Responsible AI & Analytical Governance

Your AI Training Needs a Failure Drill

Use evidence checking, decision boundaries and escalation to connect AI learning with responsible workplace practice. Identify whether the next improvement belongs in learning, policy, source access or reviewer support.

5 min read

PsychFlo Editorial

Recognising a transcription error and recognising a mistaken interpretation are different skills. A 2026 study in Harvard Data Science Review makes that difference visible. In an experiment involving 2,784 participants reviewing emissions-table suggestions, straightforward digit errors were corrected much more often than errors requiring conceptual understanding.

That is a challenge for workplace AI training. Knowing that an output needs checking does not establish that someone can interpret the underlying material. Leaders need to observe decisions in the workflow where the knowledge will be used.

The distinction is timely. In IBM’s September 2026 study, 71% of surveyed chief human resources officers identified supervising, validating and overriding AI outputs as the most essential workforce skill; only 29% of employees ranked judgement as important. The study surveyed 1,500 HR leaders and 8,800 employees. These are reported priorities, not observed measures of competence. They suggest that leaders should make the expected behaviours concrete.

One practical approach is a controlled failure drill: a short assessment that tests evidence checking, handoffs and operating rules together.

Start with an actual decision, not a catalogue of AI risks

Choose one bounded task the team already performs. It could be reviewing an account summary, checking an internal briefing or approving a proposed administrative change. Avoid live customer information and consequential real-world actions.

Create a small packet with the source material, a proposed AI output and the ordinary operating instructions. Include an error whose significance depends on understanding the task. A fabricated quotation or conspicuous typo is less useful than a plausible statement supported by the wrong source, an ambiguous instruction or an action that exceeds the tool's approved purpose.

The emissions-table research provides a concrete starting point. Its test material included values assigned to the wrong reporting year or emissions scope. Checking whether a number appears in a source is insufficient when the number belongs to a different category. An effective assessment needs to test that distinction.

Define the expected reasoning before running the drill. Otherwise, a facilitator may reward a confident answer or reinterpret the exercise around what participants happened to do.

Observe four decisions

First, can the participant find the evidence? Ask them to identify which source supports the disputed claim and what remains unknown. Repeating a general rule about checking AI is not the same as checking this output.

Second, can they identify the limit of their authority? The participant should distinguish a draft they may edit from a commitment they may not approve. If the instructions are unclear, recognising that ambiguity is a useful result.

Third, can they find a workable escalation route? Does an alternative exist when the designated reviewer is unavailable? Does the employee know which action may continue and which requires approval?

Fourth, can they leave a usable handoff? Ask for a short record of the issue, the evidence, the proposed next step and the responsible role. A concern that disappears into an unmonitored inbox has not completed its journey.

These decisions test more than recall. They reveal whether knowledge, authority, information access and support fit together.

Include a correct output

A drill made entirely of mistakes can teach participants that the safest answer is always to reject the system. Include at least one correct output and ask for the same evidence-based assessment.

This matters because human review is not automatically reliable. In the emissions-table experiment, requiring corrections changed both missed errors and unnecessary corrections, with a small improvement in overall accuracy. The authors caution against generalising from one task and participant population.

The managerial lesson is to test both missed mistakes and unjustified rejection. More corrections do not necessarily mean better judgement, and a compliant-looking process does not establish a useful result.

For the drill, record why an output was accepted, changed or escalated. A well-supported acceptance should count as a successful decision. So should a justified pause when essential information is missing.

Diagnose the process before blaming the participant

Real organisations are already connecting training to how tools are made available. The European Commission's account of Ingka Group's AI literacy approach describes foundational learning as a condition of access to advanced generative AI tools, with targeted training for higher-risk roles. It documents a company practice, not an independent demonstration of effectiveness.

An assessment should examine the operating conditions around that learning: access to current sources, compatible performance expectations and a reviewer who can actually respond. A failure in one of those conditions calls for an operational fix.

More training might help in some of these cases, but it will not repair an unavailable reviewer or an inaccessible document.

Debrief the exercise with two questions: what did the participant need to know, and what did the organisation need to provide? Assign an owner to each change. Correct the source, clarify the authority or repair the handoff before asking people to repeat a course.

Keep results developmental. A short exercise is not a validated assessment of an individual's overall competence, and it should not become a hidden employment-ranking tool. Explain the purpose and how observations will be used before people participate.

Make the first drill small enough to repeat

A first session could take forty-five minutes: introduce the task, allow independent review, compare decisions and agree process changes. That schedule is a suggested starting point, not a research-established optimum.

Use a fresh, equivalent case when checking improvement. Repeating the same planted error mainly tests whether people remember it. Look for evidence that the revised process makes a sound decision possible without coaching from the facilitator.

An exercise may reveal disagreement among experienced colleagues. Resolve whether the disagreement comes from unclear policy, different interpretations or genuinely missing evidence. Do not force a neat score where the correct organisational response is further review.

A failure drill is not a substitute for assurance, specialist review or appropriate formal training. It is a way to find the gap between knowing a principle and acting on it under realistic constraints.

Before buying another round of AI training, ask what happens when the output is convincing, the evidence is incomplete and the usual approver is unavailable. The answer will tell you whether the next investment belongs in a course, a process or the support people need to do the work responsibly.

AI assistance was used in research discovery, structure and drafting. Research findings and organisational examples are attributed to the linked sources.

← All research and insights

PsychFlo guide

Guided chatbot · prepared answers, no language model

Find the right course, software product or consulting review. Messages stay in this page’s memory; don’t enter personal or confidential information.

Full decision copilot