- Seven ways to use AI without getting flagged as AI, and every one of them is about provenance rather than evasion. No detector-beating tricks here, on purpose.
- A detection score is a probability, not a finding. What settles the question is a record of how the work got made, and that record is something you can build on purpose instead of reconstructing under pressure.
- Detectors measure how predictable your word choices are, which means careful writing in a second language sits in the same band as machine output.
- Scan your own unassisted writing once, write down the number and the date, and treat it as a baseline rather than a target to chase.
- The build item is the evidence pack that assembles itself, so every deliverable ships with its own paper trail instead of a promise.
Using AI without getting flagged as AI is a search phrase with two very different people behind it. One wants to pass off generated work as their own and would like a trick. The other does their own thinking, used a tool somewhere in the middle, and is now worried a score in a dashboard will override the truth. This piece is for the second person and has nothing to offer the first. Every idea below makes the real story of your work easier to prove, which is the only defense that holds when somebody pushes back.
Here is why that distinction matters more in 2026 than it did two years ago. Detection moved out of classrooms and into the places adults get paid: client contracts, platform labels, employer policies. Substack now ships a reader-facing scan that labels posts and comments as human or AI-assisted, which TechCrunch covered when it rolled out in July. Meanwhile the detectors are still guessing. Turnitin, which has every incentive to make its numbers look good, puts the sentence-level false positive rate of its own tool at around four percent, and Stanford HAI's summary of seven detectors run against 91 TOEFL essays found 61 percent of essays by non-native English writers called machine-written on average, against nearly zero for a native-speaker control group. The tool judging you is wrong in a patterned way, and the pattern is not in your favor if you write plainly.
You cannot fix a classifier. You can make your own account of the work impossible to doubt, which is a different and better goal. Seven ideas, each with something you can set up this afternoon.
1. Work Where the Draft Trail Records Itself
The strongest evidence that you wrote something is a timeline showing it getting written. Not a claim, not a screenshot, a sequence: the blank page, the bad first paragraph, the sentence you moved three times, the section you cut on Thursday and put back on Friday. No generated draft has that shape, because generated text arrives whole.
Almost nobody keeps it, because most people draft in whatever window is closest and paste the result into the real document at the end, which destroys the only record that mattered. The fix costs nothing. Draft in one place, from empty to finished, in a tool that keeps history. Google Docs does it automatically and lets you name specific versions so the milestones stay findable, up to forty per document. Word and most modern editors do the equivalent, and a folder of dated copies works fine if you write in plain text.
Dara writes marketing copy for three agencies and used to draft in a notes app, then paste finished pieces into the client's system. When one client's new review process asked her to "evidence her process," she had nothing but her word. She now starts every piece in the shared document itself, badly, on purpose. The messy middle is the asset.
Start today: Open the document you are working in right now and find its version history. In Google Docs that is File, then Version history, then Name current version. Give today's state a name like "first full draft." Then make a rule for yourself: the first keystroke of any deliverable happens in the document you will hand over, never in a scratch window.
2. Learn What the Detector Is Measuring
Most people arguing about detection have no idea what is being measured, which makes them easy to overrule. Two properties do most of the work. Predictability, often called perplexity, is how likely each next word is given the ones before it. Variation, often called burstiness, is how much your sentence lengths jump around. Machine output tends to be smooth and even on both counts. So does disciplined professional writing, and so does writing done carefully in a second language.
Knowing that changes the conversation from "I promise I wrote it" to "here is what your tool reacted to," and a mechanism is arguable in a way a denial is not. It also tells you who is most exposed: people who write in short declarative sentences because they were trained to, and people writing in a language they learned as an adult. The edition on writing when English isn't your first language covers that second group, and this is one more cost they pay that nobody mentions. One hard line, since this is the item easiest to get wrong: understanding the measurement to contest a bad result is fair, and editing your sentences until a score moves is not a writing technique. If you catch yourself adding deliberate clumsiness, stop.
Start today: Paste two or three paragraphs of your own unassisted writing into any model and ask which properties of the text a statistical AI detector would react to, and why. Ask for the explanation only, no rewrite and no suggestions. You are building vocabulary for a conversation you may have to have, not editing anything.
3. Scan Your Own Writing First, as a Baseline
Everybody runs the scan after the accusation, when the number has already become a weapon. Run it before, on work you know you wrote alone, and it becomes a measurement of how this particular tool reads you, taken at a time when nobody was arguing.
Do it on three or four pieces from different parts of your work, since a technical spec and a newsletter read very differently to a classifier. Write down the tool, the date, the percentage, and the fact that the piece was unassisted. Three lines in a text file. If the numbers come back high on writing that is entirely yours, you have not found a problem with your writing. You have found the detector's error rate as it applies specifically to you, which is far more useful to hold than a general complaint about false positives. The baseline is not a target, though. The moment you edit to lower it, you have handed your voice to a machine nobody asked to be your editor.
Start today: Pick one piece you wrote with no AI involvement at all, run it through whichever detector your client or platform uses, screenshot the result, and file it with today's date next to the piece. That one file is more useful than any argument you could make later from memory.
4. Hand Off the Provenance Pack
Everything above is a habit, and habits break exactly when you are busiest, which is also when the awkward question tends to arrive. This is the item to stop doing by hand. For every deliverable there is a small bundle of evidence that should exist by default: the draft history, the sources you consulted, which parts a tool touched, what you verified yourself, and the date it shipped. Assembled after the fact it takes an hour and comes out thin. Assembled as you go it takes no time and comes out complete.
That is a workflow, not a discipline. A standing job that snapshots the version history on delivery, pulls the source links out of the document, and files the set under the client and the date. When somebody asks how a piece got made eight months later, you send a folder instead of an explanation. Same shape as any other deliverable handoff you would automate, except the output is a record rather than a document.
Start today: Write down what your evidence pack would contain, one sentence per item, for the kind of work you deliver most. Four or five lines is plenty, and that list is the whole specification.
5. Put Something in It Only You Could Have
This is the honest center of the list, and it is not a detection tactic. It is the reason the question stops mattering. A model can produce a competent paragraph about almost anything. What it cannot produce is the number you pulled from your own system last Tuesday, the sentence a named customer said on a call, the measurement you took on site, the photograph of the thing in the state you found it, the objection you have heard eleven times from one specific industry. Original material is unfakeable by definition, and putting enough of it in a piece of work does two things at once: the statistical texture shifts, because real specifics are not predictable, and no reader who finishes it will believe a machine wrote it, whatever the dashboard says.
Marcus is an analyst at a logistics firm who kept getting told his quarterly summaries read like a template. They did, because he was summarizing a report that was itself a summary. He started each one with a single operational detail nobody else had written down: the depot that missed its window twice in one week, and the reason. Nobody has questioned the authorship since, because now there is something in there to question instead.
This is where AI is genuinely useful, just not for the writing: use it to find the specific thing already sitting in your own material, which is retrieval rather than generation. The edition on using AI without generating text is entirely about that mode.
Start today: Take the last thing you delivered and count the facts in it that could only have come from you. If the answer is zero, that is the finding. Add one to the next piece before you add anything else.
6. Claim the Space the Scanner Leaves You
Detection surfaces are starting to ship with a door for the writer, and almost nobody walks through it. Substack's scan is the clearest example: alongside the result, a reader sees whatever you put in your publication's "How I make this" statement, and you can report a detection error on a specific analysis. A place to explain your process, attached to the exact moment somebody is wondering, sitting empty on most publications.
The general move is to put your process statement where the question gets asked rather than where you would prefer to answer it. On a platform with a scan, fill in the field the platform gives you. For client work it belongs in the proposal and the delivery email, not buried in a terms document. Keep it short and specific about accountability: what tools are in your process, what you check yourself, who is responsible for the claims in the work. The earlier edition on using AI without stealing creative work covers the disclosure line for a single deliverable; this is the standing version, placed at the surface where the scan happens.
Start today: Write your process paragraph. Name your tools, name what you verify by hand, name who stands behind the facts. Then find one place it belongs today, your Substack settings, your proposal template, your site's about page, and paste it in.
7. Write the Calm Reply Before You Need It
The accusation never arrives on a good day. It arrives on a Friday afternoon, from a client you like, phrased carefully, and the honest reaction is a flash of anger that makes everything you type next worse. People lose this one on tone far more often than on evidence.
So write the reply now, while nothing is at stake. Four moves, in order: thank them for raising it directly, state plainly what your process was, offer the specific evidence you can produce, and ask what would resolve it for them. That last question ends the conversation, because the honest answer is usually "I wanted to know somebody had checked the facts," and that you can address in a sentence. Keep the draft wherever you keep reusable client language, next to your scope boilerplate and your standard document templates. And the harder case takes the same four moves, with the first one naming which part a tool touched. A denial that collapses two weeks later costs you the relationship; naming it yourself almost never does.
Start today: Draft the reply in six sentences and save it. Ask a model to read it cold, as the client, and say where it sounds defensive. Defensive is the failure mode here, not inaccuracy.
Pick One
Don't try all seven. If nothing jumped out, take number one, because it costs you nothing and it is the only one that stops being possible once the work is already delivered. Every other idea here is easier to do later. The draft trail is not.
The thread through all of it: you cannot control what a classifier says about you, and you can control whether the true story of your work is documented. Sunday we report what changed in AI, Wednesday we take one workflow apart and rebuild it, and Monday is for ideas like these. When keeping your own records turns into a weekly chore with a day attached to it, the workflow directory shows what other people handed off rather than kept doing.
Make the Record, Not the Argument
Describe what proof you would want to have on hand if somebody questioned a piece of your work. BYOBot turns that into an agent that collects it for you, every time, without being reminded.
Frequently Asked Questions
- It helps as a measurement, not as an edit. Scanning work you wrote by yourself tells you where your normal writing sits, which is worth knowing before anyone else runs a scan on you. What it should not become is a loop where you keep rewriting your own sentences until a score drops, because at that point you are writing for a classifier instead of a reader, and you have given up the one thing that was yours. Record the result with the date and move on.
- Check the contract first, because AI-use clauses are now common in freelance agreements and some of them require you to name the tools and the portions involved. If nothing in the agreement covers it, a short standing statement about how you work is still the cheaper option, since being asked after delivery is a much worse conversation than saying it up front. Disclosure is about who stands behind the claims in the work, not about confessing to a tool. The edition on using AI without stealing creative work has the wording for a single deliverable.
- Say which part, show the trail, and name what you verified yourself. An accusation is usually not really about the tool, it is a question about whether anybody checked the facts and whether the thinking is yours. A calm answer with a version history attached ends most of these conversations, and the edition on what to do after it gets something wrong covers the checking habits that make the answer easy to give. A denial that later falls apart does far more damage than the original use ever would have.
- Because the properties most detectors measure, how predictable each word choice is and how much sentence length varies, are also properties of careful writing in a second language. A Stanford HAI summary of seven detectors run against 91 TOEFL essays found an average false-positive rate above 61 percent on essays written by non-native English writers, against near zero on a native-speaker control set. If this is your situation, keeping a draft trail matters more for you than for anyone else.
