AI-generated text is simple to produce these days, and that makes it harder for teachers, editors and clients to know whether a piece of writing was written by a person or a model. A number of detectors exist to help — each uses different signals and each has limits. Here are five tools worth knowing about, followed by practical tips for using them sensibly.
Five tools to try
- OpenAI’s AI Text Classifier (historical example) - OpenAI released a classifier that labelled inputs on a likelihood scale (for example: very unlikely, unclear, possibly, likely). It was intended to flag likely machine-written passages, but tests showed it could be uncertain on short or mixed texts. Think of it as one signal, not a final judgement.
- Originality-style commercial detectors - There are commercial services that combine AI-detection with plagiarism checks and browser extensions for reviewers. These tools are aimed at publishers and agencies, and they typically run models tuned to spot patterns common in AI output. Because they’re paid services, they often include exportable reports and bulk-checking features.
- Content-at-scale-style percentage detectors - Some detectors return a single percentage or a “real vs fake” score for a submitted block of text. They’re easy to use and fast, and often work well on clear-cut examples. Keep in mind such scores depend on the detectors’ training data and thresholds, so they’re best used alongside other checks.
- GPT-2 / older-model detectors - Early detectors were built specifically to spot text from older models such as GPT-2. Those tools can be reliable for detecting older AI outputs but are less useful as models evolve. They remain handy if you suspect an older generator was used.
- GPTZero-style tools for educators - Some tools were developed with education in mind and use linguistic signals such as “perplexity” (how predictable text looks to a model) and “burstiness” (variation in sentence patterns). These indicators can catch boilerplate machine prose, though they’re not foolproof.
How to use detectors — and what they won’t tell you
- Use multiple signals. Run text through more than one detector and combine results with plagiarism checks and a close read of the writing itself.
- Check context. Look for an author’s drafts, timestamps, style consistency, references, and whether the work answers specific, local questions that a generic generator would struggle with.
- Watch short snippets. Many detectors struggle with very short pieces of text or with mixed human/AI edits. Length, editing and prompt engineering affect detectability.
- Expect uncertainty. No tool gives a definitive legal or academic verdict. False positives and false negatives happen — especially as models and detectors both evolve.
- Educate and verify. In classrooms and workplaces, set clear policies about permitted assistance, ask for drafts or explanations of process, and use oral checks or supervised tasks when assessment is critical.
Bottom line: detection tools can be a useful part of your toolkit, but they’re not a substitute for human judgement and good process. Treat their output as one piece of evidence, not proof on its own.
