Run the same brief through an AI writing tool five times and you'll see the same five failures recur, whichever tool you use. The brief below is used for every example that follows, so you can compare the AI's output directly: a product page for a £140 waterproof hiking jacket, made by a small UK outdoor brand, sold direct to consumers, aimed at people who walk in wet weather and want something that will actually keep them dry.
AI defaults to a house style nobody chose
Left without a specified tone of voice, an AI tool reaches for the register that dominates its training data: upbeat, adjective-heavy, and interchangeable with a thousand other product pages.
Experience unparalleled protection from the elements. Our premium jacket is engineered to elevate every outdoor adventure, whatever the weather throws at you.
Nothing in that paragraph is false, but nothing in it is this jacket either. It would sit unchanged on a page for a tent, a pair of boots, or a car. A rewrite grounded in the brief reads differently:
The seams are taped, so rain does not find its way through at the stress points. Wear it on a wet Saturday walk and the inside stays dry after three hours.
The second version could only describe this jacket. That is the test to apply: if a sentence would survive being pasted onto a competitor's page unchanged, it has not earned its place.
AI pads without adding information
A language model is built to predict a plausible next sentence, so left unedited it will restate a point two or three times in slightly different vocabulary.
This jacket keeps you dry. It offers reliable waterproof protection. You can count on it to shield you from rain and wet conditions.
Three sentences, one fact. A version that adds information does more with less:
Taped seams and a storm flap over the zip keep rain out even on the walk back when the wind has turned.
Check each sentence in a draft against the one before it. If it makes the same point in different words, delete it; if it adds a fact the reader did not have, keep it.
AI invents specifics that sound plausible
AI does not know your product. Where a brief is silent on a detail, the tool will often supply one anyway, because a specific-sounding number reads more convincingly than an admission of not knowing, and you will not always notice unless you check it against the spec sheet.
Built from a 20,000mm hydrostatic head fabric with fully recycled ripstop nylon, this jacket meets the toughest waterproofing standards.
If the brief never supplied a hydrostatic head figure or confirmed the fabric was recycled, both claims are fabricated, however plausible they sound to a reader who does not know the category. The safe version says only what is known:
The outer fabric is fully seam-taped and treated to shed rain.
Treat every number, material name, and certification in AI output as unverified until you have checked it against the actual product. This is the failure most likely to embarrass you later, because the reader who does know the category will spot it immediately.
AI's tone drifts across a longer piece
Over a page of any length, an AI tool will often swing between registers sentence to sentence, because each sentence is generated with only loose reference to the ones either side of it.
This garment provides superior protection against adverse weather conditions. You'll love how cosy and snug it feels on those chilly autumn walks!
The first sentence is formal to the point of stiffness; the second is chatty enough to belong in a different piece of copy altogether. A consistent register holds the reader's attention:
The fabric sheds rain without feeling like a bin bag. Underneath, the lining stays warm even when the outside is properly wet.
Read a full draft aloud in one pass and listen for the join. If a sentence sounds like it was written by someone else, it probably was, in the sense that the model drew it from a different part of its training data.
AI makes claims it cannot substantiate
Left unchecked, an AI tool will often reach for a superlative because superlatives are common in the marketing copy it was trained on.
This is the best waterproof jacket you will ever own, guaranteed to keep you dry in any conditions.
Nobody can substantiate "the best" or "any conditions", and under the CAP Code, an objective claim needs to be capable of substantiation before it runs. A claim built from what is actually known holds up:
Taped seams and a storm-flap zip keep the inside dry through a typical wet-weather walk.
Before publishing an AI draft, go through it line by line and ask what evidence would back each claim. Where there is none, the claim comes out. The plain-English guide to ASA and CAP rules sets out how this works in more detail. This is a practical summary for copywriters, not legal advice; for definitive guidance, consult the CAP Code and the ASA's own resources.
The same AI tendency that inflates a product claim also weakens the line that asks the reader to act. Left to its own defaults, an AI tool will often produce a call to action along the lines of "Shop now" or "Discover the range", a phrase that could sit on any page selling anything. If you are weighing up the line at the bottom of this product page, the call-to-action tester takes the phrase you are considering, along with a note on what the page or email is about, and rates how strong the verb is, how specific the wording is, how the character count sits against common platform limits, and how the structure compares with patterns that tend to hold up, then it suggests three alternatives worth trying. It runs entirely in your browser, so nothing you type leaves the page or gets logged anywhere. It checks the structure of the phrase, not its grammar, and it can rate a line as strong for reasons that do not suit your particular reader, so treat the score as a way of testing your own reasoning rather than a verdict.
None of this makes AI useless for a first draft. It makes the first draft exactly that: a draft, to be checked line by line against what you actually know about the product. The line-by-line pass for editing AI copy goes through this checking process in more detail. Once you have a headline you are weighing up, the headline scorer runs entirely in your browser, so nothing you type is sent anywhere or logged. It scores the line out of 100 against word count, the presence of any power word, the balance of emotional against rational appeal, plain clarity, and length against common platform limits, and it names three specific changes worth trying. The number it gives is a way of checking your own reasoning against a stated method, not a verdict to defer to, and it can be wrong about a headline that works for reasons the method does not capture.