Prompts & Writing
When Not to Use AI for Writing
By Jim Vernon, Editor, AI Intelligence International · Published 12 February 2026 · Reviewed against our editorial standards · About the author
Enthusiasm for drafting assistance has produced a lot of writing that should not exist and some that actively damages the writer. Knowing where the tool is wrong is as valuable as knowing where it is right.
Seven categories account for most of the harm.
Key takeaways
- Anything that must be your own thinking: Strategy documents, positions you will defend, performance feedback, and academic work exist to demonstrate that you thought.
- Condolence, apology and personal messages: Generated sympathy reads as sympathy-shaped, and being caught sending it is a lasting reputational cost.
- High-stakes factual content without verification capacity: Medical, legal, financial and safety content where you cannot verify every claim is a category to avoid entirely, not to use carefully.
- Anything with confidential or personal data: Client information, employee records, unreleased financials and health data should not enter a tool without an appropriate agreement and an internal policy that permits it.
Anything that must be your own thinking
Strategy documents, positions you will defend, performance feedback, and academic work exist to demonstrate that you thought. Outsourcing the thinking defeats the purpose and is usually visible in the discussion afterwards.
Use the model to critique your draft instead. Critique preserves authorship and improves the result.
The test: will you have to defend every sentence live? If yes, write it yourself.
Condolence, apology and personal messages
Generated sympathy reads as sympathy-shaped, and being caught sending it is a lasting reputational cost.
Short and awkward in your own words beats fluent and hollow, every time.
The same applies to apologies, where the fluency itself reads as evasion.
High-stakes factual content without verification capacity
Medical, legal, financial and safety content where you cannot verify every claim is a category to avoid entirely, not to use carefully.
If you lack the expertise to catch a plausible error, you cannot supervise the output, and unsupervised output in these domains harms people.
Commission an expert or publish nothing.
Anything with confidential or personal data
Client information, employee records, unreleased financials and health data should not enter a tool without an appropriate agreement and an internal policy that permits it.
The convenience is real and the exposure is permanent. Redact first, or use an approved environment.
Volume content nobody asked for
The ability to produce fifty articles a month is not a reason to. Thin content damages search performance, dilutes brand, and consumes review capacity that would be better spent on three good pieces.
If a piece has no audience and no unique material, cheap production does not make it worth publishing.
Work where the process is the product
Learning, first drafts in a craft you are developing, and writing that clarifies your own thinking all lose their value when the output arrives without the struggle.
Use assistance after your draft exists, not before. Order matters more than abstinence.
The last category is anything you would be embarrassed to disclose. That instinct is usually correct, and disclosure is the simplest ethical test available.
Cases where the cost of being wrong is asymmetric
Legal notices, safety instructions, medical or financial guidance, regulatory filings and anything contractual. A plausible-sounding error in these does not cost an edit; it costs a liability, and fluency makes the error harder to spot.
The same applies to anything published under a professional licence. The signature carries the responsibility regardless of how the text was produced, and 'the model wrote it' is not a defence anyone has successfully used.
Where these documents must be produced at volume, the workable pattern is templated language reviewed by a qualified person, with the model restricted to filling defined variables rather than composing prose.
Cases where the writing is the relationship
Condolence notes, apologies for something you did, personal recommendations, feedback to someone who works for you, and anything where the point is that you spent the time. Recipients can usually tell, and the damage when they can outweighs the minutes saved.
There is a narrower legitimate use here: asking a model to check whether your own draft reads as defensive or unclear. That keeps the words yours and borrows only the second opinion.
The test is whether the reader would feel differently knowing how it was produced. If yes, write it yourself.
Cases where it is simply slower
Very short text you already know how to write — a two-line reply, a subject line, a commit message. Prompting, reading and editing costs more than typing thirty words.
Highly idiosyncratic internal writing that depends on context the model has no access to: team history, ongoing negotiations, unwritten politics. Supplying enough background takes longer than writing the thing.
And first drafts of arguments you have not thought through yet. Writing is how you find out what you think; outsourcing it produces a fluent version of an argument nobody has actually made.
Frequently asked questions
Is using AI for a cover letter dishonest?
Not if the content is true and the judgement is yours. It becomes a problem when the letter claims qualities the process itself contradicts.
What about academic writing?
Follow the stated policy, which increasingly permits assisted editing and prohibits generated content. Undisclosed use is a conduct risk, not just a quality one.
Can I use it for performance reviews?
For structure and clarity, yes. For the assessment itself, no — the judgement is the entire content of the document.
How do I decide in a grey case?
Ask whether you would be comfortable stating exactly how the text was produced. If not, change the process rather than the disclosure.
Is it ever fine for sensitive messages?
For structure and tone review of your own words, yes. For composing the sentiment on your behalf, the risk to trust rarely justifies the time saved.
How do I set a team rule?
List three or four document types that always require human authorship and review, and leave everything else to judgement. Long prohibition lists get ignored wholesale.