The tool that promises certainty is the one to distrust.
DetectionDE
There is a particular comfort in a number. You paste a paragraph into a box, a meter swings, and a percentage appears: 94% AI. The judgment feels delivered, settled, outsourced. That is precisely the feeling to distrust. The detector has not read the text the way you would. It has measured a few statistical properties and compared them to a threshold, and on any single document, that comparison is unreliable enough that no serious institution should hang a decision on it.
This is not a fringe position. It is roughly where the evidence has landed, and it includes the people who built these tools.
In January 2023, OpenAI released an AI Text Classifier, a tool meant to flag writing produced by its own models. By its own published numbers, the classifier correctly identified only 26% of AI-written English text as "likely AI," while wrongly flagging human-written text as AI about 9% of the time. Six months later, in July 2023, OpenAI quietly withdrew it, citing its "low rate of accuracy."
Sit with that. The organization with the deepest possible knowledge of how its models write (the weights, the training data, the sampling behavior) could not reliably tell its own output apart from a human's. Every detector you'll find advertised today is attempting the same task from the outside, with less information, against a moving target of models it has never seen. When the manufacturer can't detect its own product, skepticism toward third-party tools is the baseline, not the cynical extreme.
False positives flag real human writing as machine-made. This is the dangerous one, because it accuses. The most-cited evidence comes from a 2023 Stanford study of seven popular GPT detectors: more than half of essays written by non-native English speakers (real people, taking the TOEFL exam) were misclassified as AI-generated. Nearly 98% of those essays were flagged by at least one of the seven detectors. The same detectors were near-perfect on essays by U.S. eighth-graders. So the tool is measuring something that happens to correlate with not being a native English writer, not "AI-ness."
The mechanism is worth understanding, because it explains a lot. Many detectors lean on perplexity (roughly, how surprising the word choices are). Simpler, more predictable phrasing scores as "low perplexity," and detectors treat low perplexity as a tell for AI. Non-native writers tend to use more common words and steadier constructions, so they trip the wire. The same Stanford team showed the reverse, too: prompt ChatGPT to write in more elaborate language, and the detectors started calling its output human. The signal these tools rely on is gameable from both sides.
False negatives let machine text through, and they require almost no effort to produce.
If a detector could be reliably beaten only by experts with custom tools, that would be one thing. The reality is that the standard countermeasure is paraphrasing, and it works alarmingly well. In one widely cited evaluation, running AI text through a paraphraser dropped a strong detector's area-under-the-curve score from 96.5% to 25.2%, from "quite good" to worse than a coin flip. Recursive paraphrasing, where the output is run through again, pushes the numbers lower still, and human raters judged most of the rewritten passages (about 77%) to be high quality.
You don't even need a dedicated tool. Light editing, a different prompt, re-running the text through another model, or for images a simple re-compression or crop: all of it perturbs the surface statistics detectors latch onto. The detector is catching "this has the fingerprints of a particular generation process," not "this was generated by a machine." Fingerprints wipe off easily.
They don't generalize to models they haven't seen#
The harder, structural problem is generalization. A detector is trained to recognize the artifacts of the models that existed when it was built. New models train on fresh data, get better, and shed those artifacts. Benchmark studies in 2025 and 2026 found the same pattern across the board for AI-image detectors: tools that score impressively in the lab (on the exact data they were tuned for) degrade sharply on real-world images and on outputs from generators they weren't trained against. The phrase researchers keep using is that detectors overfit to superficial artifacts. They learn the tells of yesterday's models, and yesterday's models are not what you're up against.
This is why detection is a treadmill that the generators always win. Every new model release is a fresh out-of-distribution problem for every existing detector. The defender has to retrain and recalibrate constantly; the attacker just ships.
Not nothing, but far less than the marketing implies. Across a very large population of documents, a calibrated detector can surface rough trends: this corpus has more machine-written text than that one. That's a statistical instrument used at statistical scale.
The failure is specific: treating a confidence score as a verdict on one person's one piece of work. That is the use case where false positives ruin someone's week and where, as schools and universities have learned the hard way, the accusation can land on exactly the people least able to defend against it. A number that is wrong 1-in-10 times (or worse) is, at most, a prompt to go look, not evidence of misconduct.
The honest move is to stop asking a meter for a verdict and start doing the two things that actually hold up.
First, provenance. Where did this come from, and can the claim be traced? Content-credential standards like C2PA attach a signed, tamper-evident history to a file: who made it, with what tool, what was edited. It's not yet everywhere, and absence of a credential proves nothing, but a verifiable trail is worth more than any black-box score. We unpack the strengths and gaps in how content provenance and C2PA work.
Second, lateral reading, the move professional fact-checkers use. Instead of staring harder at the suspicious item, you leave it: open new tabs, find the original source, check who is saying it elsewhere, and read sideways across the web. It's slower than pasting text into a box, and that slowness is the point. More on the habit in media literacy for the AI era.
If you came here hoping for a better detector to swap in, that's the wrong frame. The skill is noticing, sourcing, and refusing to be rushed, and it can't be outsourced to a confidence score. Conveniently, that is the exact muscle our daily puzzle is built to train. (And if you want the human-eye tells that still work on machine writing, start with how to spot AI text.)
Can an AI detector prove a specific piece of writing was made by AI?
No. A detector returns a probability, not a verdict, and on any individual document that number can be wrong in either direction. It is evidence at best, never proof. Treat a high score as a reason to look closer, not a conclusion.
Are AI detectors completely useless then?
Not completely. At large scale (measuring rough trends across thousands of documents), a calibrated detector can be informative. The failure is specific and important: using a confidence score to judge one person's one piece of work. That is exactly where the false-positive and bias problems do their damage.