publication and confirm its title, author and date.
- Does it support this claim? Locate the evidence, rather than relying on a relevant-looking headline or summary.
- What limits the conclusion? Check the population, methods, timeframe and uncertainty. Watch for an association being rewritten as a cause.
- Does the claim belong here? Decide whether the evidence is relevant to this audience and substantial enough to carry the argument.
The depth of review should reflect the consequences of an error. A working list of headline ideas needs a different level of scrutiny from a public report or a recommendation someone will act on. When a reviewer lacks the expertise to judge a claim, involve someone who has it.
Keep the supporting passage or calculation with important claims, along with the source link and any unresolved questions. This gives the next reviewer something concrete to examine. It also helps catch changes introduced during editing: a deleted qualification or a broader word can alter a previously checked statement.
Before a document leaves the team, someone should be able to point to the evidence behind its central claims. If nobody can, the review is unfinished.