AI Content Detection

AI content detection is the attempt to determine whether text, images, audio or video were produced by an AI model rather than a person (see AI content generation). There are three main approaches: statistical classifiers that guess from the content itself, watermarks that AI providers embed in their outputs, and provenance metadata that records how a file was created. They differ a lot in reliability, especially for text.

Classifiers such as GPTZero, Originality.ai and Turnitin's detector analyze features like perplexity (how predictable each word is) and return a probability. Their weaknesses are well documented. OpenAI withdrew its own text classifier on July 20, 2023, citing its 'low rate of accuracy'; it had flagged only 26% of AI-written text correctly while mislabeling 9% of human text as AI. A 2023 Stanford study in the journal Patterns found that seven popular detectors labeled 61% of essays by non-native English writers as AI-generated, while classifying US students' essays almost perfectly (Liang et al.). Light editing or paraphrasing also defeats most classifiers, so a score is evidence to weigh, not proof.

Watermarks work differently: the provider alters its output in a way only a matching detector can recognize. Google's SynthID marks images, audio and video from Google models and some partners, and Google's public SynthID Detector checks files for those marks. On October 5, 2026, OpenAI announced an invisible text watermark called textGrain for ChatGPT and Codex output in the EU, with opt-in watermarking for API customers worldwide (OpenAI). Its own tests show the limits: at a 1% false positive rate the detector found the watermark in about 80% of 200-token passages and 95% of 400-token passages, and replacing 25% of the words with synonyms cut detection to 17%. OpenAI is limiting the text detector to approved researchers for now. Provenance metadata based on the C2PA Content Credentials standard records a file's origin and edits, but it can be stripped.

Regulation is now the main driver. Article 50 of the EU AI Act, applicable from August 2, 2026, requires providers of generative AI systems to mark outputs so they are machine-readable and detectable as AI-generated, and requires deployers to disclose deepfakes and AI-generated text published to inform the public on matters of public interest, unless that text has had human review and editorial responsibility. A Code of Practice published in June 2026 sets out how to comply; about 190 organizations had signed it by the end of July. AI regulation covers the wider rules.

For search, detection matters less than many people assume. Google says it judges content on quality rather than production method, although its rater guidelines allow the Lowest rating when nearly all main content is AI-generated or paraphrased with little effort or originality, and its spam policies target mass-produced pages without added value (see AI spam). No search engine has said it uses a detector score as a ranking signal. The practical goal for content teams is content worth reading, with original information and accurate facts, rather than content that passes a detector. Content quality signals and E-E-A-T say far more about whether a page will be cited than any detector does.