AI literacy has to reach the decision people actually make

A training session can explain that AI invents citations without teaching a reviewer how to assess a real citation used incorrectly. The reviewer may remember the warning and still approve the recommendation. Preparing someone to use a system in professional work takes a closer account of the mistakes they could encounter and the decisions they are expected to make.

Helen Smith and Jonathan Ives’ rapid scoping review in AI and Ethics, published on 19 October 2026, examines proposed professional ethical guidance for AI use in healthcare. It identifies 124 relevant items and synthesises nine themes containing 33 normative claims. It maps what the literature says practitioners ought to do, rather than demonstrating which training methods improve outcomes.

The review does not establish how marketing or compliance professionals should be trained. It does provide a useful comparison for teams trying to specify what a person must understand before relying on an AI recommendation.

What the healthcare review can establish

The review draws on peer-reviewed work, grey literature and formal guidance, with research undertaken from December 2025 to June 2026. It includes English-language material relevant to UK professional practice. Its scope concerns healthcare professionals using AI, rather than autonomous systems operating without them.

The authors identify gaps in the diversity of voices represented, including patients and people outside English-speaking academic or professional settings. These limitations matter when interpreting the apparent convergence of guidance. The review offers a foundation for developing professional standards, not a validated curriculum or a finding that 33 separate interventions produce safer practice.

For a compliance leader, the distinction between using the interface and judging its output is worth retaining. Training-completion figures will not reveal whether staff can recognise a mistaken interpretation or know when to ask a specialist. Those abilities need to be observed in work that gives people a reason to exercise them.

Different roles need different forms of competence

A senior leader deciding whether to deploy AI needs to assess the business use and the evidence for it. Technical details become relevant where they affect that decision: a limitation might justify further testing or a narrower scope. Learning how to maintain the model would serve a different role.

Consider an AI assessment of a mortgage advertisement that cites a real rule but applies it to the wrong product category. A specialist reviewer would need enough domain knowledge to notice the mismatch and inspect the source. Checking that the rule existed would leave the mistake intact.

A technical team, meanwhile, may need to investigate why the wrong context was supplied, how knowledge is updated and whether similar cases are affected. Giving all three groups the same introductory session would leave important parts of each role unaddressed.

The FCA’s October 2026 AI practice publication makes role-specific training a concrete example of good practice, distinguishing senior stakeholders, technical teams and human reviewers. It also identifies generic awareness training that does not support effective oversight as an area for improvement. Its examples do not create a universal training syllabus.

The three roles would benefit from different exercises. A deployment decision, a disputed recommendation and an investigation of faulty retrieval each require knowledge that a general introduction to AI may never test.

Checking the evidence behind a recommendation

For a marketing reviewer, the central skill may be reconstructing the relationship between a claim and its evidence. That means checking whether the material concerns this product, remains current and supports the entire promise. The same examination is useful whether the initial recommendation came from a colleague or an AI system.

A model’s explanation needs scrutiny too. It may offer a fluent account that is unsupported by the source, and it should not be taught as a trustworthy view of the model’s internal reasoning. A reviewer needs to examine whether the cited material justifies the conclusion.

People using a system need to understand the scope of its application-level performance evidence, including the tasks and situations it does not cover. General confidence in a supplier cannot resolve every local use decision.

A training exercise could present a plausible recommendation alongside incomplete product information. The learner would have to identify the missing evidence before reaching a conclusion. Another could pair a valid citation with an interpretation that goes beyond it. Both would reveal more about the learner’s judgement than asking them to recite a warning about hallucinations, although their effectiveness as training would need evaluation.

The quality of the answer key matters. Historical approvals may contain inconsistent decisions, and some cases genuinely allow reasonable disagreement. Specialists should examine the underlying evidence and preserve uncertainty where it exists. Otherwise, training can reward conformity to precedent rather than sound judgement.

Staff need somewhere to take a disputed finding

A reviewer may recognise a problem and still have nobody available to resolve it before a deadline. An unclear escalation route or pressure to accept recommendations can defeat otherwise sound training. The organisation needs to make room for the behaviour it is teaching.

The authority to intervene in AI-assisted work includes permission to hold a decision while evidence is checked. Without it, training someone to identify uncertainty may simply leave them more aware of a problem they cannot act on.

The same applies to retaining skills. If a team sees only completed AI assessments, junior colleagues may receive less practice in forming an independent interpretation. Selected case discussions or initial unaided assessments can preserve opportunities to learn, although the appropriate balance should be tested in the workflow.

Training can become obsolete even when its account of AI remains accurate. A model update or an extension of the application’s permissions may introduce failures the original exercises never covered. Staff then need guidance on the changed task, not another run through the same introductory module.

Missed qualifications in live work could supply material for a focused exercise. Unnecessary escalations might instead expose unclear source material or uncertainty about who can decide. Examining those incidents would help the team choose between training, better information and a change to the process.

The test comes when a reviewer meets an answer that looks plausible but is wrong. Training has done useful work if they can identify the problem, explain what they need to check and obtain help before the recommendation is used. A completion certificate cannot show whether any of that happened.