How to check a generated text, in two minutes
Echipa HR 365 · reviewed 2026-09-10 · 5 min read
The problem with a generated text is not that it is badly written — it is usually better written than what you would have produced in five minutes. The problem is that it is equally fluent whether it is right or wrong, so you cannot use writing quality as a signal. The check has to be deliberate, on five fixed points.
The five controls
| Control | What you look for | Why |
|---|---|---|
| Figures and dates | Any number, percentage, deadline | Plausible is generated as easily as correct |
| Claims about you | What it says about the company, product, process | The model cannot know and will fill in |
| References and sources | Titles, authors, legal provisions | The most frequent form of convincing invention |
| What is missing | Exceptions, edge cases, conditions | Generated text goes for the general case |
| Tone | How it would sound read aloud by you | Correct and wrong-toned does more harm than clumsy and authentic |
What is missing: the most valuable control
A generated text almost always describes the general case and skips the exceptions — because the exceptions are exactly the part that belongs to your context. And the reader looking for an answer usually has an edge case; which is why a text that is “correct and complete on the surface” is often useless.
The practical test: read the text from the perspective of the most complicated case you know. If its answer is not there, you add it — and that is usually the only thing the text is worth reading for.
What cannot be checked at all
- Whether the text reflects how your company actually works. Only you know.
- Whether the example given ever happened. It sounds real and usually is not.
- Whether the nuance that mattered for your situation was lost along the way. It does not show in the text.
- Whether someone on the team will unfairly recognise themselves in a generic example.
The second point produces the most unpleasant situations: an invented example, published as though it came from the company’s experience. It is not the model’s error — an example was requested — but the error of whoever let it pass as real.
When checking costs more than writing
The point at which it stops being worth it
A one-page text. Written by hand: 40 minutes. Generated: 5 minutes. Checking depends on how many verifiable claims it contains.
- Internal process text, no figures
- 5 + 10 min checking = 15 min
- Text with five figures and two references
- 5 + 30 min checking = 35 min
- Text describing specific internal rules
- 5 + 40 min = 45 min, more than by hand
- Practical rule
- the more the text depends on your context, the smaller the gain
Nu intră în calcul:
- the cost of a published error, which cannot be averaged — it is rare and disproportionately expensive
The useful conclusion: AI is most valuable for well-written generic text — a first draft of an advert, a standard email, a structure. It is least valuable exactly where you would expect it to save the most: describing your own processes.
How do I tell that a figure is invented?
You cannot tell from the text — that is the problem. The only method is to check each figure at source, or remove the ones you cannot check. The second option is often enough.
Is it acceptable to publish a generated text, if I checked it?
Yes, and it is a question of honesty rather than quality: if the reader would expect to know how it was produced — for instance for a text signed by a person — then you say so.
Where to start
On your next generated text, mark every figure and reference first, then verify each one. The rest of the check takes two minutes; that is the part that matters.
Then read it from the perspective of the most complicated case you know. What you add there is usually the only original part of the text — and the part worth reading it for.
Texts that reach people, not just you
A wrong internal text gets corrected. One that reaches candidates, employees or clients produces an effect that cannot be withdrawn. For those, the check has one extra step: someone who did not write the text reads it before it goes out.
| The text | Who checks it | What they look for |
|---|---|---|
| Email to candidates | A second recruiter | Tone and implied promises |
| Internal announcement to the whole company | The manager of the area touched | Claims about the process that might be false |
| Training material | Whoever does the work daily | Accuracy and missing exceptions |
| Public text about the company | Whoever owns communications | Consistency with what was said elsewhere |
The third row is the most valuable and the most often skipped: training material checked only by its author almost always contains the general case and no exceptions — and the exceptions are exactly what learners ask about.
Content generation inside the platform, with a check before publishing
What is generated stays as a draft until you confirm it, and the sources and figures sit alongside — so checking is a step in the flow rather than a personal discipline.
Free account, every module for 7 days, no card required.