Anyone trying to work out how to tell if text was written by AI runs into an awkward fact early: the tools sold to answer that question are the least trustworthy part of the process. Teachers, editors, hiring managers and ordinary readers all want a quick verdict, and a whole industry has grown up to supply one. The verdicts, though, are probabilities dressed as certainties, and the research on them is not flattering.

This guide covers what detectors actually measure, where they fail, what the better alternatives look like, and what a careful reader can do without any software. Established findings are separated from guesses, because in this area the two get blurred a lot.

How do AI text detectors work?

Most text detectors are classifiers. They are trained on large piles of human writing and machine writing, then asked to guess which pile a new passage resembles. Some lean on predictability: language models tend to pick likely next words, so text that is smooth and statistically expected can score as machine-made. Others look at structure, such as how much sentence length varies, how often certain transition phrases recur, or whether the rhetoric follows a familiar template.

Notice what is missing from that description. Nothing in the process reads a signature or finds a fingerprint. A detector never sees who or what produced the words. It sees the words, compares them with its training data, and reports a score. That is why the output is best read as a hunch with a number attached.

The distinction matters because a human can write predictable prose and a model can write odd prose. A careful, formal essay by a student working in a second language may look “smooth” to a classifier. A machine told to be quirky and uneven may not.

How accurate are AI detectors?

Not very, according to independent testing. A 2023 study by Weber-Wulff and colleagues examined 14 detection tools, including Turnitin and GPTZero, and reported that all of them scored below 80 percent accuracy, with only five above 70 percent. The tools also leaned toward calling text human, and their accuracy dropped further once the text was paraphrased.

Vendors tell a more favorable story. Turnitin has said its false positive rate is below 1 percent and has acknowledged a false negative rate of about 15 percent. A Washington Post test in 2023, using a much smaller sample, reported a far higher false positive rate, around 50 percent. Small samples are shaky evidence in either direction, so the honest summary is that nobody outside the vendor can confirm the low figures, and independent checks have turned up worse ones.

OpenAI offers perhaps the most telling case. In early 2023 the company released its own classifier, and its announcement said the tool flagged only 26 percent of AI-written text in a test set while wrongly labeling 9 percent of human writing as AI. It also warned that the tool was very unreliable on texts under 1,000 characters. By July 20, 2023, OpenAI had pulled the classifier because of its low accuracy. If the company that builds the writing models could not build a dependable detector, that says something about the difficulty.

Why do detectors flag real human writing?

Because plenty of human writing looks, statistically, like the machine kind. The best-known demonstration came from a Stanford team (Weixin Liang, James Zou and colleagues), whose paper on detector bias was published in the journal Patterns in 2023. They tested several widely used detectors on essays by native and non-native English speakers. The detectors handled native writing well but kept labeling non-native writing as machine-generated. Wikipedia’s summary of the work puts the average false positive rate for non-native essays at 61.3 percent across the seven detectors tested.

The likely reason, which the authors themselves frame as a suggestion, is that writers with a smaller working vocabulary and plainer grammar produce text that is more predictable. Predictability is what the detector reads as machine-like.

Other groups appear to be affected too. A September 2024 report from Common Sense Media, as summarized on Wikipedia, found false positive rates of 20 percent for Black students, 10 percent for Latino students and 7 percent for white students. Neurodivergent writers have also been described as flagged more often. These second-hand summaries are worth checking against the original reports before anyone quotes them, but they point the same way as the Stanford result.

There are individual cases as well. The author Janelle Shane saw passages of her own book flagged as AI-generated in 2023. At UC Davis that year, two students were referred for misconduct after detector flags and were cleared after an investigation. At Texas A&M–Commerce, an instructor asked ChatGPT whether it had written his students’ essays; it said yes, which proves nothing, since chatbots cannot recall what they have produced for other people.

Can AI writing be disguised?

Easily, and that is the other half of the problem. Paraphrasing a passage, editing it by hand, or mixing human sentences with machine ones can sharply reduce detection. A 2023 study in the journal Eye tested Originality.AI on GPT-4 text and found a mean accuracy of 91.3 percent. After the same text was run through a rewriting tool called Undetectable.ai, accuracy fell to 27.8 percent.

So detectors err in both directions at once. They accuse honest writers and miss anyone who bothers to rewrite. A tool with both flaws is hard to use for anything with consequences. That is the reason several universities stepped back: members of the UK’s Russell Group, including Cambridge, opted out of Turnitin’s AI detection feature in 2023, citing reliability concerns, and the University of Texas at Austin followed months later.

Is there a better method than detection?

Two other approaches exist, and both work differently from detectors.

The first is watermarking. Here the model’s maker builds a hidden statistical pattern into the output as it is generated. In 2023, Kirchenbauer and colleagues described a text watermark that can survive light paraphrasing. Google DeepMind’s SynthID applies the same broad idea to images, and Google has applied it to other media too. A watermark is far more convincing than a guess, since the check looks for a planted signal rather than a style.

It has real limits, though. It only covers models whose makers participate, which leaves out most open-source models. Heavy editing, or passing text through a second model, can weaken it. And a detector checks only for its own company’s mark: in July 2026 the fact-checking outlet Lead Stories reported that Gemini’s SynthID check repeated its answer about the first file in a chat when asked about later ones, and Google said a fix had rolled out globally by July 16. Lead Stories also pointed out the simple logic problem: no watermark does not prove something is not AI-made.

The second approach is provenance, mainly the C2PA standard behind “Content Credentials.” Instead of guessing from content, it attaches cryptographically signed records of how a file was created and edited. Adobe, Google, Sony and others support it, and it applies mainly to images, audio and video rather than prose. Its weak spot is that metadata can be lost when a file is screenshotted or re-encoded, and implementation flaws happen: Nikon revoked certificates in 2025 after a vulnerability was found in Z6III firmware that let a fake image keep a valid signature.

Neither approach is a universal answer. Both are more principled than reading style tea leaves.

How can you tell if an image was made by AI?

Visual inspection used to be fairly productive. Garbled hands, nonsense lettering and melted backgrounds were common giveaways. Those flaws have faded as models improved; Wikipedia’s overview of AI art notes that text rendering and hand quality have been specifically targeted in recent releases. An image made today may show none of them, so a clean-looking picture proves little.

What tends to help more is context rather than pixels. Where did the image first appear, and who posted it? Does a reverse image search turn up an earlier source? Does the claimed event have coverage from outlets that sent a photographer? Are there Content Credentials attached, and do they verify? Those questions take minutes and do not depend on a classifier’s mood.

The same logic applies to text. Ask about the source before you ask about the style.

What patterns suggest machine-written text?

Some patterns are commonly reported, though every one of them is a hint, not proof. Descriptions of detector features mention consistent, tidy grammar, formulaic transitions, repeated discourse markers, and recurring rhetorical templates. In practice readers also mention a smooth evenness in sentence length, a habit of summarizing a point right after making it, vague claims without a concrete example, and lists of three.

Here is the catch: all of those traits also show up in human writing, especially corporate, academic and school writing, which is where much of the training data came from. Models learned the habits from us. A reader who treats “sounds like a chatbot” as evidence will sometimes be right and sometimes accuse an innocent person.

Fabrication is a sturdier signal than style. Invented citations, quotes that cannot be found, or details that fall apart when checked are problems whatever produced them, and they are verifiable. Checking a few specifics beats scoring a tone.

What should teachers and editors do instead?

The research points toward process rather than product. Asking for drafts, version history, notes and a short conversation about the work gives far better evidence than a percentage. Guidance from several sources, including the article summaries cited below, warns against using detector scores as the sole basis for a penalty. That seems a fair reading of the evidence, though policy choices are judgments, not facts.

For readers who simply want to know how much to trust something online, the practical approach is the older one: check the author, check the sources, and look for corroboration. Whether a sentence came from a person is less important than whether it is true.

Common questions

Can any AI detector be trusted completely?

No. Independent tests have found accuracy below 80 percent for every tool examined in one 2023 study, and false positives remain a documented problem. A score can be one input among several, but it should not decide a case alone.

Why do detectors flag my own writing?

Plain, formal or highly consistent writing can look statistically similar to model output, and the Stanford study found non-native English writers hit particularly hard. Being flagged says something about the tool’s limits, not about your honesty.

Does ChatGPT know whether it wrote something?

No. A chatbot has no memory of what it has produced for other users, so asking it “did you write this?” yields a plausible answer, not a fact. The Texas A&M–Commerce case from 2023 is the standard warning.

Is watermarking the solution?

It is promising but incomplete. It works when the maker embeds a mark and the content has not been heavily altered, and it cannot cover models whose makers do not participate.

Can you spot AI images by looking at hands or text?

Sometimes with older images, rarely with new ones. Those flaws have been steadily reduced by newer models, so source checking and provenance data are more reliable.

Will detection get better?

Possibly, but it is an arms race, and rewriting tools already exist. Predictions here are guesses, and the structural problem, that the tool never sees who wrote the text, does not go away.

What to remember

Detectors give probabilities, not proof, and independent testing shows both false accusations and missed detections. Non-native English writers have been shown to be flagged unfairly often. Watermarks and provenance records are sturdier ideas but cover only some content. When it matters, look at where the material came from and whether its claims check out.

Sources