Before Trusting an AI Answer: Six Checks

First check: does it answer your real question?

Start by rereading your request and formulate the expected result in one sentence. Did you ask for a list of ideas, a factual answer, a calculation or a decision? An AI can produce a general explanation when you were expecting a comparison between two specific options. In this case, the fluidity of the text masks a simple problem: the requested work is not completed.

Let’s take a fictitious example. You ask to compare two hosting offers for a site. The answer describes the usual hosting benefits, but does not indicate what is included in each offer. Even without a visible factual error, it does not allow a choice. The first check therefore consists of finding your criteria in the response: price, scope, limits and information still missing.

You can ask: “Rephrase my request in one sentence, then indicate which points your answer covers and which remain open. » This step helps spot a discrepancy. It does not prove that the content is accurate, but avoids checking at length a text that responds next to it. To understand the possibilities and limitations of a recent model, consult our Astra file.

Second check: do the sources really exist?

When a response advances a test result, rule, or product feature, look for the source. Open the link and find the passage that supports the statement. A credible title or an address that looks like the right domain is not enough. The document can exist without containing what is assigned to it. The verification concerns the content, not just the presence of a link.

Choose the document from the organization responsible for the information for the characteristics of a product or the terms of a service. For a performance announced by a supplier, keep this attribution: “the publisher announces” does not mean “we measured”. If several articles repeat the same announcement, they do not necessarily constitute several independent validations. You have to look at where the initial number comes from.

A useful guideline is: “For each verifiable statement, give the source and indicate the relevant passage. Separate sourced information from your analysis. » If the tool cannot view a page, it must say so. You can retract an unverified statement instead of asking for a more convincing restatement. This is especially important before publishing an article on your company website.

Third check: are the date and context appropriate?

Information accurate last year may be outdated. Look at the publication date, update date, and context of the source. A feature may only be available for certain accounts or in a given country. A rate may concern annual billing. A study may focus on a category of users very different from yours. These details change the scope of the answer.

Fictitious example: an assistant recommends a “free” directory with a link to your site. The home page still talks about free, but the form specifies that new registrations are suspended. The recommendation is unusable today. The control must go as far as the step necessary for your action. Just looking for a sentence that confirms the initial idea is not enough.

Note the date you verified rapidly changing information. For a shared document, add a person responsible for reviewing it. The guide for public officials published by DINUM recalls that the person remains responsible for their productions and decisions. This working principle is useful well beyond administration. Read the user guide.

Fourth check: are the numbers recalculated?

For a total, percentage or price comparison, redo the calculation with a calculator or spreadsheet. First check the starting data: period, units, possible taxes and included items. A correct formula applied to the wrong base produces a false result. Asking the same AI if it is sure of its calculation is not an independent verification.

Here is a fictitious example. A task usually takes forty minutes. With AI, you spend eight minutes preparing, four minutes waiting, and sixteen minutes correcting. The total is twenty-eight minutes. The saving is therefore twelve minutes, or 30% of the initial time. To say that labor is “ten times faster” when comparing only the four minutes of treatment would be misleading.

On a series of files, also look at the differences between the cases. An average can hide many costly failures. Maintain the number of files, total time and important reworks. If you use an estimate, display it as such. This transparency allows a colleague or client to understand your reasoning and change an assumption without starting from scratch.

Fifth check: what is missing from the answer?

Visible errors attract attention, but omissions can be more difficult to spot. A summary may leave out an exception. A comparison can overlook a recurring cost. A debrief can turn an open question into a decision. So go back to the source with a short list of elements to keep: commitments, dates, amounts, reserves and people responsible.

Ask explicitly: “What information are you missing to conclude? What passages in the document contradict or limit your answer? » A response that recognizes a gap is often more useful than a text that seems to solve everything. In a documentary assistant, this behavior should be tested with questions whose answers are not in the files.

For public content, also check what could be misunderstood. A phrase like “this technology improves care” is broader than “a tool assists in note writing.” The first may suggest a demonstrated clinical benefit. The second describes limited use. The article on AI in eight sectors applies this distinction between observed uses and possible benefits.

Sixth check: What happens if you take action?

Adapt the level of control to the consequence. A title idea can be easily replaced. A sent email, a publication, a payment or the deletion of a file require further verification. Before an external action, look at the final content, destination and scope. A general instruction does not guarantee that the tool has understood all the details of your situation.

For personal or confidential information, also review what you transmit to the Service. The CNIL brings together practical resources for small businesss and small businesss, particularly on the choice of uses and precautions. View these resources. For an important medical, legal or financial question, use competent sources and a suitable professional for the decision rather than treating an AI response as validation.

A simple work rule consists of separating preparation and validation. The tool can produce a proposal; one person controls what involves the company. The system must also allow a return to a manual method. If no one knows how to explain how a decision was made or correct an error, automation has exceeded the level of mastery available.

An example of control in a few minutes

Let's imagine that you are preparing an article on new software. Start by marking the factual statements: date, functions, price, announced results and access conditions. Open the corresponding sources. Recalculate the comparisons. Remove what is not confirmed. Finally, reread the text like a beginner reader: would he understand that a result comes from the publisher and not from your own experience?

Then keep a verification note: sources consulted, date, calculations and unresolved points. This trace facilitates a future update. It also prevents a colleague from reintroducing a statement that was dismissed because it seemed convincing. The goal is not to create a cumbersome procedure for each sentence, but to focus attention on the elements that guide a decision.

Frequently asked questions about AI trustworthiness

Can a very powerful AI still make mistakes?

Yes. Good performance on tests does not guarantee the correctness of a particular answer. The context may be incomplete, the source outdated or the request ambiguous. Model power can improve the job, but it is no substitute for control fit for your purpose. Especially check the items you are actually going to use.

Is asking a second AI enough?

This may reveal disagreement, but two identical answers do not prove a fact. The tools can rely on the same erroneous information. Return to the original document or independent calculation to decide. A second reading is an aid to review, not final authority.

Should we check everything with the same attention?

No. Focus efforts on the facts, commitments and actions that would result in consequences for error. For an idea session, the goal is exploration. For a publication or a customer response, the demands rise. Defining these levels in advance helps the team work smoothly with each use.

Transform these reflexes into a team method

You can summarize the method on one page: question, sources, date, figures, omissions and consequences. Add one accepted and one rejected example from a mock test. This support makes expectations concrete and facilitates the arrival of a colleague. The AI implementation plan explains how to integrate this control into a first driver.

ISS-AGENCY supports companies that wish to clarify their digital uses. Present us a task and an expected result via the contact page. We can discuss the points to test and the necessary controls. Useful AI should help you move forward with an outcome you understand, not just text that feels confident.