GPTZero and Copyleaks: an honest review for educators

GPTZero and Copyleaks are two of the most frequently mentioned AI detection tools in education. They're genuinely different products with different histories and emphases, but they share a common methodological foundation - and a common set of limitations that matter significantly in school and university settings. Here's an honest assessment of both.

GPTZero: background and method

GPTZero was launched in January 2023 by Edward Tian, then a student at Princeton University, in direct response to the rapid spread of ChatGPT in academic settings. It attracted significant attention quickly and became one of the first widely used AI detection tools in education. The timing was important: it arrived before institutional policy frameworks had formed, which gave it substantial early adoption among schools and universities looking for any available response.

The tool works primarily on two measures: perplexity (how predictable each word choice is) and burstiness (the variation in perplexity across different sections of the text). The underlying premise is that AI-generated text tends toward consistent, high-probability word choices, whilst human writing varies more. GPTZero has evolved substantially since its initial release, adding features for institutional use and refining its underlying scoring models.

GPTZero: what the evidence says

GPTZero's detection performance on clearly AI-generated text is reasonable in benchmark conditions. The more significant concern for educational use is its false positive rate - specifically for non-native English writers. The same characteristics of careful, formal non-native writing that trigger false positives in other text-based detectors also affect GPTZero, because it relies on the same fundamental perplexity-based approach.

The general finding from multiple independent studies of text-based detectors - that ESL writers are flagged at significantly elevated rates - almost certainly applies here, as it does for any perplexity-based tool. For schools with meaningful proportions of international students, this is a significant practical concern that shouldn't be minimised.

Copyleaks: background and method

Copyleaks is a longer-established product, founded in Israel in 2015 as a plagiarism detection tool and subsequently expanded to include AI content detection. Unlike GPTZero, which started specifically as an AI detector, Copyleaks positions itself as a combined academic integrity platform covering both plagiarism similarity checking and AI generation detection.

The AI detection component uses text analysis to estimate AI generation likelihood. The combination with plagiarism detection makes it an appealing single-platform option for institutions that want both capabilities in one tool. It offers API access and integrations with several learning management systems, which eases deployment in institutional settings.

Copyleaks: what the evidence says

Copyleaks has published accuracy figures, but as with any vendor-provided data, these should be contextualised. Third-party testing has found similar limitations to other text-based tools: accuracy on clearly AI-generated text is reasonable, but false positive rates for non-native writers and for content that mixes AI assistance with human revision are a meaningful concern.

The combined plagiarism and AI detection offering is genuinely useful from a workflow perspective, but the AI detection component doesn't resolve the fundamental methodological limitations that affect the whole text-analysis category. The plagiarism detection is the stronger and more established part of the product.

The shared limitation

GPTZero and Copyleaks share a fundamental methodological constraint: both rely on text analysis to detect AI generation. This means they both face the same core problems: accuracy that erodes as models improve; false positives for non-native and EAL writers; vulnerability to paraphrasing circumvention; and probabilistic rather than factual output.

Neither tool can tell you how a piece of work was produced. They can only tell you something about how the finished text reads statistically. For making fair, defensible decisions about specific students, this distinction matters. A score from a text-analysis tool answers a narrower question than most teachers think they're asking.

What process-based detection offers instead

The most significant practical alternative to text-analysis approaches is process capture - recording the writing session itself rather than analysing the output. Tools that capture typing behaviour, paste events, and session timing during submission give you evidence about authorship rather than text patterns. This is a categorically different kind of information.

For the specific concerns about GPTZero and Copyleaks - false positives for ESL writers, vulnerability to paraphrasing, degrading accuracy over time - process-based detection largely avoids these problems. It doesn't read the text, so it can't flag students on the basis of writing style. It doesn't care whether the text was paraphrased. And the signals it produces - observed events with timestamps - don't erode as AI models improve.