AI Detection Research & Benchmarks

Accuracy varies significantly by tool, document type, and test conditions. On academic writing specifically, our sentence-level system achieves a 0.2% document-level false positive rate, with 97.6%-99.9% accuracy across modern LLMs – documented in our Sentence-Level AI Detection for Academic Integrity study.

False positives occur when human writing shares statistical properties with AI output – particularly low perplexity and low burstiness. Technical writers, non-native English writers, and writers who follow rigid structural formats are most at risk. Stanford researchers (Liang et al., 2023) found seven commercial GPT detectors misclassified TOEFL essays by non-native English speakers as AI-generated at an average rate of 61.3%, with at least one detector flagging 97.8% of those essays. We addressed this specifically with iterative hard negative mining on TOEFL/IELTS corpora and formulaic academic prose – dropping our non-native English false positive rate to 0.5%, compared to 6.4% for an equivalent document-level baseline. Proofademic is the only AI detector calibrated specifically for academic writing across native and non-native English populations.

Detection accuracy drops significantly after paraphrasing across all detection paradigms. Krishna et al. (2023) demonstrated that DIPPER-paraphrased text dropped DetectGPT’s accuracy from 70.3% to 4.6% at a 1% false positive rate. Proofademic’s sentence-level analysis offers more resilience than perplexity-only methods – per-sentence stylometric signals are harder to uniformly disguise than document-wide statistics – but we treat humanizer robustness as an active area of work, not a solved problem. We disclose this explicitly in our published study’s limitations section.

Document-level scores collapse a whole submission into one number (e.g., “68% AI probability”) – useful when a document is entirely AI or entirely human, but uninformative when a paper is part-student-part-AI, which is how students actually use these tools. Sentence-level scoring flags specific passages, which lets an educator have an evidence-based conversation with the student rather than rely on a binary verdict. Proofademic was designed for this from the start – every submission returns a per-sentence breakdown with confidence scores, calibrated for academic writing. The same signal is documented in our published study and available to try free.

Frontier LLMs produce text that’s statistically closer to human writing than earlier models, making detection harder industry-wide. Proofademic was trained on outputs from eight current frontier LLMs across the major providers – OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, Alibaba, and xAI – with two of those held out from training as blind-tested models to measure generalization. Per-LLM detection accuracy ranged from 97.6% to 99.9% on our internal evaluation, with only a ~2 point accuracy drop on the unseen models. That generalization suggests the system learns general distributional properties of machine text rather than memorizing model-specific artifacts, which buys us time between retraining cycles. We retrain on a rolling basis as new frontier models reach commercial adoption.