As AI writing tools spread, demand for a way to tell whether a piece of text was written by a machine has soared, especially in schools. A range of AI detection tools now offer a verdict, often as a percentage. Unfortunately, the technology cannot reliably do what many people expect of it.
Updated September 2026: we added the Stanford research on detector bias, the accuracy figures for OpenAI’s withdrawn classifier and details of Google’s open-source text watermark.
How AI detectors work
Most detectors look for statistical patterns. AI-generated text tends to be predictable: words that a language model would be likely to choose, with fairly even sentence structure. Detectors measure properties such as perplexity, how surprising the word choices are, and burstiness, how much sentence length and structure vary. Some are themselves trained models that classify text as human or AI.
Why they fall short
- False positives. Clear, simple or formulaic human writing can look “predictable”. Stanford researchers found that seven popular detectors flagged 61% of TOEFL essays by non-native English speakers as AI-generated, while classifying essays by US eighth graders almost perfectly.
- Easy to evade. Light editing, paraphrasing or asking the AI for a different style can change a detector’s verdict.
- Mixed authorship. Much writing now combines human drafting with AI editing, or the reverse. There is no clean line to detect.
- Moving target. Detectors trained on older models perform worse on newer ones.
- Unclear error rates. Vendors’ accuracy claims are often based on test conditions that do not match real use.
OpenAI discontinued its own AI text classifier in July 2023, citing its low rate of accuracy. When it launched, OpenAI reported that it correctly identified only 26% of AI-written text and wrongly labeled 9% of human-written text as AI-written.
What about watermarking?
A more promising approach is for AI providers to embed an invisible statistical watermark in generated text at the moment of creation. Google has developed SynthID, including a text watermarking method it made open source in October 2024 after testing it on about 20 million chatbot responses. Watermarks can help identify content from a specific participating model, but they only work if the provider applies them, and heavy editing or paraphrasing can weaken them. They cannot identify text from models that do not participate.
If you are a teacher or editor Never treat a detector score as proof. At most, it is a reason for a conversation. Decisions about misconduct should rest on evidence you can explain.
Better approaches for educators
- Set clear policies for each assignment about what AI use is allowed, as discussed in our students guide.
- Assess the process: drafts, notes, outlines and version history.
- Include in-class and oral components where students explain their work.
- Design assignments around personal experience, local context or class discussions.
- Teach responsible AI use rather than only policing it.
For editors and businesses
Focus on what matters to readers: accuracy, originality, sourcing and accountability. We made this argument in our column on AI-written content. A strong editorial process catches weak work regardless of how it was produced.
If you are wrongly accused
Ask what evidence the decision is based on. Share drafts, notes, research history and document version history. Offer to discuss your work in person. Point out the documented limitations of detection tools, and follow your institution’s appeals process.
Process beats detection
Today’s AI detectors are not reliable enough to decide anyone’s grade or reputation. Clear rules, attention to the writing process and a focus on quality are far better tools than a percentage score.
Sources
- AI detectors biased against non-native English writers, Stanford HAI, 2023
- New AI classifier for indicating AI-written text, OpenAI, updated July 2023
- SynthID, Google DeepMind
- Google DeepMind is making its AI text watermark open source, MIT Technology Review, October 2024
