The objections worth answering, including the ones that go against the essay.
Is it true that most of what I read online was written by a machine?
No, and the essay does not say so. It depends heavily on where you are reading. On LinkedIn, roughly two in five long-form posts appear to be written entirely by a machine. Across newly published web articles, about half. On Reddit, somewhere between 2.5% and 11.6% depending on who counted and whether you mean top-level posts or replies — replies are about 98% human. Medium and Quora sit near two-fifths, Substack around a fifth.
The one place a majority claim holds up reasonably well is LinkedIn long-form, and even there the phrasing that survives scrutiny is “roughly half, and rising” rather than “most”.
Two firms give completely different figures for LinkedIn. Why should I believe either?
Because once you line up what they were counting, they agree. One reported 81% and one reported just over 40%, which looks like a flat contradiction. The first counted posts of 100 words or more and asked whether AI was likely involved at all; the second counted posts of 250 words or more and asked whether the post was written entirely by a machine. Apply the stricter test to the first firm’s own data and it lands at 41%.
That is worth more than a disagreement would have been. It means the estimate is firmer than the spread suggests — and that almost nobody quoting these numbers says which question was asked.
What about the figure that 90% of online content will be synthetic by 2026?
Do not use it. It comes from a Europol report of 2022, whose own sentence says “experts estimate” and whose footnote leads to a single trade paperback published in 2020. No measurement was performed. The claim in the book was about video. Anyone who checks the footnote can take apart an argument that leans on it.
What about Facebook, Instagram and WhatsApp? They are missing from the table.
Deliberately. Every platform in that table is one an outsider can read; Meta’s are not. Every platform in that table is one an outsider can read. Meta’s are not, so the table is really a map of what can be crawled.
For Facebook and Instagram, no measurement of the machine-written share has been published that we could find. The nearest study touches only Facebook — DiResta and Goldstein, peer-reviewed in 2024 — and is an investigation of 125 Pages found by hand, and its authors state that those Pages “are not necessarily reflective of how unlabeled AI-generated images are used on Facebook as a whole”. It did find one unlabelled AI-generated image among the ten most-viewed Facebook posts of Q3 2023, with 40 million views. That is a real finding about reach. It is not a percentage, and any percentage citing it was invented downstream.
WhatsApp cannot be measured at all. End-to-end encryption means nobody outside the conversation can survey the contents — no researcher, no regulator, no third party. The platform could only do it by inspecting text on the handset before it is encrypted, which is what the encryption exists to prevent. Only public groups a researcher has joined can be studied, and that is a self-selected sample. If you are shown a figure for how much of WhatsApp is AI-written, it could not have been produced.
Does Facebook not label AI content? I have seen the badge.
For images, video and audio — yes, since 2024. For text, no. Every scope-defining statement Meta has published names image, video and audio; generated text is not among them. So the badge you have seen does not, and was never designed to, tell you whether a post was written by a machine.
Meta also publishes no figure for how much AI content is on its platforms. Its enforcement reporting covers twenty-six policy categories across Facebook and Instagram, none of them synthetic media. The one number it has released — 590,000 — counts requests its own image generator refused to make images of named politicians around the 2024 elections, which is a measure of Meta declining to make something, not of what is circulating.
Does the EU AI Act not require disclosure now?
It requires two different things of two different parties, and the difference is the point.
Whoever built the model must, under Article 50(2), mark synthetic “audio, image, video or text” so that it is machine-readable and detectable. Text is covered.
Whoever runs the platform must, under Article 50(4), disclose deep fakes — which the Act defines as “image, audio or video content”. Text reaches a platform’s duty only where it is “published with the purpose of informing the public on matters of public interest”, and not even then if it “has undergone a process of human review or editorial control”. The Digital Services Act names “a generated or manipulated image, audio or video” among the risk-mitigation measures a very large platform may take. Text is not among them.
Article 50 has applied since 2 August 2026; the DSA obligations have been in force for longer. Text is named at the point of manufacture, and not named at the point of carriage — the same division this essay describes in the watermarking section.
Do these limits change the figures at the top?
It bounds them. The figures describe the open, crawlable internet. They are silent about the everyday writing people do to one another, on apps Meta reports 3.60 billion people opening daily, and on the encrypted surfaces they are silent permanently — not until detection improves, because detection is not what is missing there. Access is.
Every instrument in the piece except one requires somebody to be able to inspect the text. Inside end-to-end encryption nobody can, and nobody should be able to. What still works is a person saying what they did.
Are these figures not just detector output? Detectors are unreliable.
Mostly yes, and that is stated throughout. The exceptions matter. The strongest measurement in the field — the excess-vocabulary work across 15.1 million biomedical abstracts — used no detector at all. It counted how often particular words appeared and compared that against how often they should have appeared. Nothing was classified.
Where detectors were used, the essay names who ran them and on what. Three of the table rows come from one vendor measuring posts in the feeds of people who installed its own browser extension, which is not a random sample of anybody, and that firm’s chief executive says so himself.
If I see four or five of these signs in one post, does that mean something?
Less than it feels like. Signs that share a cause do not add up — five observations of one habit are one observation. The essay’s own worked example is a person writing at length, formally, repetitively, refusing to concede and sounding annoyed. That looks like five findings. It is one person arguing hard.
The more useful question is whether a single ordinary explanation would cover the whole cluster. A house style guide prescribes vocabulary, register and post template from one document — three apparent findings, one cause. So do a scheduling tool, a translation workflow, an employer’s template and accessibility guidance. Go looking for that explanation first. If one fits, you have found it.
So what is a marker actually good for?
Reading more carefully. Check a claim before repeating it. Look for the source behind a confident summary. Notice the difference between sounding authoritative and demonstrating expertise. That is all it licenses, and it is genuinely useful.
What it cannot do is tell you who you are talking to.
Are em-dashes a sign of AI writing?
Not usefully. Across twelve models the rate ranges from 10.6 per thousand words down to zero — Llama 3.1 uses none at all. The human baseline in the same study was 3.2, with individual human writers running from 0.3 to 17. There is no AI em-dash rate; there are model rates, and the human range covers most of them.
This essay runs at 10.2 per thousand words, which is above the model that drafted it. Make of that what you like — the point is that you cannot tell from the page which explanation applies.
What about hidden characters?
Almost entirely folklore as an AI marker. No major provider hides characters in text as a watermark, and Anthropic states in writing that Claude does not: “Nothing is added to the text and there are no hidden characters.” The reason is structural — any character-based mark dies to one find-and-replace, so nobody serious designs one.
Real invisible characters do turn up. A LinkedIn post examined during this research carried ten invisible characters interleaved with ten visible ones in its timestamp. That was the platform’s anti-scraping obfuscation, inside a post written by a person. Anyone running the popular hidden-character check would have got a confident hit on a human.
Which marker is actually the strongest?
Structural rather than lexical: machine prose runs noun-heavy and informationally dense, leaning on nominalisation. It is the dearest to remove, because it is a property of how the writing is built rather than of which words were chosen. It is also the one nobody has tested against a writer actively trying to remove it, which is where the claim stops.
If AI companies watermark their output, why can’t I just check?
Because the marks are readable only by the company that applied them. European law requires machine-generated text to be marked in a machine-readable form. It does not require anybody but the provider to be able to read the mark.
Google’s detector exists and is a waitlist, for journalists and researchers, covering Google’s own content. Anthropic has announced a detection interface and not published one. OpenAI built a tool, measured it internally at 99.9%, and did not ship it. The mark exists; the key does not travel with it.
Won’t that be fixed?
Possibly, and not by the mechanism people expect. The Act asks for marks that are “interoperable”, but nothing yet defines what that requires. A European code of practice does ask providers to give researchers and journalists detection access — and it is voluntary, while the marking obligation is not. So the binding part requires marking and the part that would let you read the mark is the part nobody has to sign.
Can watermarks be removed?
Yes, and cheaply. One paraphrase takes detection from 99.8% to around 80%. Repeated paraphrasing takes it under 10%. Writing in another language and translating afterwards drops every scheme tested to around a fifth. A round trip through a translator is more variable — one scheme survives at 82.5%, another collapses to 26.3%, and nobody reading a post knows which scheme they are dealing with.
Is this not just a manual for accusing people?
It would be if the markers worked, and the essay’s central finding is that they do not work well enough to accuse anybody. Seven detectors flagged 61% of essays by second-language writers as machine-written, against about 5% for native speakers. Rewriting the same essays to sound more like a native speaker dropped that to under 12%. The detectors were measuring how restricted the writer’s English was, not who wrote it.
A list that does not say so becomes a weapon aimed at people writing in a second language. This one says so repeatedly, and the arithmetic is the reason rather than the manners.
You say the marks justify caution. Is that not just a suspicion nobody can dispel?
That is the strongest objection to the piece and it deserves a straight answer. The safeguard is the same discipline applied to content credentials: the absence of a declaration means nothing. Most people have never heard of any of this. Most posts involving no machine will carry no note saying so, because there was nothing to say.
If “they did not declare” ever becomes evidence, the markers have been handed back the job the essay spends five sections taking off them.
The essay was written with AI and argues that AI writing is fine. Is that not convenient?
It is, and it is declared at the top of the page rather than left to be discovered. The argument does not depend on who wrote it — every figure is sourced, and the sources page records the ten errors found in earlier drafts, including one statistic that had no source at all. Check the working rather than the byline.
Are you not just describing a problem you helped cause?
Partly, yes. The essay measures its own em-dash rate against the models it discusses and reports that it exceeds them. The companion piece argues that many AI vendors converging on one register is a worse outcome than the monopoly people worry about, and that the register is already leaking into human speech. Neither of those is comfortable for the tool that drafted them.
What if I am wrong about someone?
Then you have accused a person of something they cannot disprove, using a test that is wrong more often than it is right about exactly the people it lands on hardest. That is the outcome the whole essay is built to prevent, and it is why the only thing it asks you to do with a marker is read more carefully.