Key points
- AI content detection looks a piece of content over and assigns a score for how likely it is to be AI-made. It doesn't come back as yes or no — it comes back as a probability.
- A machine looks at things a person can't. Fine regularities in how pixels sit, or how predictably one word follows another, are the kind of trace it watches for.
- A check can fail in two ways, and which hurts more depends on where it's used. Where a person is being judged, as in a text check, many views weigh flagging something a person made as AI-made more heavily.
- The score is easy to shake. Just resaving a file or lightly rewording a passage can flip a result.
- Because of that, the center of gravity is shifting from catching fakes after the fact toward marking and recording content at the moment it's made.
The analogy
Set an item on a pawnbroker's counter and the appraiser takes out a loupe and looks over the seams. They weigh it, scratch the surface to check the material, look at whether the finish is even. And what comes back is never "this is genuine" — it's "this appears genuine." It's an opinion, not a certainty.
A well-made forgery can pass that appraisal. The reverse happens too: a genuine piece worn down by years of use, carrying old repair marks, can get mistaken for a fake. The first kind of mistake leads to a loss. The second leaves the owner feeling wronged.
Put the same item on a different counter and you can get a different opinion. That's why big transactions don't lean on one appraisal alone — they're checked against a receipt and a warranty too. Working out whether something is AI-made runs into that same situation.
In detail
Human eyes have already lost half their edge
A few years back there were tells a person could catch by eye: an odd number of fingers, garbled text, a background that seemed to melt. Those tells have largely disappeared.
In detection experiments, people's accuracy on well-made face images has dropped to coin-flip level in some results. It varies with the medium and how the content was made, but trusting the eye alone is hard. Answers given with confidence are often wrong too, and experiments suggest many people rate their own ability to tell real from fake more highly than it really is.
Text is harder still. A photo at least leaves pixels behind as a trace. A well-written passage leaves almost nothing on the surface for an eye to grab onto.
What a machine looks at instead
A detection model watches for something different than a person does. For an image, it looks for fine regularities left behind by how the image was produced — a photo shot through an actual lens and sensor carries a particular texture of noise, while a generated image reads differently.
For text, it measures how predictably one word follows the last. Human writing tends to have spots that jump outside what's expected here and there; generated text tends, on average, to flow more smoothly than that.
The catch is that none of these traces stay fixed. As generation methods change, the traces change with them. A detector trained on last year's output loses its footing in front of this year's.
The two kinds of mistake weigh differently depending on where they're used
A detection result carries two possible kinds of mistake: missing AI-made content, and flagging human-made content as AI-made. Lower the threshold and the first kind drops while the second climbs. Raise it and the reverse happens.
Where a person is being judged, many views weigh flagging something a person made as AI-made more heavily. Someone whose own writing gets marked generated has no good way to argue back — a detection score doesn't come with a reason attached. Some also see missed cheating as unfair to people who wrote honestly, so how much to reduce each kind of mistake is decided case by case. Conversely, where a miss does harm the moment it happens, as with deepfake fraud or impersonation, missing it is serious too.
There's a further trap here. Where AI-made content is genuinely rare in a pool, even an accurate-sounding detector can end up flagging mostly human content, because most of what gets flagged can turn out to be human-made once the true rate is that low. Rare things magnify even a small false-positive rate.
Small changes shake the score
Because detection leans on surface traces, disturbing the surface shakes the reading. Resaving a photo, resizing it, or photographing it off a screen can move the score substantially.
Text is even more sensitive. Reworking a few sentences or passing a passage through another tool once can flip the reading. Going the other direction, writing whose words and phrasing are easy to predict — text written in a language that isn't the writer's first, or something in a fixed, formal format — can score high even when a person wrote every word.
That instability is why a detection score is safer read as a signal for where to look closer, not as the basis for a conclusion. Wherever one score alone decides the outcome, someone is likely to end up wrongly accused.
From catching fakes to proving what's real
Catching fakes is always a step behind. Every new way of generating content means detectors have to learn its traces from scratch. So the center of the response has been shifting.
The newer approach is marking content at the moment it's made and attaching a record to it as it moves. A mark too subtle to notice gets embedded inside the content itself, and the file carries a record of where it came from and what was done to it. Instead of chasing fakes after the fact, the trail gets laid down from the start.
These marks can be stripped too, of course, which is why detection stays necessary. Running marks, records, and detection together, without leaning the whole conclusion on any one of them, is the most realistic approach available right now.
More precisely
Many detection models are classifiers trained on a mix of genuine and generated examples to tell the two apart. For text, some approaches skip that training and measure only the prediction probabilities a language model assigns. Its output is a probability score, and the threshold for calling something AI-made is a line a person chooses, not something the model decides on its own. Looking at raw accuracy alone is misleading — precision (how often a flagged item is actually right) and recall (how much of the real generated content actually gets caught) need to be read together, and moving the threshold trades one against the other.
The analogy breaks down in places too. An appraiser looks at a maker's skill and materials; a detection model doesn't look at meaning at all, only at statistical texture. So content can make perfect sense and still have nothing to do with the verdict, and content that makes no sense can pass if its texture resembles a human's. An appraiser can also explain their reasoning and point to a specific seam or mark; a detection model returns a score and nothing more, with no way to ask it why.
Above all, detection only answers "was this AI-made." Whether the content is true, or whether it harms someone, is a separate question entirely.
Last verified: 2026-10
Try it yourself
- Which Face Is Real? ailearn.space Pick the real person's photo out of a pair and get a feel for how hard telling them apart by eye really is
- Detect Fakes (real-or-fake research experiment) ailearn.space A research experiment on how well people tell real from fake, or where a speech came from. Use the top menu to switch to the image, video or audio detection experiments and try them yourself (English, ages 18+)
Common misconceptions
It's easy to think a detector calling something AI-written is certain, but actually it's only a probability score, and human writing scoring high is not rare, especially in text written in a language that isn't the writer's first, where the phrasing is easier to predict.
It's easy to think a sharp eye is enough to tell the difference, but actually, depending on the experiment, people's accuracy can fall to coin-flip level.
It's easy to think better detection technology will settle this, but actually generation improves right alongside it, which keeps detection a step behind.
One-line summary
AI content detection is closer to a probability-based opinion than a verdict, and where one score alone is used to judge a person, many views weigh mistaking their own work for an AI's more heavily.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-18