AI Text Detection Accuracy — What the Research Says

Overview¶

This page summarizes publicly available research on AI text detection accuracy published between 2023 and 2026. The goal is to give writers, educators, and researchers a clear-eyed picture of what these tools can and can't do.

Key Findings from Published Research¶

Detection Accuracy Varies Widely by Model and Domain¶

Research published in 2023 by Mitchell et al. ("DetectGPT") and follow-up studies showed that detection accuracy depends heavily on: - Which LLM generated the text (GPT-4 is harder to detect than GPT-3.5) - The domain (scientific/technical writing is harder to detect accurately) - Whether the text was edited post-generation

For text generated without any editing, leading detection tools achieve 85–95% accuracy. For text that was lightly edited by a human, accuracy drops to 60–75%.

False Positive Rates Are Significant¶

Multiple studies have found that AI detectors flag human-written text as AI-generated at rates between 3–20% depending on the tool and writing style. Academic writing in technical fields is particularly prone to false positives because it naturally has lower perplexity and burstiness.

A 2024 study found that international students whose first language was not English were disproportionately flagged by AI detectors — their writing style, which tends toward more formal and less idiomatic language, shares statistical properties with AI-generated text.

Paraphrasing Significantly Reduces Detection Rate¶

Research consistently shows that paraphrasing or rewriting AI-generated text substantially reduces detection tool accuracy. Studies testing paraphrase attacks on detection tools found that accuracy dropped from 85%+ to below 50% after a single paraphrase pass.

This doesn't mean detection is useless — it means it's probabilistic, not definitive.

Implications for Educators¶

AI detection scores should be treated as one signal among many, not as definitive proof of AI use. Educators using detection tools should: - Treat flagged content as warranting a conversation, not automatic punishment - Be aware of the false positive rate for their student population - Combine detection results with assessment of whether the content matches the student's known capabilities

Implications for Writers¶

AI detectors are imperfect tools. A low AI-probability score doesn't guarantee your writing will be perceived as human. A high score doesn't mean you used AI. The scores are statistical estimates, not ground truth.

References¶

• Mitchell, E. et al. (2023). "DetectGPT: Zero-Shot Machine-Generated Text Detection Using Probability Curvature." arXiv:2301.11305
• Sadasivan, V. et al. (2023). "Can AI-Generated Text be Reliably Detected?" arXiv:2303.11156
• Liang, W. et al. (2023). "GPT Detectors Are Biased Against Non-Native English Writers." arXiv:2304.02819
• Chakraborty, S. et al. (2023). "On the Possibilities of AI-Generated Text Detection." arXiv:2304.04736