The AI detection false economy: write in Markdown, don't fight a flawed tool

AI detection has become a multi-million-dollar industry, and in the past two weeks its contradictions moved from a niche concern into mainstream debate. The New York Times ran a first-person test of a popular detector on August 13 and called the experience "empowering" — the case for the tools. The Verge wrote that AI detectors are creating a new era of distrust — the case against. LinkedIn has been labeled the most AI-saturated platform on the internet, and its editor in chief flatly dismissed the detection industry in the process. For a person who writes and publishes content to LinkedIn, X, Medium, or Substack, this is not abstract: a number from a black box can now cost you work. Here is what is actually going on, and why writing in Markdown is the cleanest way out of the trap.

What the "false economy" actually is

Betts's experiment is the clearest description of the problem. He took articles he knew were entirely human — one from 2014, one from 2019 published years before ChatGPT existed, one from 2021, and a fresh 2026 piece — and ran them through multiple well-known AI detectors. The results: one tool scored his 2026 article 100% AI-generated, another 78%, a third 42%, and a fourth said a human wrote it. His 2019 technical SEO piece, written before generative AI existed, came back 35% AI from one tool, 16% from another, and one claimed it had "found traces of GPT" — a product that did not exist yet. A 2021 article split 86% human versus 76% AI. Same words, opposite verdicts.

Betts calls the industry built around these scores a false economy: money flowing into a verification layer that cannot reliably verify anything. The research backs him up. Studies that pushed historically human-written documents, including newspaper opinion pieces, through leading detectors found many falsely flagged as machine-written. Stanford researchers found the problem was worse for non-native English speakers — some detectors misclassified large portions of their genuine work as AI, which is an uncomfortable thought for any business with international writers. And a 2026 ScienceDirect study found the opposite failure too: small human edits let genuinely AI-generated content bypass detection entirely. The tools wrongly accuse human writers while missing content they are supposed to catch — and a few human tweaks beat them anyway.

The cost: FOW, fear of writing

The real damage is not the software; it is the effect on writers. Betts names it FOW — the fear of writing. Writers second-guess prose they know is good, worried a tool will decide it "looks AI." Agencies panic-run scans on client work and treat the percentage as gospel. Freelancers lose assignments because someone ran a detector and, wrongly, got a false positive. Clients pay per scan, per word, per seat for the privilege of being told whether they are human — and then pay again for a "humanizer" that promises to fix the very thing the detector flagged.

The humanizer upsell makes the economics explicit. In Betts's write-up, a detector scored his fully human 2014 article as 100% human, then immediately offered him a paid upgrade to "humanize" it. Humanize what? The human? That is the toll nobody needs: an industry charging writers to fix text that was human to begin with. The same pattern now reaches publishing. This week, editor Mary Kole wrote on Jane Friedman's site that she unknowingly accepted two AI-assisted novels because clients did not disclose their chatbot use, and is rewriting her client agreement to protect against AI "scope creep." When even polished, professional prose triggers suspicion, everyone downstream pays.

The case for detection (it is not nothing)

To be fair to the tools, they are responding to a real problem. AI content has genuinely flooded public feeds — LinkedIn became the most AI-saturated platform by several studies, and platform-level measures like the "seems like AI slop" report button and Substack's partnership with Pangram to flag AI-generated posts are attempts to preserve the value of human writing. A detector that works "well enough" can keep an office or classroom from drowning in low-effort generated text. Even the skeptical writers concede the goal: nobody serious wants to defend lazy, low-quality AI output — Betts himself says AI "has earned its place" in every modern marketing workflow. The dispute is not over whether bad content exists. It is over whether point-in-time detector scores are a trustworthy way to decide who wrote what. On that, the evidence is damning.

Why institutions are walking away

The most telling sign is how quickly institutions that would benefit most from detection are abandoning it. Indiana University's Kelley School of Business bans AI detection tools outright in its faculty AI playbook, calling them unreliable and telling staff not to upload student work to them at all. Universities around the world have dropped Turnitin-style detectors over false-accusation fears. If a leading business school will not trust these scores on student essays, the reasoning goes, why would a brand trust them on marketing copy?

Search engines reached the same conclusion from a different direction. Google's guidance has been steady for years: reward helpful, reliable content, however it was produced, and enforce against low-quality content at scale rather than trying to decide whether each sentence came from a person or a model. LinkedIn's editor in chief, Dan Roth, delivered perhaps the clearest statement of the problem: "A big share of LinkedIn's content is AI-assisted, but nobody can know exactly how much, including LinkedIn, and anyone claiming they do is selling a detection tool."

What actually proves you wrote it

Here is the practical question for any writer: if you cannot trust the detectors, what evidence can you offer that your work is yours? Betts's answer is telling — he points to humans and the things humans naturally create: draft history, version control, editorial review, a writer's body of work. Each of those carries more context than a percentage spat out of a black box. That is the whole ballgame, and it is exactly what a Markdown workflow gives you for free.

A Markdown file is the purest form of evidence-of-authorship available. It is plain text: every heading, paragraph, list, and link is literal, readable, verifiable characters — no hidden formatting layer, no invisible output, nothing you cannot inspect. And when you version that file in Git — even a simple personal repo — you accumulate a draft history and an edit trail that documents how the writing developed over time. That is the "version control" and "draft history" provenance even the skeptics recommend, and it cannot be faked by running text through a detector. You are not trying to pass a score; you are pointing at the thing itself.

A Markdown workflow that stays out of the arms race

The trick is to structure your writing so that the clean, human artifact is the source of truth from the start — not a post-hoc scrubbing exercise. That flips the detector problem on its head. If the words are yours from the beginning, there is nothing to hide and nothing to "humanize." A simple, repeatable routine:

1. Draft in Markdown. Open your editor and write your own headings, sentences, and examples in plain .md files. These files are your authorship record.
2. Use AI as a think-partner, not a co-writer. Brainstorm angles, generate an outline, or pressure-test a draft — but keep the model's output as input to your thinking, not as sentences merged into your file. Your final text is yours.
3. Version it in Git. Commit as you go. The commit log becomes an honest, human-verifiable edit trail that a detector score can never match.
4. Convert at publish time. Turn that clean Markdown into the rich text your platform wants — LinkedIn, X, Medium, Substack — using a converter. Nothing about the source changes.

Here is that idea in practice. You keep your draft as Markdown, and paste it into a converter at the end:

# Why I stopped fighting AI detectors

Three detectors read my post three ways.
I stopped worrying about the scores and
started keeping the source — my writing,
in Markdown, in Git. It is my own draft
history, plain and verifiable.

## What I do instead

1. Draft in Markdown
2. Keep AI for outlines and review
3. Version every commit
4. Convert to rich text at publish time

And the formatted result you paste into LinkedIn, where the source is now styled with headings and structure rather than pasted as a flat wall of text:

Why I stopped fighting AI detectors

Three detectors read my post three ways.
I stopped worrying about the scores and
started keeping the source — my writing,
in Markdown, in Git. It is my own draft
history, plain and verifiable.

What I do instead
- Draft in Markdown
- Keep AI for outlines and review
- Version every commit
- Convert to rich text at publish time

The point is subtle but important: you are not optimizing for a detector. You are removing yourself from the detector economy entirely. You do not care whether a tool thinks you sound like a model, because your authorship is established by the artifact you control — the versioned Markdown source — and not by some vendor's percentage.

Why md2rich fits here

The one missing piece in a clean-provenance Markdown workflow is getting from plain text to the styled rich text platforms actually want. That is where md2rich comes in. It is a client-side Markdown-to-rich-text converter: you paste your Markdown, it produces clean, formatted text — headings, bold, lists, links — ready to paste into X Articles, LinkedIn, Medium, or Notion. Because the conversion happens entirely in your browser, nothing is uploaded anywhere. Your draft never leaves your machine, which means there is nothing to intercept, store, or hand to a third party to process. The privacy angle matters for the same reason provenance does: both are about you staying in control of your own words.

Your Markdown file remains your single source of truth, and the rich text you publish is a rendering of it — one you produced yourself at publish time. You get the format the platform rewards and the provenance the detector debate keeps reminding you to protect, from the same written source. That is the write-once, publish-everywhere model, and it is the most honest answer to the false economy.

Frequently asked questions

Bottom line

The AI detection industry is a false economy, and the evidence this month is overwhelming: the same human writing scores 100% AI and 100% human depending on which tool you run, honest non-native writers are falsely accused at higher rates, and institutions from a top business school to the world's largest professional network are declining to trust the numbers. But behind all the noise is a real ask that keeps coming back: prove your words are yours. The tools cannot answer that question responsibly, and neither can a percentage.

Your drafting format can. When you write in plain Markdown, keep it versioned, and convert it to rich text only at publish time, your authorship is established by an artifact you control instead of a score you cannot predict. You leave the fear economy behind because you were never trying to pass its test. Write your next post as Markdown and see the difference. Try md2rich — paste your Markdown, get clean rich text for LinkedIn or X, and know that nothing was ever uploaded, because the conversion happens in your browser.