Your AI-generated training looks convincing. That’s the problem

AI-generated training promises speed, but unchecked inaccuracies can quickly become accepted practice across an organisation. Aidar Akmaev outlines five ways to improve reliability, including grounding content in trusted documents, verifying claims, using independent AI reviews and involving subject experts. His approach turns quality checks into a repeatable process for teams.

AI can now put together training material in minutes. But it has one dangerous habit: it gets things wrong in exactly the same confident tone it uses when it gets them right. A mistake doesn’t look like a mistake. It looks like a plausible figure, a tidy sentence, a reference to a study that doesn’t exist. In training, that kind of error is costly, because dozens of people will repeat it, and they will repeat it as a rule.

Reliability has to be built in from the start

I’ve worked as a product designer for more than ten years, and for the last few I’ve been building tools in which AI writes documents. The main lesson is this: checking at the end is not enough. Reliability has to be built in from the start. Here is how I go about it.

1. Build the system before you write the prompt

Most errors are born before the AI writes a single word, at the moment we let it “know” the topic on its own. Give it the source documents instead: the policy, the procedure, the report. Ask it to rely only on them, and to flag anything it can’t support rather than fill the gap.

If your team produces this kind of material regularly, turn this into a shared prompt template. Then the quality no longer depends on who happens to be at the keyboard that day.

2. Every claim needs a source you can open

Not “research shows”, but a specific document and the place in it where the point is made. The rule is simple: no source, no claim. It keeps both the AI and the person honest.

This also turns checking into a simple question, “does this match the source?”, instead of an endless “is this actually true?”.

3. Give the text to a second, independent AI

The model that wrote a text is poor at spotting its own mistakes. It tends to agree with itself. Take a different model, give it the source documents and the finished draft, and set it one job: find inaccuracies, contradictions and claims with nothing behind them.

Different models have different blind spots, and the second one often catches what the first one missed. It’s a cheap check, and a very useful one.

4. A person reads all of it. Every time

No automated check takes away responsibility. Someone who knows the subject has to read the whole thing, from the first line to the last. Not skim it, but read it as if their name were on every slide, because in effect it is. AI cuts the work down. It doesn’t replace the person who answers for it.

5. Every error you find should improve the system

When the second model or the reviewer finds a mistake, fixing it in that one file isn’t enough. Ask why it happened. Was a source missing? Was the prompt too open? Did the topic need an expert?

Adjust the template, and that mistake won’t come back next time. Over time, checking stops being a chore and becomes a system you can trust.

A short list before anything goes out

• The AI worked from our documents, not from memory

• Every claim has a source, and the source can be opened

• A second, independent AI has reviewed the text

• Someone who knows the subject has read all of it

• Any errors found have been fed back into the template

AI’s speed is only worth something if people can trust what it writes. A little discipline at the start and at the end saves both time and trust.


Aidar Akmaev is Founder of NOXI


illuminated must read neon sign

Get Training Journal’s Thursday newsletter straight to your inbox