Where AI Writing Help Works

Discussion of AI writing tends to be either promotional or anxious. A plainer inventory is more useful.

Straightforwardly good

Routine correspondence. Scheduling, confirmations, polite declines, follow-ups. Nobody expects voice and the value is clarity.

Formatting and conversion. Turning notes into a structured document, prose into a table, a transcript into minutes.

Boilerplate. Standard clauses, repeated sections, templates.

Summarising material you supply. Long documents, threads, transcripts. Reading comprehension is what these models do best.

Explaining a difficult passage in something you are reading.

Naming things. Titles, headings, subject lines, variable names. Generating twenty options and picking one is genuinely faster than thinking of three.

Starting when you are stuck. A bad first paragraph you can react to beats a blank page.

Second-language support. Grammar, idiom, register. The strongest use case there is.

Editing your own writing. Critique, cutting, clarity.

Adapting register. The same content for a colleague, a customer and an executive.

Good with verification

First drafts of factual material. Fine, provided every fact is checked.

Research orientation. Finding the vocabulary and the main positions. Verify everything.

Technical documentation. Good structure, plausible errors.

Code explanation. Usually right, occasionally confidently wrong.

Data analysis in prose. Reliable where the tool executes code; unreliable where it does arithmetic in text.

Bad ideas

Anything you will sign without reading. Obviously, and it happens constantly.

Legal, medical or financial documents for actual use. Fluent, plausible, and wrong in ways that matter. Professional review is not optional.

Text where the point is that you wrote it. Condolences, apologies, personal messages, references. The recipient's response to discovering it was generated is worse than any awkwardness in your own words.

Content on topics you know nothing about, for publication. You cannot verify it, so you cannot take responsibility for it.

Anything containing confidential material, in a tool whose data policy you have not read.

Assessment work where it is prohibited.

The category people get wrong

Personal writing. There is a strong temptation to use these tools for messages that matter to another person — a note to a grieving friend, an apology, a message to a partner.

The output is fluent and it is empty, and the value of those messages is entirely that you sat down and wrote them. If it is discovered, and it frequently is, the damage substantially exceeds whatever awkwardness you were avoiding.

Write those yourself, badly. Badly written and sincere lands better than fluent and generated, and the recipient can tell the difference more often than people expect.

The honest test before using it

Would I be comfortable if the recipient knew? If not, that is information about whether to do it.

Can I verify what it produced? If not, do not publish it.

Is my voice part of the point? If yes, edit rather than generate.

Would I sign my name to every sentence? You are about to.

Four questions, ten seconds. They sort almost every case correctly and they are more useful than any policy.

The register problem

One tendency worth naming separately, because it affects almost every use above.

Models drift toward a formal, neutral, slightly promotional register. Left alone, output reads like corporate communication regardless of what was asked for.

This flattens everything. A note to a colleague comes back sounding like a press release. A technical explanation acquires enthusiasm nobody asked for. A simple refusal becomes three paragraphs of appreciation and regret.

Counter it explicitly. Say what register you want and what you do not. "Plain, direct, no enthusiasm" works better than "casual".

Give a sample. Two paragraphs of your own writing in the target register does more than any description.

Watch for the additions. Generated text tends to add a warm opening, a hedge and a closing offer of further help. In most professional contexts all three should be cut.

Cut before sending, always. The single most reliable improvement to any generated message is deleting the first and last sentences.