What is AIGC detection?
AIGC detection means checking whether a piece of content — text, an image, audio or video — was generated by AI; AIGC is short for AI-generated content. An AIGC detector works in one of two ways: it analyses the content itself for statistical traces a model leaves behind, or it reads a watermark or signed provenance record the generating tool attached.
The first gives a probability, the second a reliable yes, but only for tools that took part. Neither can prove that something was made by a person. The term is most common in China, where 'AIGC' is the everyday word for generative AI output and many universities now check theses for an 'AIGC rate' alongside plagiarism; in Western marketing the same checks surface as the AI labels social platforms add to posts and ads.
How AIGC detectors work
Every detector on the market uses one of three methods, and they fail in different ways.
- Artifact analysis (classifiers): A model trained on large sets of real and generated samples scores new content for the patterns generators leave: in images, regularities in noise and frequency spectra, texture that is too clean, inconsistent lighting, warped text or hands; in text, how predictable each next word is (perplexity) and how evenly that predictability is spread (burstiness). The output is a confidence score, not a fact.
- Invisible watermarks: The generator embeds a signal in the output as it is made. Google's SynthID marks images, video, audio and text from Gemini, Imagen, Veo and Lyria; it is designed to survive cropping, resizing, compression and filters, and a matching detector reads it back. A watermark only proves anything for the tools that embed it.
- Provenance metadata (C2PA Content Credentials): The creating tool signs a manifest into the file recording what made it and what edits followed, including whether AI generated it. It is cryptographically verifiable while it survives, but a screenshot, a re-encode or an upload pipeline that strips metadata removes it. A related, unsigned flag is the IPTC 'digital source type' field, which some platforms read.
AIGC detector tools
| Tool | Method | Checks | Limits |
|---|---|---|---|
| SynthID Detector (Google) | Watermark | Images, video, audio, text from Google AI models | Only finds Google's own watermark; says nothing about other generators |
| Content Credentials Verify (C2PA) | Provenance metadata | Any file carrying a C2PA manifest | Empty result once metadata is stripped; absence proves nothing |
| Hive AI-generated content detection | Classifier | Images, video, text, audio | Probability score; accuracy varies by generator and edits |
| Sightengine | Classifier | Images and video | Probability score; heavy compression and edits weaken it |
| AI or Not | Classifier | Images and audio | Probability score; lags behind newly released generators |
| GPTZero, Originality.ai, Turnitin | Classifier | Text | False positives on human writing, especially non-native English; easily defeated by paraphrasing |
How reliable is AIGC detection?
- Watermark and metadata checks are the only methods that give a definite positive, and only for content from participating tools. A negative result is not evidence of human origin.
- Classifier scores are estimates. They are strongest on unedited output from generators they were trained on, and weaker on new models, compressed or resized files, and content a person has edited.
- Text detection is the weakest case. OpenAI withdrew its own AI text classifier in 2023, citing low accuracy, and a 2023 Stanford study (Liang et al.) found popular detectors routinely flagged essays by non-native English writers as AI-written.
- Treat any single score as a lead, not a verdict: combine a watermark check, a metadata check and a classifier, and weigh the context the content came from.
What is an AIGC rate (AIGC check)?
An AIGC rate is the share of a document that a detector scores as likely AI-written, shown as a percentage. It comes from China's universities, which run a thesis through an AIGC check alongside the plagiarism check (services such as CNKI's AIGC detection report one), and some set a maximum rate a thesis may have before it is sent back. The rate is a classifier's estimate summed over paragraphs, so it inherits every limit above: two services can report very different rates for the same text, edited AI text can score low, and plain human writing can score high.
How to lower an AIGC rate honestly
- Write the argument yourself and use AI for research and checking, not for the paragraphs that get scored. Detectors flag uniform sentence length and predictable word choice, which is what unedited model output has.
- Keep your drafts, notes and version history. If a score is challenged, a document's history is better evidence than any second detector.
- Run the same checker your institution uses and rewrite the passages it marks, in your own words and with your own examples. Paraphrasing tools sold to lower a rate often make the text worse and break the rules they are meant to get around.
- For marketing images and video the question is different: not passing a detector, but not looking fake. Our guide to reducing AIGC detection covers the visual tells (plastic skin, warped labels, impossible lighting).
Can you remove an AIGC watermark or label?
A visible label, such as the AI tag a phone gallery or app stamps on an edited photo, can be cropped. That leaves the invisible marks: SynthID is designed to survive cropping, compression and screenshots, and C2PA or IPTC metadata goes only when the file is re-encoded or stripped. Where labels are required, removing them is the problem: China's labeling measures forbid deleting or tampering with AIGC labels, and Meta, TikTok and YouTube ask you to disclose realistic AI content whether or not a label survived. For brands, keeping the label costs far less than being caught without one.
Put it to work
- Product UGC photos — AI-made creator-style product photos; label them as AI-generated wherever a platform asks.
- AI content disclosure — what Meta, TikTok and YouTube expect you to label, and how.
How it relates to AI UGC
For a brand running AI UGC, detection matters less than disclosure. Meta and TikTok read C2PA metadata (Meta also IPTC) to apply AI labels, and Meta, TikTok and YouTube all ask creators to disclose realistic AI content themselves, so the safe position is to label AI-made ads yourself rather than count on a detector missing them. Google's Gemini image and Veo video models embed SynthID in everything they generate, so content made with them stays identifiable as AI even after the file's metadata is gone.
Related terms covered here
- AI watermark
- An AI watermark is a signal embedded in AI-generated content that identifies it as synthetic. Visible marks (a logo in the corner) are easy to crop; invisible ones such as SynthID are woven into the pixels, audio or word choices and need a detector to read.
Sources
- Google DeepMind: SynthID
- Content Credentials: Verify (C2PA)
- C2PA: Coalition for Content Provenance and Authenticity
- OpenAI: New AI classifier for indicating AI-written text (withdrawn July 2023)
- Liang et al. (2023): GPT detectors are biased against non-native English writers
- CNKI (China National Knowledge Infrastructure)