AI words to avoid, and what detectors really measure

The AI words to avoid are the inflated verbs and stock phrases that language models overuse, such as “delve”, “leverage,” and “in today’s digital world”. ProfileTree’s editors cut them from every draft written with AI help because they make copy harder to trust. Detector scores matter far less, since independent studies show detectors disagree widely on lightly edited text.

AI words to avoid

A short list of words turns up again and again in machine-written drafts, and readers have learned to spot them on sight. None of these words is wrong in itself. The trouble is that they sound weighty while saying very little, so copy full of them reads as if nobody stood behind it. For a company, that costs trust with the visitors you most hope to convert.

If you would prefer to hand this editing to a team, our content marketing Belfast service covers drafting with AI and a person editing every page. The sections below give you the list so you can apply it yourself first.

Inflated verbs and adjectives

The most common tells are verbs that dress up a simple action, such as “leverage”, “delve”, “harness”, “elevate” and “foster”. Adjectives work in a similar way: “robust”, “seamless”, “comprehensive”, “cutting-edge” and “transformative” all claim a quality without showing it. “Leverage your analytics” means “use your analytics”, and “a robust SEO plan” says nothing a reader can picture.

The fix is to swap the word for the plain one, or better, for the specific detail it was hiding. A “seamless checkout” becomes “a checkout that takes two steps and saves the card for next time”, which a customer can test.

Stock transitions and sentence openers

Model output tends to open almost every paragraph with a connector, and a few favourites repeat: “Moreover”, “Furthermore”, “Additionally”, “It is worth noting” and “In conclusion”. Human writers start most paragraphs with the point itself. Deleting half the connectors in a draft usually does more for its rhythm than finding softer replacements for them.

Closing lines deserve a look too. A paragraph that ends by restating what it has just said, often opening with “Ultimately” or “Overall”, can drop that last sentence and lose nothing.

Marketing phrases that promise without proof

The third group is the advertising language that AI reaches for when a brief asks for persuasive copy: “unlock the potential of”, “in today’s digital world”, “game-changing” and “take your business to the next level”. These phrases make a claim and then move on without supporting it. A sentence with a real number or a named example persuades more, because the reader can check it.

Product copy shows the difference clearly. “Experience unparalleled quality with our innovative range” gives a buyer nothing to hold on to, while “stitched in Belfast from 400gsm cotton, and machine washable at 40 degrees” tells them what they are buying.

AI sentence patterns

Word swaps fix only part of the problem, because AI drafts also share a handful of sentence shapes. The clearest is the negation contrast, “it is not just X, it is Y”, which sets up a point nobody made in order to knock it down. Another is the three-item list built for rhythm, where the third item adds sound more than meaning.

Watch too for trailing “-ing” clauses that inflate a sentence, as in “the team rebuilt the site, reflecting its commitment to quality”. Vague attributions like “experts agree” or “many consider” belong on this list, and so do rhetorical questions that the next sentence answers. Each pattern is easy to fix once you see it: state the point directly, name the source, or cut the clause.

Rhythm is the subtler tell. AI tends to write sentences of similar length and shape, paragraph after paragraph, while people mix a short sentence with a long one without thinking about it. Our guide to making AI drafts read naturally works through a full draft with these patterns in mind.

How detectors work

Most detectors score a passage on how predictable it is to a language model. The first measure is perplexity, which means how likely each next word is given what came before it. Low perplexity suggests machine text, because a model tends to pick the likeliest word. The second is burstiness, meaning how much sentence length and structure vary across a passage, and people’s writing usually varies more.

Some newer tools add a classifier, a model trained on large sets of text already labelled as human or AI, and give its verdict alongside or in place of these measures. Either way, the software is estimating a probability from the text alone. It cannot see who wrote the words or how the draft was produced.

That limit explains most false flags. Careful formal writing, technical writing, and text from people writing in a second language can all be predictable on the surface without any AI involvement. A detector reading those texts has no way to tell disciplined human prose from a model’s output.

Detector accuracy

Detector accuracy depends heavily on which tool runs the check and what kind of text it reads. The strongest commercial tools rarely flag fully human writing in controlled tests, but every tool tested struggles once someone has touched up a model’s output or blended it with their own. Three recent studies show the spread.

The Maryland study matters most for businesses, because light AI polishing of a human draft is the most common way AI is used in company copy. The Chicago Booth paper shows that the better commercial detectors have become far less likely to accuse a human writer outright. The 2026 study, run on writing from students of English as a foreign language, shows how quickly accuracy falls on mixed and technical text.

The practical lesson is that a single detector score is weak evidence in either direction. A clean result from one tool tells you little about what another tool would say, and a flag on something you wrote is no proof that AI wrote it. A clear editorial process, with version history showing who drafted and who edited each page, is a stronger defence than any score.

Google’s view

Google’s published guidance says that using AI or automation to produce content is not against its guidelines when the content is helpful and written for people. What breaks its spam policies is using automation mainly to manipulate search rankings. Its ranking systems look at the quality of a page, and they do not check which tool wrote the first draft.

For business owners, that puts the risk in the quality of the page and away from the software that drafted it. Thin, repetitive pages written to fill a keyword list carry risk, however they were made. Our article on the SEO risks of AI-written pages covers where that line sits in more detail.

Editing process

A repeatable edit turns an AI draft into copy a business can stand behind. Running the same five checks on every page works better than catching problems on the drafts that happen to look wrong.

1. Check every fact and figure against a named source, and cut any you cannot trace.
2. Run the draft against your list of words to avoid, and replace each hit with a plain word or a specific detail.
3. Read the opening paragraph and the first section aloud, and rewrite any sentence a person would not say.
4. Vary sentence length within each paragraph, without breaking sentences into fragments.
5. Check the tone against your brand voice, and add one example from your own work that no other business could write.

ProfileTree is a Belfast web design and digital marketing agency, founded in 2011 and based at the McSweeney Centre, and this is the edit our writers run on client drafts that started with AI. The word list in this article comes from the banned list in our own writing guide, which our writers check every new and updated article against.

“In ProfileTree’s own content and in the content marketing we do for Northern Ireland businesses, the AI does the research, the structure and the first draft, and a person who knows the business decides what is true and what the customer needs. Take the person out, and the result reads like everyone else’s,” says Ciaran Connolly, founder of ProfileTree.

Being open about that process also helps. A business that explains how AI fits into its content, with an editor checking and signing off each page, gives clients a clearer answer than a flat denial that AI was ever involved.

Next steps

The next step depends on whether you want to do the editing in-house or hand it over. If your team publishes a large volume of AI-drafted content and needs it edited to a consistent standard, our content marketing service takes on the drafting and the human edit, and it is quoted on request.

Teams that want to build the skill themselves can learn these editing checks through our digital training, which starts from £350 plus VAT and is built around your own pages. Businesses looking at AI more widely across their marketing can see how we use it in drafting and research on our AI in marketing page.

FAQs