Teachomatic what to automate, and what not to

The Skill of Checking the Output

If you teach students one thing about these tools, teach them to check the answer.

It is the skill with the longest shelf life. Prompting technique is already changing under everyone's feet. Which tool is best will be a different answer next year. But the gap between a plausible answer and a correct one is a permanent feature, and closing it is a discipline that transfers to sources, statistics, headlines, and anything else that arrives looking finished.

For a workplace-oriented comparison outside education, this reference offers another way to look at measurement, workload, behaviour, or accountability.

It is also, conveniently, teachable and assessable.

Why students do not check

Not laziness, mostly. Three specific reasons, and each has a countermeasure.

For an external perspective on student digital literacy, privacy, and learning, see Media Literacy Now.

The output has no uncertainty signal. A correct answer and an invented one arrive in identical prose, with the same confidence, the same structure, the same tidy formatting. Every cue humans normally use to gauge reliability has been flattened.

They do not know what would count as checking. "Verify it" is not an instruction anyone can follow without a method.

Checking feels like it defeats the purpose. If the point was to save time, re-doing the work seems absurd. This is the real barrier, and the answer is that checking is much cheaper than producing — but only if you know where to look.

The method: check the checkable

The core move is that you do not verify everything. You verify the parts that are both load-bearing and cheap to check.

Give students a concrete hierarchy, in this order.

Things that exist or do not. Sources, quotations, titles, cases, formulae, dates, people. These are the highest-risk items — fabrication is the characteristic failure — and the cheapest to check, usually seconds. A citation you have not opened does not exist. That single rule prevents most of the trouble students get into.

Numbers. Any figure, and especially any figure that arrives without a year or a source. Ask what kind of number it is: measured, estimated, or asserted — the same three questions we apply to claims about AI in education.

Anything you are going to build on. If a later step depends on this being right, check it now, because an error at step two makes steps three to seven worthless.

Anything that sounds too neat. Round numbers, tidy contrasts, satisfying reversals. Reality is lumpier than a generated summary.

And what to skip: general explanation of well-established material that you can sanity-check against what you already know. That is where the tool is strongest and where checking costs most relative to the risk.

Teach it as a subtraction exercise

The most effective version, and it takes fifteen minutes.

Hand out a generated passage in your subject containing three planted errors: one fabricated source, one wrong number, one plausible-sounding claim that is subtly wrong. Do not say how many.

Ask them to find what is wrong and, more importantly, to say how they decided where to look.

The second question is the lesson. You are not training them to spot these three errors; you are training them to allocate attention. The students who do well are the ones who checked the citations first, and saying that out loud teaches the rest.

Run it once a term with a new passage. It stays interesting because the errors change.

Make it assessable

Verification only becomes a habit if it counts for something, and it is easy to assess.

Ask for the check, not just the answer. "Give me the answer and tell me one thing you verified and how." A line at the end. Cheap for them, informative for you.

Mark the fabricated citation as a fact error, not as misconduct. This matters. A student who submits an invented source has usually not tried to deceive you — they trusted output that looked exactly like the real thing. Treating it as cheating teaches concealment; treating it as an accuracy failure teaches checking. The integrity conversation is a separate one and merging them here does damage.

Give credit for catching something. If a student reports that the tool got something wrong and they corrected it, that is a better demonstration of understanding than a clean answer, and it should be scored that way.

The habit that generalises

Say this to them explicitly, because it is the actual point and it is not obvious from inside the task.

The question how would I know if this were wrong? is not about AI. It is the question that separates people who can be told anything from people who cannot. It applies to a statistic in an article, a confident answer from an adult, a claim in a group chat, a graph on a slide.

Generated text is simply the most convenient practice material anyone has ever had for it: unlimited, plausible, and wrong often enough to reward attention. The rest of this section builds on the same idea. That is a genuinely good reason to bring it into the room rather than keep it out.

What this does not solve

Be honest with them about the limit.

Checking works well for facts and badly for judgement. If the output offers an interpretation, an argument, or a recommendation, there is nothing to look up. All you can do is evaluate the reasoning — which is the harder skill, the one your subject already teaches, and the one that gets quietly skipped when the fluent median answer is available.

So: check the facts, and argue with the reasoning. The second is what the lesson was always for.

The short version