How to spot AI content
The visual tells everybody learned are already obsolete, and the detectors are unreliable in a way that harms specific people. What is left is less satisfying and more useful.
Key figures
Counting fingers and looking for odd text stopped working; those were bugs, not signatures.
Detectors are biased against non-native English writers, so a high score is not evidence.
Invisible watermarks such as SynthID are the only method designed to survive editing, and they only cover content from participating generators.
Why the old tells stopped working
Six fingers. Warped text on a shop sign. Teeth that do not divide properly. Smooth plastic skin. Every one of these was a genuine artefact of a particular generation of image models, and every one has been fixed.
That is the structural problem with artefact-spotting: the tells are defects, defects get engineered out, and each generation of models removes the exact signals people just finished learning. A checklist of visual tells is a description of last year's models.
Why detectors are not the answer
The obvious alternative is a tool that scores content. For text this is actively harmful, and the reason is documented: Stanford researchers found detectors consistently misclassify writing by non-native English speakers as AI-generated while correctly identifying native writing.
Careful prose written in a second language is flatter, more regular and less idiomatic, which is exactly the texture a detector reads as machine-written. So the errors are not random; they land on the people least able to contest an accusation. There is a fuller treatment of that evidence on our page about whether AI detectors work.
What actually carries a signal
The one technical approach that is not a guessing game is watermarking at the point of generation. Google DeepMind's SynthID embeds imperceptible watermarks in images, video, audio and text produced by Google's consumer AI products. For images and video they are designed to survive cropping, filters, frame rate changes and lossy compression; for audio, added noise, MP3 compression and speed changes. Content can be checked by uploading it to Gemini or through a detector portal.
| Method | What it tells you | What it misses |
|---|---|---|
| Visual artefacts | Something about an older model generation | Anything made by a current model |
| Text detectors | A probability from an undisclosed method | Reliability, and it misfires on non-native writers |
| Watermarks such as SynthID | A positive signal that content was generated | Everything from generators that do not participate |
| Provenance and process | Where a claim came from and who stands behind it | Nothing, but it takes effort |
Note the asymmetry in the third row. A watermark hit is strong evidence that content is AI-generated. The absence of one is almost no evidence at all, because most generators do not embed anything.
What to do instead
The workable approach abandons the question “was this made by a machine” and asks the one that actually matters.
Does it contain anything checkable? A specific number with a named source, a date, a method, a stated limit. Generated filler avoids all four, because they are the parts that cost something to produce and can be shown to be wrong.
Who is accountable? A named person or organisation that can be contacted and contradicted. This does not change when the tooling changes.
For an image, where did it come from? Reverse image search and the original posting date beat pixel-level inspection, because provenance is about the chain of custody rather than the pixels.
The reframing matters because the original question is becoming unanswerable and was always the wrong one. A machine-drafted article that cites checkable sources is more trustworthy than a human-written one that cites nothing.
Sources and method
- SynthID, Google DeepMind. Source of the content types covered, the imperceptibility of the watermarks, their robustness to editing and the detection routes.
- GPT detectors are biased against non-native English writers, Liang, Yuksekgonul, Mao, Wu and Zou, Stanford. Source of the finding that detectors misclassify non-native English writing.
- Guidance on AI Detection and Why We are Disabling Turnitin’s AI Detector, Vanderbilt University, August 2023, on what a detector score is worth in practice.
- The characteristics table is this page's own summary rather than a quotation.
- Every source here was opened and checked on 25 September 2026. Where nothing has been measured, this page uses a table and says so rather than drawing a chart of an opinion.