Common Mistakes People Make When Using AI Writing Tools

You can usually tell within two sentences. The email restates your question back at you, says the project is exciting, mentions a commitment to quality, then asks you to let them know if you have questions. It could have been sent to anyone, about anything.

Nobody set out to write that. Somebody typed one line into a chat box, got something that looked professional, and sent it. That is the shape of nearly every mistake here.

Mistake one: giving no context#

This causes most of the others, so fix it first. A vague request gets a vague answer, and the tool has no choice about it. It produces the most likely continuation of what you gave it — so feed it something generic and generic is what comes back. That falls straight out of how these tools work.

Here is the version that produces the email above:

Write a follow-up email to a client about a late invoice.

And here is the version that produces something you can send:

I am a freelance illustrator. A repeat client — friendly, usually pays on time — is 26 days late on a £1,400 invoice, and the one reminder I sent went through the accounting system, which they may not have seen. Write a short personal email, four sentences maximum, that assumes an oversight rather than a refusal, names the amount and date, and asks for a payment date. Warm, not passive-aggressive, and no thanking them in advance.

The second takes about forty seconds to write, and what changed is not clever phrasing. You supplied who you are, the history, the constraint, the tone, the length, and one thing you specifically do not want.

The habit: tell it what you would tell a competent freelancer who has never met you. Situation, goal, constraints, tone, length.

Mistake two: accepting the first draft — or editing it into nothing#

Stopping too early#

These are opposite errors from one source: not knowing what the second round is for. Most people ask once, get something mediocre, and either send it or give up. But the first output is a guess at the average version of what you asked for. The value is in round two, where you say what is wrong.

Worth saying back: "the second paragraph says nothing, cut it". "Too formal — write it the way I speak". "Give me three openings, not one". "That is 300 words; I need 120".

Sanding it down until it is smooth and empty#

The subtler failure, common in people who have used these tools a while. Another pass, then another, then a style checker with every suggestion accepted, and you get prose with no rough edges and nothing in it.

Specific claims soften into general ones. A concrete number becomes "significant". A blunt sentence becomes balanced. Each edit looks like an improvement, and the cumulative effect is text nobody could disagree with — which is text nobody needs to read.

A stopping rule: if no sentence in the draft is one someone could argue with, you have edited too far.

Mistake three: believing it about facts#

The failure is not that these tools are sometimes wrong. It is that wrong output looks exactly like right output: same confidence, same tone, no hedge, no tell.

Citations are the worst case, because a fabricated one is assembled from real parts: a plausible author in the right field, a title in the right register, a real journal, a sensible year. It survives a glance. It does not survive a search.

A second version: assuming it knows about recent events. Training stops at a cut-off date, and unless the tool is searching, it answers about last month's news from a picture of the world that predates it — without warning you.

Mistake four: letting it flatten your voice#

This is the one readers notice, though they rarely name it. They do not think "this was written by AI". They think this person is boring, or is hiding something.

The tools default to a register best described as inoffensive corporate: the average of an enormous amount of business writing, and averages are bland by construction.

There is a vocabulary tell too. A small set of words now reads as a signature: delve, seamless, robust, leverage, unlock, game-changer, elevate. Always corporate filler; now corporate filler with a source attached. Search your draft and cut every one.

The fix that beats all of these: write the first two sentences yourself, then hand over. What you write first sets the register for the rest.

Mistake five: what you paste in, and what you ask for#

Text you paste into a consumer AI tool has left your control, and some providers use submitted content to improve their models unless you turn that off. So: no client contracts, no unpublished work belonging to someone else, no personal data about third parties, nothing under an NDA, no credentials.

If a real document needs work, strip the names, figures and company first, then put them back. It takes a minute. Our safety and privacy guide covers the settings worth changing today.

The other half of this mistake is asking for work the tool cannot do well: current information without search, anything needing real knowledge of your organisation or last week's meeting, precise arithmetic, legal judgement. And repetitive rule-based work — moving the same data between two systems every week is an automation job, with different tools.

Frequently asked questions#

How do I stop AI writing from sounding like AI writing?#

Three things do most of the work. Write the first two sentences yourself so the register is yours. Cut the signature vocabulary. Replace every vague claim with a specific one: a number, a name, a real example. Flatness comes from having nothing concrete in the text, not from the tool itself.

Is it safe to paste work documents into a chat tool?#

Not by default. Assume anything you submit may be retained and, depending on the provider and your settings, used to improve their models. Keep client data, information about other people, unpublished work and anything under an NDA out of consumer accounts. If you need help with a real document, remove the identifying details first, or use a business plan your employer has approved.

Can I trust the sources and citations it gives me?#

No, not without checking each one. Fabricated citations are common and convincing because they are assembled from realistic parts: a plausible author, a real journal, a sensible year. Search for every reference before you use it. If the tool has web search on, follow the links rather than trusting its summary of them.

Will my writing get worse if I use these tools?#

It can, if you hand over the thinking as well as the typing. People who stay sharp use these tools for form — restructuring, tightening, second opinions on phrasing — while keeping the argument and the examples. Handing over the whole job repeatedly is where skill fades, in writing as in anything else you stop practising.

A test to run this week#

Take the last thing you sent with AI help in it and read it as the recipient. One question: could this have gone to someone else, about something else, with three words changed?

If yes, the problem was context going in, not the tool. Rewrite the request with the situation, the constraint, the tone and the length, and see what comes back.

And if you are still choosing a tool, the beginner guide sorts the options by job, not brand.