The email is impeccable, the summary crystal clear, the recommendation almost irresistible. Then a colleague asks: “Where did that information come from?” Silence. With generative assistants, the trap is not just believing an error: it is no longer feeling the need to check. Critical thinking is becoming an everyday workplace skill, provided it does not turn every paragraph into a formal inquiry.
Fluency is not proof
An assistant can explain a complex subject, adopt a profession’s vocabulary and present a conclusion with confidence. That fluency makes reading less demanding. It can also lead us to confuse polished wording with sound reasoning. A well-organized answer can easily seem better supported than a tentative note, even when the latter rests on verified observations.
The problem was apparent well before September 2026. In 2023, in the US case Mata v. Avianca, lawyers submitted nonexistent legal citations generated by ChatGPT to a court. Checking them with the same tool did not correct the error. The lesson extends beyond the law: asking an AI to confirm its own answer does not constitute independent verification.
This documented case serves as a reference point here. As for the September 2026 context, the developments discussed are forward-looking analysis, not an account of future events established in advance. As assistants become embedded in office software, one risk can be anticipated: their answers become less recognizable as machine-generated output, making them easier to reuse without scrutiny.
Check according to the consequences, not your anxieties
Checking everything would be a poor approach. If every suggested headline requires ten minutes of research, the productivity gain disappears. Conversely, treating medical advice like a slogan suggestion risks far more serious consequences. The right starting point is therefore not “Is this AI reliable?” but “What happens if this answer is wrong?”
Three criteria allow for a quick decision: the severity of an error, how easily it can be reversed and how widely the answer will be shared. An approximation in a personal draft is easily corrected. The same approximation in a client presentation puts a reputation on the line. In a hiring decision, it can directly affect someone’s life.
- Low stakes: ideas, rewording, alternative presentations. A careful reread is often enough, with attention to any added facts.
- Medium stakes: document summaries, comparisons of solutions, meeting preparation. Verify the claims that support the conclusion.
- High stakes: health, law, safety, finance or decisions affecting people. Require appropriate sources and qualified review before acting.
This graduated approach avoids two dead ends: automatic trust and paralyzing distrust. Above all, it puts responsibility where it belongs. The assistant’s confident tone should not determine the level of scrutiny; the intended use of its answer should.
Identify the sentence holding the entire argument together
Imagine a sales manager preparing a proposal. The assistant recommends a substantial discount on the grounds that “competitors generally offer contracts with no commitment.” The table is neatly presented, and the arguments flow smoothly. Yet a single claim underpins the entire recommendation: the terms competitors actually offer. That is what needs checking before discussing the discount.
This habit involves looking for the pivotal claim: the one that would change the decision if it proved false. It might be a figure, an effective date, a software feature or an assumption about customers. You can ask the tool to identify it, but it is up to the reader to judge whether it is truly decisive.
Three questions to help you pause at the right point
- What is being presented here as a fact when it could be an assumption?
- Which piece of information would change my decision if it were wrong?
- What external source could quickly settle the question?
These questions shift the focus. Instead of rereading the entire answer with a vague sense of suspicion, you isolate a testable point. In our example, consulting competitors’ commercial terms will be more useful than asking for five rewordings of the recommendation.
Step outside the conversation to verify
A displayed reference is not validation. It may be fabricated, outdated, misinterpreted or simply unrelated to the conclusion. Even when an assistant searches the web, it remains necessary to distinguish three things: whether the document exists, what it actually says and what can reasonably be inferred from it.
Effective verification prioritizes a source close to the fact: an official text for a regulatory obligation, the vendor’s documentation for a software feature, the original publication for a scientific finding. The next step is to open the document, locate the relevant passage and check its date and scope. A rule that applies in one country does not automatically apply elsewhere.
Two websites repeating the same press release do not constitute two independent confirmations. Conversely, a single explicit primary source may be enough for a simple question. Verification is measured by its relevance, not the number of tabs open. If the point remains uncertain, it is better to flag that uncertainty than to keep searching without a stopping criterion.
Challenge constructively, without playing prosecutor
The assistant can nevertheless help prepare a challenge. “Separate the facts, assumptions and recommendations.” “What serious objection could invalidate this conclusion?” “What data is missing to make a choice?” These instructions make the answer easier to scrutinize. They do not guarantee its truth: a model can also produce an appealing but unfounded objection.
You also need to watch how you frame your own questions. “Explain why our strategy is the best” invites the tool to build a case in its favor. “Compare this strategy with a credible alternative using the same criteria” opens up the discussion. Critical thinking sometimes starts before the answer, with a refusal to use AI as a machine for reinforcing a belief.
Make verification a collective habit
Within a team, a simple rule can change practices: every AI-assisted recommendation should identify its pivotal claim, the source consulted and the remaining uncertainty. There is no need for a compliance file for every email. A brief note is often enough to open the reasoning to discussion and prevent an assumption from becoming a fact through repeated copying.
Management also has a role to play: treating “I haven’t checked yet” as useful information, not an admission of incompetence. An emphasis on speed should not reward whoever is quickest to pass along an answer that looks impeccable on the surface. It should account for the time needed to verify what matters.
What next? If assistants become even more prevalent in everyday work, knowing how to challenge them could matter as much as knowing how to use them. The habit to build remains simple: identify the stakes, isolate the decisive claim, consult an external source, then decide or suspend judgment. Neither blind trust nor endless investigation: proportionate vigilance that protects both time and the quality of decisions.
