How can I tell if something is AI?
There is no reliable test. AI detectors guess from statistics and get it wrong often: Stanford researchers found seven detectors flagged 61% of essays by non-native English speakers as AI. Your best tool is provenance, not forensics — check who posted it, where it first appeared, and whether the source can be traced.
Why — the first-principles explanation
Generative models work by predicting the most likely next piece of content — the next word, the next patch of pixels — over and over. Averaged across a whole paragraph or image, that means the output tends to sit near the middle of the distribution of everything the model learned. It avoids weird word choices, rare sentence shapes, and odd compositions, because those are by definition unlikely.
That statistical averageness is the only signal any detector has. Nothing is stamped into the text. There is no hidden serial number in a normal sentence. A detector is just measuring "does this look more typical than a human would usually be?" — and then drawing an arbitrary line.
The problem is that the two populations overlap. Plenty of humans write in the middle of the distribution: people writing in a second language, people writing formal or technical prose, people following a template, students taught to write plainly. Meanwhile, anyone can push a model out of the middle by asking it to write loosely. So the signal that flags AI also flags careful, plain, or non-native human writing — and misses AI that was told to be strange.
This gap shrinks every year. As models get better at imitating human variation, the statistical daylight a detector needs disappears. That is why the durable answer is provenance — evidence about where a thing came from — rather than analysis of the thing itself. A traceable source, a photographer with a camera roll, a document with an edit history, a poster with years of consistent output: those are hard to fake and don't depend on out-guessing a model.
An example that makes it click
Imagine trying to tell whether a cake was baked by a factory or a person. The factory cake is suspiciously even — every crumb the same, frosting perfectly flat, no scorched edge. So you decide: "too even means machine."
Now a careful baker brings you a beautiful, perfectly even cake, and you call her a factory. Meanwhile the factory sets its machine to add a wobble and a burnt corner, and you call it homemade. You were never testing where the cake came from. You were testing how tidy it was — and tidiness is something both a machine and a person can choose. The only real answer is to check the receipt and the kitchen.
How to do it
- Ask what kind of thing it is. Text, image, video, and a chat account all need different checks — there is no universal AI test.
- Check provenance first. Who posted it? Do they have a long, consistent history? Does the content exist anywhere earlier, from a traceable original source?
- For images and video, reverse image search it. AI slop is usually posted by anonymous, high-volume accounts with no original source and no photographer.
- Look for Content Credentials (C2PA) metadata if the file came to you directly — but note most social platforms strip metadata on upload, so its absence proves nothing.
- For text, read it yourself before you use any tool. Look for confident claims with no specifics, invented citations, and paragraphs that say very little at length.
- If you use a detector, run the same text through two or three. Treat agreement as a weak hint and disagreement as a non-answer.
- Never treat a detector score as proof, especially about a real person. Ask for the source, the drafts, or the raw file instead.
Key facts
- Seven widely used GPT detectors flagged 61.22% of 91 TOEFL essays written by non-native English speakers as AI-generated; all seven unanimously flagged 19% of them (Liang et al., Patterns, 2023).
- The same detectors were near-perfect on essays written by U.S.-born eighth-graders — the errors were concentrated on non-native writers, not spread evenly.
- OpenAI withdrew its own AI Text Classifier on July 20, 2023, citing its 'low rate of accuracy.' The company that built ChatGPT could not reliably detect ChatGPT.
- Detectors score text using 'perplexity' — how surprising the word choices are — which correlates with writing sophistication, not with authorship (Prof. James Zou, Stanford).
- Vanderbilt University disabled Turnitin's AI detector in August 2023, noting that even a claimed 1% false positive rate across its 75,000 annual submissions would mislabel about 750 papers.
▶ The 60-second explainer (script)
Can you tell if something is AI? Honestly — not reliably. And anyone selling you a tool that says otherwise is overselling. Here's why. AI generates content by picking the most likely next word or pixel, over and over. So its output lands in the safe middle: tidy, average, unsurprising. That averageness is the only thing detectors can measure. The problem is that humans live in that middle too. Stanford researchers ran seven detectors on essays by non-native English speakers and 61 percent got flagged as AI. The same detectors were near-perfect on American eighth-graders. The tools weren't detecting machines. They were detecting plain writing. OpenAI shut down its own detector in 2023 for low accuracy — the makers of ChatGPT couldn't spot ChatGPT. So what actually works? Provenance. Not what it looks like, but where it came from. Who posted it? Do they have a real history? Does an original exist? Can you reverse image search it? A traceable source is hard to fake. A vibe is not. Use detectors as a hint if you like. Never as proof — especially about a person.
What authoritative sources say
People also ask
Is there any tool that just tells me for sure?
No. Every text detector returns a probability, not a verdict, and those probabilities are wrong often enough that no serious institution treats them as evidence. OpenAI retired its own detector rather than keep shipping an unreliable one.
What about the 'AI look' in images — extra fingers, glossy skin?
Those tells are real but fading fast, and they only catch careless output. A well-made AI image in 2026 usually has correct hands. Absence of glitches is not evidence of a human.
Do AI images have hidden watermarks?
Some do. Content Credentials (C2PA) metadata and invisible watermarks like SynthID are added by several major generators, but most social platforms strip metadata during upload, so a missing credential tells you nothing.
Is it safe to accuse someone based on a detector score?
No. Detector errors fall hardest on non-native English speakers and plain, formal writers. If the stakes are real, ask for drafts, version history, or a conversation about the work instead.
The same question, asked other ways
- How can I tell if something is AI-generated?9,900/mo
- How can you tell if something is AI-generated?1,900/mo
- How to check for AI?1,000/mo
- How to tell if something is AI generated?720/mo
- How to detect AI?720/mo