By Andreas Ehstand
AI upskilling can teach people to prompt, draft, summarize, and automate. These are useful capabilities. Training also needs to address a different task: deciding whether a particular output is sound enough to use.
Producing an answer and evaluating it require different actions. A learner may create a fluent draft while overlooking an invented source or an unsupported conclusion. Learning and development teams can make those checking actions explicit, give learners opportunities to practice them, and examine what happens on new tasks.
My work connects human-AI collaboration with questions of learning, judgment, and knowledge transfer. A useful coaching question is whether a skill can be demonstrated under specified conditions. That question also helps frame AI training: What should a learner be able to check, using which sources, and with what support? The routine below can be incorporated into an existing curriculum, with time allowed for preparing the practice materials and reviewing the results.
Why ‘Checking’ Is a Trainable Skill—Not a Disposition
Three useful error types to include in workplace practice are:
1. Invented or misrepresented sources. A citation may point to a nonexistent source or to a real source that does not support the attributed claim.
2. Outdated states presented as current. The answer was true two years ago. An answer may present an older state as current without making its age clear.
3. Plausible-but-wrong conclusions from correct sources. The documents are real and accurately quoted, but the confident conclusion drawn from them does not follow.
These errors may be difficult to notice in a fluent answer. Checking can be taught as a set of observable actions:
- Locating the source
- Checking its date
- Comparing the claim with the source
- Recording what remains uncertain
The exercise below is a proposed way to practice and assess those actions; its training effect has not yet been established.
The 5-Step Review Routine
The routine below makes the checking steps explicit while allowing time for verification.
The time required depends on the claim and on access to reliable sources. A short triage may identify the claim that needs checking; completing the check can take longer. Measure the time in your own workflow rather than promising a fixed duration. Learners apply it before accepting any AI result that will leave their hands.
- Origin. Which source supports the core claim—and does it exist? (Open the link. Actually open it.)
- Currency. What is the date of that source? Could there be a newer version or source?
- Support. Does the source really say what the output claims—or merely something adjacent?
- Cross-check. Verify the disputed point against an independent source with relevant authority, or ask a qualified colleague to explain the evidence. A second AI answer or a search result can suggest where to look, but agreement alone is not verification.
- Note. Write one line: What was checked; what remains open.
The note makes the learner’s stated checking process available for review. It is evidence of what was recorded; its accuracy still needs to be checked against the underlying material.
Introduce the routine in a short session, then practice it on realistic, synthetic, or appropriately cleared material. Two or three checked outputs per session is one possible starting schedule; it is not an established optimal dose. Adapt the time and difficulty to the workflow.
The Part Coaching Adds: Verify the Skill with Seeded Errors
The coaching analogy suggests a practical next step: observing what learners can check on a specified task. To find out whether the checking skill actually exists, you test it the way a coach tests a serve: under realistic conditions, with a known target.
Concretely:
- Take a practice task that mirrors real work—summarizing a document, drafting a client answer, researching a factual question.
- Seed it with one known error of the three types above: an invented source, an outdated figure, or a conclusion the material does not support.
- Announce clearly that practice materials in this program may contain planted errors. (Transparency is non-negotiable—this is training, not surveillance.)
- Count how many learners catch the planted error.
For each item, record how many learners identify and correctly explain the seeded error. In a hypothetical example, 4 of 20 learners do so before practice and 15 of 20 do so afterward.
That would be an observed difference on those assessments, not proof that training caused a lasting capability change. Use new, comparably difficult items, specify the conditions, and include correct items so that unnecessary error flags are visible. A later assessment can examine whether the performance persists.
Three proposed design rules guide the exercise:
- Use practice tasks only, never live customer work or real personnel data.
- Rotate the error types so learners cannot pattern-match to “the trick is always a fake citation.”
- Debrief every seeded task by walking the routine backward: Which of the five steps would have caught this one?
Keep the denominators explicit. Report detected errors out of all seeded errors, and incorrect error flags out of all correct items, separately. Record unresolved cases as unresolved. These measures describe performance on the assessed material; they do not by themselves establish performance in every workplace task.
What the Catch Rate Does for L&D
Alongside completion counts and learner feedback, these assessments provide information about observed checking behavior. Report it together with the time required, task difficulty, support allowed, and unnecessary error flags.
If comparable assessments show faster work but fewer detected errors, that is a reason to investigate. If speed and detection both improve, that is encouraging evidence about the measured tasks. It does not, on its own, establish the overall success of an AI rollout or the cause of the change.
A 5-Step Process to Start Next Month
- Pick one workflow and prepare a baseline set and a different, comparably difficult, follow-up set. Include both correct items and the selected error types
- Before teaching the routine, run the transparent baseline assessment. Record correct detections, incorrect flags, unresolved cases, and time.
- Teach the five-step routine and practice it on separate, synthetic or appropriately cleared examples.
- Introduce the one-line checking note and review a sample against its sources.
- At a preselected follow-up date, use the new assessment set under comparable conditions. Report the separate measures and the limitations. A four-week interval is a practical choice for this proposal, not a validated optimum.
The exercise can be run with ordinary documents and a record sheet. Its value has to be examined in use: Define the task, observe the checks, and adjust the training in light of the results.
Image credit: phuttaphat tipsana
