This summer, the pattern is visible almost everywhere: newsletters opening with a sweeping statement about how quickly everything is changing; a restaurant's Facebook post with three perfectly balanced sentences and a neat moral; a "thoughtful" LinkedIn post that could describe any company in the world. After two lines, most readers move on.
They move on because the text sounds generated. Nearly everyone now has this reflex, even without being able to say exactly what triggers it. We share it, and it shows every sign of lasting.
This is a notable position for a team that has used these tools since GPT-3 and whose founder wrote a beginner's guide to ChatGPT in 2023, in English and French. We value these tools and use them extensively. We have, however, become strict about what is published under our name.
Why does AI-sounding text lose trust?
Because readers interpret it as a lack of care. When a reader identifies generated text that no one took the time to edit, the conclusion is immediate: the author did not care enough to write it personally. And a business that does not care about its words may not care about its service either.
The judgement is sometimes unfair. The business owner may have used AI to overcome a blank page, with good intentions. The outcome is the same: trust is decided in a few seconds, and readers do not extend the benefit of the doubt.
For a small business, this is a real cost. Its main advantage over larger competitors is that it is run by people customers can know. Text that sounds like everyone else's forfeits that advantage.
Which patterns reveal generated text?
No list is complete, and the models evolve. Certain patterns, however, recur consistently in unedited AI text:
- Everything comes in threes, for rhythm, whether or not there are actually three things.
- Sentences that set up a contrast only to dismiss it, the "it's not about this, it's about that" construction.
- Paragraphs of nearly identical length, all at the same measured pace.
- A handful of elaborate words that almost no one used in conversation before 2023 and that now appear everywhere.
- A neat summary at the end that repeats what was just read.
- Long dashes, in large numbers.
- No opinion: every side considered, no one offended, nothing actually said.
Any one of these on its own is acceptable; skilled writers use threes and dashes. It is the accumulation that reveals the origin. Searching for banned words with a checklist does not solve the problem. A text can avoid every item on this list and still read as though no one wrote it, because the underlying issue is that no one decided what the text was meant to say.
Which editing habits restore a genuine voice?
We edit with a short set of rules, applied to all our writing, whether or not a machine was involved. In summary:
Remove anything rehearsed
A sentence that sounds written to impress probably was. Remove it or state the idea plainly. "We are passionate about delivering exceptional experiences" becomes "We like this work and we're good at it." Alternatively, delete it, since readers do not believe that sentence in any case.
Allow the rhythm to vary
People do not think in paragraphs of equal length. A short line, then a longer one that takes its time because the idea requires it, then perhaps a list, if the content is genuinely a list. Break the pattern wherever it feels too controlled.
Keep your opinions
State what you think and where you are uncertain. If you changed your mind about something, say so. Small contradictions are human. A text that agrees with everyone reads as though no one wrote it.
State the specific detail
General sentences are where generated text hides. "We serve the region" conveys nothing. "We go as far as Duhamel in winter and Montebello in summer" says something only your business could say. Details are also what readers remember.
Read it aloud
This step catches nearly everything. If you would not say a sentence to a customer standing in front of you, rewrite it until you would. If you stumble over it, readers will too.
Why does plain language help more than stylistic techniques?
Because clarity itself reads as human. Much sound advice on this subject predates AI by years. The Government of Canada's Canada.ca content style guide is one of the best free references we know: short sentences, common words, the most important information first, active voice. It was written for government pages, but it is excellent training for anyone writing a website.
Plain language is also the most effective safeguard against sounding generated. Unedited AI text tends to be wordy and ornate. Clear, direct writing with a specific point of view reads like a person because it usually comes from one.
How do we use AI for writing today?
As a support, never as the final author. We continue to use it to produce a rough first draft when progress stalls, to identify gaps in an argument, to check a French phrase against how people in the region actually speak, and to catch our own habitual errors. Then we rewrite. Every sentence is reviewed by a person before publication.
This is slower than pasting the output, but much faster than writing from nothing. The result reflects our own voice, which is the only reason anyone reads it.
For teams beginning to use these tools, AI training focused on editing rather than prompting delivers results faster than any other approach we know. For businesses where writing is not a strength, we recommend engaging a professional for bilingual copywriting who knows the region and reviews every draft before sending it.
In either case, read the text aloud first.