Accuracy & Limitations

We believe that transparency about accuracy is not optional — it is a responsibility. If you are going to use an AI detection tool to inform decisions about student work, content quality, or hiring, you deserve to know exactly how reliable that tool is, and where it falls short.

This page provides our current accuracy benchmarks, false positive rates, known limitations, and guidance for responsible use. We update this information as our models improve and as new data becomes available.

Our Commitment to Transparency

Many AI detection tools publish bold accuracy claims without context — “99% accurate” headlines that collapse under scrutiny when you examine the testing methodology, dataset composition, or definition of “accurate.” We take a different approach.

Every benchmark we publish comes with context: what dataset was used, what text types were included, what constitutes a correct classification, and what the error margins are. We believe this level of disclosure is essential for any tool whose results may influence consequential decisions.

Our accuracy figures are derived from internal testing on held-out evaluation datasets that our models have never seen during training. We also cross-validate against publicly available benchmark datasets used in academic research on AI detection.

Current Accuracy Benchmarks

The following figures reflect our model performance as of our most recent evaluation cycle. Accuracy varies significantly depending on the type of content, the AI model that generated it, and whether the text has been edited after generation.

Detection Rates by AI Model

GPT-4 / GPT-4o generated text: ~92-95% correct classification on unedited text of 200+ words.
GPT-3.5 generated text: ~94-97% correct classification. Older models produce more detectable statistical patterns.
Claude-generated text: ~90-93% correct classification on unedited text of 200+ words.
Gemini-generated text: ~89-92% correct classification on unedited text of 200+ words.
Open-source models (LLaMA, Mistral): ~88-93% correct classification, varying by model size and fine-tuning.

Detection Rates by Text Type

Academic essays and research writing: ~93-96% accuracy. Structured academic formats provide strong signals.
Blog posts and articles: ~90-94% accuracy. Casual but structured content is well-suited to detection.
Business and professional writing: ~87-91% accuracy. Formulaic business writing can resemble AI patterns, slightly increasing false positives.
Creative writing and fiction: ~82-88% accuracy. Creative text is inherently more variable and harder to classify reliably.
Technical documentation: ~84-89% accuracy. Highly structured, domain-specific text poses challenges for all detectors.

Impact of Post-Generation Editing

Unedited AI text: Highest detection rates (figures above).
Lightly edited (minor corrections, formatting): Detection rates drop by approximately 3-5 percentage points.
Moderately edited (sentence restructuring, some rewriting): Detection rates drop by approximately 10-15 percentage points.
Heavily edited (substantial rewriting, added personal content): Detection rates drop by 20-35 percentage points. At this level of editing, the text reflects significant human authorship and may legitimately classify as mixed or human-written.

False Positive Rates

A false positive occurs when human-written text is incorrectly classified as AI-generated. This is the most consequential type of error, because a false positive could lead to an innocent student being accused of cheating or a legitimate writer having their work questioned.

Our current false positive rates:

Overall false positive rate: ~4-6% across all text types and demographics when using a 50% threshold.
Native English speakers, general writing: ~2-4% false positive rate.
Non-native English speakers: ~8-12% false positive rate. This is a known industry-wide issue. Non-native speakers often write in more formulaic, predictable patterns that resemble AI output. We are actively working to reduce this disparity.
Formulaic writing (legal, medical, technical): ~7-10% false positive rate. Highly structured professional writing shares statistical properties with AI-generated text.
Using a higher threshold (70%+) for “AI-detected” classification: False positive rates drop to ~1-3% for native speakers, though at the cost of more false negatives.

We strongly recommend that users consider the context and writer demographics when interpreting scores, particularly in educational settings where non-native speakers may be disproportionately affected.

Factors That Affect Accuracy

Understanding what affects detection accuracy helps you interpret results more responsibly. The major factors include:

Text Length

Longer texts provide more statistical data for analysis. Our confidence increases substantially with text length:

25-50 words: Minimum viable analysis. Results should be treated as rough indicators only.
50-150 words: Moderate confidence. Sufficient for preliminary screening but not for consequential decisions.
150-300 words: Good confidence. Most signal dimensions have enough data for reliable extraction.
300+ words: Highest confidence. This is our recommended minimum for decisions that matter.

Language and Domain

Our models are primarily trained on English-language text and perform best on standard American and British English. Detection accuracy may be lower for:

Text written in non-English languages (our models have limited multilingual support)
Highly specialized jargon-heavy text (medical, legal, scientific)
Dialect-heavy or highly informal writing
Poetry, song lyrics, and other non-prose formats

AI Model and Prompt Strategy

Not all AI-generated text is equally detectable. Text generated with specific instructions to “write like a human” or “vary your sentence structure” is harder to detect. Newer, more capable models generally produce text that is harder to distinguish from human writing. Texts generated with high temperature settings (more randomness) are also harder to detect.

Known Limitations

We disclose these limitations not because we are obligated to, but because we believe honesty builds trust and leads to better outcomes. Every user who understands these limitations will make better use of our tool.

No detector should serve as sole evidence. AI detection results are probabilistic estimates, not proof. They should never be the only basis for academic misconduct charges, employment decisions, content rejection, or any other consequential action.
Mixed content is inherently ambiguous. Text where a human used AI for a draft and then substantially edited it occupies a gray area. Our tool may score such text anywhere on the spectrum, and any result could be considered partially correct.
The technology will get harder, not easier. As AI models improve, the statistical gap between human and AI text narrows. We expect detection to become increasingly challenging over time, and we will be honest about declining accuracy if and when it occurs.
Demographic bias exists. Non-native English speakers face higher false positive rates across all AI detection tools, including ours. We are investing in research to reduce this disparity but have not eliminated it.
Short, formulaic content is unreliable to classify. A three-sentence email or a bulleted list does not contain enough stylistic signal for meaningful detection. We will not present a high-confidence score for text that does not warrant it.
Paraphrasing tools can defeat detection. Text that has been run through paraphrasing tools or heavily rephrased will often evade detection. This is a fundamental limitation of statistical detection approaches.

Best Practices for Interpreting Results

Based on our experience and research, we recommend the following practices:

Use detection as one signal among many. Combine detection scores with your own judgment, knowledge of the writer, comparison with previous work, and any other relevant context.
Apply appropriate thresholds. For high-stakes decisions, require a higher threshold (70-80%+) before considering text as potentially AI-generated. A score of 55% is not meaningful evidence.
Consider the writer’s background. If the writer is a non-native English speaker, a technical writer, or someone who naturally writes in a formal, structured style, account for the higher false positive likelihood.
Look at the signal breakdown. Check which signals (perplexity, burstiness, consistency) contributed most to the score. A text that scores high on all three signals is a stronger indicator than one where only one signal is elevated.
Allow the writer to respond. If you are using detection results in an evaluative context, always give the writer an opportunity to explain their process before drawing conclusions.
Re-test with different sample lengths. If possible, test multiple samples from the same writer. Consistent patterns across multiple samples are more meaningful than a single result.

Industry Context

To place our accuracy figures in context: independent academic research on AI detection tools generally finds that most commercial detectors achieve between 70-95% accuracy on unedited AI text, with false positive rates ranging from 2-15% depending on the tool and the text population. Our performance falls within the upper range of these benchmarks, but we want to be clear that no tool in the industry has solved this problem completely.

Notable research papers, including studies from the University of Maryland, Stanford, and others, have repeatedly demonstrated that all current AI detectors share similar fundamental limitations, particularly regarding edited content, non-native speakers, and newer AI models. Our tool is no exception.

Our Ongoing Improvement Process

Accuracy is not a static achievement — it requires continuous investment. Our improvement process includes:

Regular model retraining: We update our classifiers when new AI models are released to ensure we can detect the latest generation of AI text.
Dataset expansion: We continuously expand our training and evaluation datasets to cover more text types, domains, and writing demographics.
Bias auditing: We regularly audit our models for demographic bias, particularly false positive disparities affecting non-native English speakers, and prioritize reducing these gaps.
Benchmark publication: We aim to publish updated accuracy benchmarks at regular intervals so users always have current information about our performance.
Community feedback integration: We review user reports of incorrect classifications and use them to identify systematic issues in our models.

If you have feedback about our accuracy or have encountered a result you believe is incorrect, we encourage you to contact us. Real-world feedback is one of our most valuable tools for improvement.