Checking an AI Answer Is Not the Same as Changing Your Decision

A person can doubt an AI answer, check it against a source and still make the same wrong decision. Those are not contradictory outcomes. They are three different behaviours, and each can fail on its own.

A September 2026 review in Current Opinion in Psychology separates scepticism, verification and reliance in goal-directed use of organisation-backed AI advisers. The distinction matters for workplace training. A warning may make someone cautious. A citation may prompt a click. Neither proves that the check succeeded or changed what the person did.

Measure three behaviours, not one confidence score

The review finds that most studies measure reliance, while scepticism and verification are less often measured directly. It also reports that general hallucination warnings have weak and mixed effects across the evidence reviewed. More targeted cues about where an AI system tends to fail may help, but the evidence remains limited. The practical lesson is modest: do not treat a disclaimer, a source link or a self-reported confidence score as proof of safe use.

Worked example: an internal travel-policy adviser

Imagine an employee asks an internal AI adviser about the dinner allowance for a trip. The answer says €60 and links to the travel policy. The current policy says €45. The employee feels uncertain and opens the document, but scans the wrong section, keeps the €60 figure and submits the claim.

This example reveals four possible outcomes. A person may not notice the risk. They may notice but not check. They may attempt a check and miss the error. Or they may find the error but fail to change the action. A single score called “trust” hides those differences and gives the training team little guidance about what to practise next.

Design the review around the decision consequence

NIST’s AI Risk Management Framework calls for documented human oversight, context-aware interpretation of outputs and feedback mechanisms that can support appeal or override. The UK Government AI Playbook similarly says people remain accountable for outputs and should validate them in proportion to impact and risk. For a low-stakes internal note, NCA may take a minute. For a public, financial, legal or people-related decision, the source, reviewer and release rule should be explicit.

The European Commission’s AI-literacy guidance also emphasises context, purpose, user knowledge and the skills needed for human oversight. That supports task-specific practice: the relevant check for a policy answer differs from the check for a calculation, a customer message or a ranking of people.

Build a ten-minute NCA exercise

Pair the method with the “nothing found” coverage check, the UNIT number check and the evidence-versus-hypothesis sales brief. Each one gives the check stage a concrete form for a different type of work.

The aim is not permanent distrust. It is appropriate reliance. Good AI practice makes the path visible: notice the risk, check the right evidence, then let the evidence change the decision.