Who Actually Wrote This? A Clear-Eyed Guide to AI Detection in 2026

For the last three years, “did a human write this?” has mostly been a question people asked each other in frustration. In 2026 it has become an operational one, with budgets, policies and — as of next month — European legal deadlines attached to it. Universities need an answer before they grade. Newsrooms need one before they publish. Marketplaces need one before they show a review to a buyer.

The tooling has matured accordingly. What has not matured is the way most organisations use it. A detector's output is routinely treated as a verdict when it is, at best, a well-informed suspicion. Understanding that difference is what separates a functioning content policy from a defamation claim.

The deadline that changed the conversation

On 2 August 2026, Article 50 of the EU AI Act becomes applicable. Its core requirement is disarmingly simple: providers of AI systems that generate synthetic text, audio, image or video must mark those outputs in a machine-readable format, so that they are detectable as artificially generated. Systems already on the European market before that date get a transitional period until 2 December 2026 to bring their marking into conformity.

The obligation sits with the providers of the models, not with the person staring at a suspicious job application. That distinction matters more than it first appears. Even under a perfectly implemented marking regime, the burden of checking falls entirely downstream — on the editor, the admissions officer, the platform moderator. Regulation creates a signal. It does not deliver it to your inbox.

What a detector actually measures

Detection tools work on statistical fingerprints rather than on meaning. Two properties do most of the work.

The first is perplexity — how predictable each word is, given the words before it. Language models are optimised to choose likely continuations, so their output tends to sit in a narrower, smoother band of predictability than human prose, which wanders. The second is burstiness — the variance in sentence length and structure across a passage. People write a long, winding sentence, then a short one. Models lean metronomic.

Modern detectors layer trained classifiers on top of these signals, fitted on paired corpora of human and machine text drawn from several model families. ZeroGPT's free AI checker applies this at both sentence and document level, returning a percentage alongside the specific passages that triggered it. The highlighted passages are by far the more useful of the two outputs, because they tell you where to look rather than what to think.

Watermarking is coming, and it leaks

The industry's preferred long-term answer is provenance rather than inference: mark content at birth instead of guessing at it later. Two mechanisms dominate. C2PA Content Credentials attach cryptographically signed metadata describing how a file was made. SynthID, developed by Google DeepMind, embeds a statistical watermark directly into the token choices of generated text. In May 2026, OpenAI joined the C2PA steering committee and committed to embedding SynthID alongside the Content Credentials it already attaches — a rare moment of alignment between two major labs.

Both mechanisms share a known failure mode, and for text it is severe. Metadata does not survive a copy-paste. And text watermarks, unlike image watermarks, live in a medium that can be rewritten without changing its meaning: swap synonyms, reorder clauses, run a round-trip translation, and the signal degrades badly. Published robustness testing on SynthID-Text has found detection accuracy falling from the 80–85% range to under 30% once a passage has been paraphrased. A February 2026 Microsoft report on media integrity conceded the broader point: no scheme prevents every attack, and platforms strip provenance data routinely in the course of ordinary processing.

Provenance and detection are therefore complements, not substitutes. Watermarking catches the honest case cheaply. Statistical detection is what remains when somebody has taken the trouble to cover their tracks.

Take the false positive problem seriously

This is the part that vendors underplay, and that anyone running a detection policy should have pinned to the wall.

In 2023, a Stanford team led by Weixin Liang tested seven widely used GPT detectors against TOEFL essays written by non-native English speakers. The detectors misclassified 61.3% of them as AI-generated, on average; all seven unanimously flagged nearly one in five. Essays written by US students, run through the same tools, were classified correctly at a far higher rate. The mechanism is not mysterious. People writing in a second language tend to use more formal constructions, a narrower vocabulary and taught sentence patterns — which is, almost exactly, the low-perplexity profile detectors are built to flag.

Tools have improved since, but the bias is structural rather than a bug awaiting a patch, and the other weaknesses have not gone anywhere either. Short samples are unreliable, because a paragraph does not contain enough variance to measure. Heavily edited text — a human draft polished by a model, or a machine draft rewritten by a person — sits in a genuine grey zone that no single percentage can honestly represent. And “humanizer” tools exist for the express purpose of defeating detection, which means a low score proves considerably less than a high score suggests.

Using a checker without doing damage

The tools are genuinely useful. Where things go wrong, the failure is almost always procedural rather than technical.

Treat the score as a trigger, not a conclusion. A high reading is a reason to read closely, ask a question, or request the working file. It is not evidence on its own, and it should never be the sole basis for an accusation, a grade or a rejection.

Feed it enough text. A few hundred words, minimum. Detectors run on statistical variance, and variance needs room to show itself.

Ask for process, not proof. Draft history, version-control timestamps, research notes, a five-minute conversation about the argument — all are more informative than any percentage, and considerably harder to fabricate convincingly.

Write the policy before you need it. Decide in advance what score prompts what action, who reviews the result, and how the person concerned can respond. A policy improvised in the middle of a dispute is a policy that will be argued with.

Look at patterns, not single documents. One flagged text tells you very little. Forty flagged submissions from the same contributor over three months tells you something worth acting on.

What this means for publishers

There is a persistent belief that running content through a detector is an SEO safeguard. It is not, and the misunderstanding wastes real effort.

Google's documented position is that it evaluates quality, not production method. Its spam policies target scaled content abuse — mass-produced pages made primarily to manipulate rankings — and they apply identically whether a person or a model produced the text. Automation has generated helpful content for years: sports results, weather summaries, transcripts. None of that is penalised.

A detector, then, is not a compliance instrument. It is an editorial one. Its value on a content team is catching copy that was generated and shipped without anyone reading it: unverified claims, invented citations, confident nonsense about a subject nobody on the team understands well enough to notice. That is the real risk, and it is a risk about editorial process rather than about authorship.

The better question

The interesting shift underway is not technical. It is that “was a machine involved?” is quietly becoming the wrong question. Nearly every professional text produced in 2026 has been touched by a model somewhere — an outline, a rewrite, a grammar pass. Answering yes tells you nothing about whether the result is accurate, original or worth anyone's time.

The question that survives is: did somebody take responsibility for this? Was it checked? Does a name stand behind the claims? Would that person defend them if challenged?

No tool answers that directly. But a good checker, used with judgement and paired with a policy somebody actually thought about, tells you where to point the question. In a year when everyone is drowning in plausible text, knowing where to look is most of the job.