At a glance: AI detectors can misread practised writing as machine-made. The answer is to show more of the real work behind the words.
AI detectors score how predictable each next word appears, so fluent, orderly prose can be flagged as machine-made.
For an experienced independent professional, the risk is that a false flag reads as a shortcut.
The answer is not to write worse. It is to put more of the real work on the page: the people, the dates, the constraints and the decisions behind the prose.
The models generate the options. The calls are mine.
A chapter written before the web had a browser
In September 2026, Richard David Hames published an essay called Guilty Until Proven Human.
His team at the Asian Foresight Institute had built a model trained on everything he had written and every piece of case work he had produced over four decades. They then fed it a full chapter from his first book, The Management Myth, published in 1994.
The model declared with total confidence that it had written the chapter.
A second, widely available model judged the same chapter to be 65 per cent "produced with the aid of AI".
Hames had written the book before the web had a browser. He was 81 when he described the experience. It was an experienced writer having his own earlier work attributed to a machine that did not exist when he wrote it.
That makes his story uncomfortably relevant if you are over 50 and building an independent practice around work you have spent decades learning to do well. Your sentences may be orderly because you have written thousands of them. Your explanations may be clear because you know which details matter. Your tone may be steady because you have handled the same difficult question in factories, boardrooms, clinics, classrooms or community organisations.
Those qualities can now attract suspicion.
Hames also complicates the common advice to leave a few rough edges so that writing appears human. He composes by dictation because arthritis makes typing difficult. Dictation leaves homophone errors and other verbal slips, which he calls "voiceprints from a mouth, not fingerprints from a keyboard".
Even those imperfections offer no reliable certificate of human authorship. Messiness proves little. A machine can produce errors, while a person can produce immaculate prose.
Hames puts the problem like this:
Richard David Hames
Competence itself has become evidence against the competent.
Why polished prose can trigger suspicion
Detectors score "perplexity", or how surprising each next word appears. So plain, orderly prose can look machine-made, while more literary language can look more human.
A 2023 Stanford study by Weixin Liang and colleagues showed how serious that bias could become. The researchers ran 91 TOEFL (Test of English as a foreign language) essays written by non-native English speakers and 88 essays written by US eighth-grade students through seven commercial AI detectors.
The non-native English essays were misclassified as AI-generated at an average false-positive rate of 61.3 per cent. At least one detector flagged 97.8 per cent of them. The essays by native English speakers were identified with near-perfect accuracy.
The researchers found that more literary language was classified as more human. Clear, predictable word choices carried a higher risk of being read as artificial.
The study concerned a particular set of essays and detectors. It did not establish that every detector fails in every setting. It showed that fluent prose can be misread at a disturbing rate, and that the burden does not fall evenly.
OpenAI supplied another warning in July 2023 when it withdrew its own AI text classifier because of its "low rate of accuracy". In OpenAI's evaluation, the tool correctly identified 26 per cent of AI-written text and falsely flagged human-written text as AI 9 per cent of the time.
At this point, I see no reason to let an unreliable software tool tell me whether my own sentences are human. A detector score is just a software guess. It knows nothing about where you were sitting, or the thirty years you spent learning your trade.
The risk for a practised professional
If you have spent 30 or 40 years learning how to explain complicated matters simply, your writing may display exactly the regularity a detector associates with a machine.
Years of practice teaches you to cut what I call the throat-clearing: the glib wording, the puffery.
"In today's fast-paced world, it is important to note that communication matters" becomes
"She answered my question before I finished asking it."
You learn to keep a paragraph on one subject, and to know when a technical term helps and when it's just showing off.
That polished output has a history. It came from proposals that failed, reports returned with questions, meetings where an explanation landed badly, later attempts that worked better, and a schooling where a computer was still science fiction.
A detector sees the finished sentences. It cannot see the years that shaped them. Nor can the client who gets a warning that your work looks machine-made. To them it reads as a shortcut, and the one thing your reputation cannot survive is the suspicion that you took one.
Adapting your writing to sound worse throws away the benefit of those years. Artificial stumbles, random fragments and forced slang would make the work harder to read. They could also make you sound as though they were performing a version of casualness for approval.
I would rather preserve the clarity and make its origins more visible.
Put the work behind the words
A polished general statement is easy to detach from its author.
Consider a sentence such as: "Successful change depends on clear communication and stakeholder engagement."
It's tidy, familiar, over used and almost empty. No reader can tell which change, whose support mattered, what was communicated, or what happened when someone disagreed.
The real work begins when the account includes the situation.
- Perhaps the decision concerned closing a regional office in 2008.
- Perhaps two long-serving supervisors opposed the proposed timetable because it clashed with an equipment relocation.
- Perhaps the original plan allowed six weeks, and someone extended it to ten after discovering that customer records had to remain accessible during the move.
Now the reader can see the decision being made. They can assess the constraint, the trade-off and the result. The writing carries evidence from a particular piece of work rather than a general claim that could sit on any website.
That level of detail does more than signal human involvement. It makes the material useful.
I can explain what I mean by "experience" through the choices a person made.
- Which warning did they take seriously?
- Which familiar method did they set aside?
- What did they notice because they had seen a similar failure ten years earlier?
- What changed after the first attempt produced the wrong result?
I can treat "good judgement" with the same care. The phrase earns its rightful place when the reader can see the actual call: the consultant who refused a job because the client could not name what success would look like, or the one who told a client the problem was not the pricing but two partners who no longer trusted each other.
The point is showing what was happening, what you decided to do, and what happened next.
Specificity needs substance
Specific detail can also be fabricated, and AI does it constantly. It will invent a name, a date, a legal precedent, and state it with total confidence. So a scattering of names and dates gives you no protection. The stronger material is connected to a real working history and can be checked where appropriate.
That history lives in the details: the day a policy changed and you reworked the advice, the report that pulled the project in a new direction, the objection that exposed a flaw you had missed, and the early result that looked promising before it failed.
Confidential work needs care. Names can be withheld, commercial figures can be expressed as ranges, and identifying details can be removed. Enough of the reasoning can still remain for a reader to understand what made the situation difficult and why a particular course was chosen.
Sources help too. A link to the regulation in force at the time, a dated quotation from a report, or a reference to an original dataset gives the reader somewhere to go next. It also separates a considered account from polished commentary floating free of evidence.
None of this requires turning every article into a case study. One concrete episode may be enough. A short account of a wrong turn can carry more weight than several paragraphs of smooth advice, especially when it shows what information was missing and how the next decision improved.
How I work with AI
I use AI heavily. Most days I feed several models the things I have been reading, and they hand back options and drafts. Then I do the only part that matters. I keep some of it, throw most of it away, and decide what goes out under my name.
I am the one who knows which highlight is worth keeping, which argument survives contact with more than twenty years of client work, and which sentence sounds like me and which sounds like a machine trying to sound like me, or just sounds silly. The models generate the options. The calls are mine.
Those calls are the trail I have been telling you to put on the page. Not the model that generated the options, but the decisions you made between them: what you kept, what you cut, what you answered for.
Keep the voice you earned
Hames's experience does not prove that detection can never contribute useful information. The Stanford study does not cover every kind of professional writing, and OpenAI's withdrawn classifier was one tool. These sources support a narrower and more useful conclusion: detector scores can produce false accusations, particularly when prose is fluent, orderly and predictable.
That is enough reason to resist reshaping a mature voice around the preferences of an opaque scoring system.
If a platform, publisher or prospective client questions a piece, the working trail may matter more than stylistic contortions.
Drafts, notes, interview recordings, source documents, version histories and dated research can show how the piece developed. Keeping that material also makes revision easier because the claims can be traced back to their foundations.
The deeper answer belongs on the page itself. A reader should encounter more than a polished conclusion. They should be able to see the actual problem, the available choices, the people affected, the decision taken and the result that followed.
I'm not going to write poorly so a detector lets me through. I'll put more of the actual work on the page: the names, the dates, the constraints and the decisions I made, so it could only have come from me.
Related reading inside Wisepreneurs
How experienced professionals avoid the AI commoditisation trap
Every Tuesday I send The Wisepreneur, a short letter for experienced professionals building independent practice: one usable idea a week. Sign up at wisepreneurs.com.au/newsletter
When you want a second pair of eyes on your own practice, the Positioning and Practice Audit is how I work with people: wisepreneurs.com.au/marketing-partnership
Frequently asked questions
Does a low detector score mean I used AI?
No. The Stanford study found a 61.3 per cent false-positive rate for non-native English essays, and at least one detector flagged 97.8 per cent of them. A detector score is a software guess, not proof of authorship.
Should I write worse to avoid detection?
No. You would throw away the clarity you spent decades earning. The better answer is to put more of the real work on the page: the people, dates, constraints and decisions behind the prose.
How do I prove my writing is human?
Show the working trail. Name the situation, the options considered, the decision taken and what happened next. Drafts, notes and dated research also help if a platform or client questions a piece.
Is using AI in my writing a problem?
Not if you stay in the loop. Use AI to generate options and drafts, then keep, cut and decide yourself. The trail of those decisions is the human part no machine can produce.
References
Richard David Hames, "Guilty Until Proven Human", The Hames Report, 11 September 2026.
Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou, "GPT detectors are biased against non-native English writers", Patterns 4, no. 7 (10 July 2023).
OpenAI, "New AI classifier for indicating AI-written text", 31 January 2023, updated 20 July 2023.