AI-generated copy reads cleanly on a first look and falls apart on a second. A line-by-line pass means going through the draft sentence by sentence and asking, of each one, whether it earns its place. That is different from a general tidy-up, and it takes longer, but it is the only way to catch the faults AI drafts share.
AI reaches for the heavier word
Ask an AI tool to write about a product and it will usually offer you "seamless", "robust" or "transformative" somewhere in the first paragraph, whether or not anything in the brief supports the claim. The fix is not to delete every adjective. It is to check each one against the fact behind it: if the product does one thing reliably, say that it does one thing reliably, and save the stronger word for the rare case that has actually earned it. The same check runs the other way. AI drafts also lean on false modesty, calling something "surprisingly simple" or "nothing fancy", which asks the reader to be impressed by the absence of complexity.
AI repeats a point instead of making it once
A pattern worth watching for: three sentences in a row that all say the same thing in slightly different words, usually because the model is padding towards a length target. Read each paragraph and ask what new information each sentence adds. If the answer is none, cut the sentence. The same applies to lists: AI tools default to bullet points even where the content is one connected idea that reads better as a sentence, so check whether the items on a list are genuinely separate steps or criteria, or whether they are one thought that has been chopped into fragments for the look of structure.
AI invents detail that was never in the brief
The more serious problem sits below the sentence level. AI does not know your product, so where a draft needs a specific, a figure, a feature, a claim about what a customer gets, the model will supply something that reads as confident and specific and may simply be wrong. This will not always announce itself: a plausible-sounding number sits in a sentence exactly where a true one would sit, and a reader skimming for tone will not notice the difference. Every specific claim in an AI draft needs checking against what is actually true of the product. This is the edit that most affects the reader's trust and the one most likely to be skipped, because the sentence itself reads perfectly well.
Read every sentence aloud before you publish it
The test that catches what the others miss is simple: read the sentence aloud and ask whether it is something a person would actually say to the reader they have in mind. AI drafts often pass every mechanical check, correct grammar, plausible claims, reasonable length, and still contain sentences nobody would say out loud, because they were built to a pattern. Where a sentence fails that test, rewrite it as plain speech first and tidy the register afterwards. The guidance on what AI is actually good and bad at covers where this drafting pattern comes from; the glossary is useful if a term in your draft needs checking against its proper definition before you decide whether it has earned its place.