AI detector accuracy: what a score can and cannot tell you
An AI detector estimates whether writing resembles patterns in its training data; it does not observe who wrote your essay. Accuracy depends on the model, passage, language and evaluation method. Use a result to start a review, not to prove misconduct or to certify a draft as human-written.
How do AI writing detectors work?
A detector is a classifier. It analyzes a supplied passage and estimates which learned category it most resembles. Predictability and sentence variation are useful introductory concepts, but modern systems are not simply lists of forbidden words. GPTZero describes an end-to-end deep-learning system with document and sentence classifications.
This distinction matters when revising. Removing a transition such as ‘furthermore’ does not establish authorship, and keeping it does not establish AI use. A student can write predictable sentences, while an AI system can produce varied prose. Neither a word list nor a detector has access to your intentions, research notes or course permission.
Drawn from [1]
Does an AI probability mean that much of my essay was generated?
Not necessarily. Model confidence in a class and the estimated proportion of generated writing are different measurements. GPTZero explicitly distinguishes these concepts. Cortex’s free detector integration reads the API’s AI-class probability when that field is available. That is not a count of copied words, a plagiarism percentage or the chance of receiving a particular grade.
Keep the provider’s own definition beside any result you save. A document classification and a sentence highlight may summarize different levels of evidence. Do not turn several highlighted sentences into an exact claim about how a person composed the document.
Drawn from [1]
What are false positives and false negatives?
A false positive labels human writing as AI. A false negative misses generated writing. A tool can reduce one type of error while increasing the other, so a headline accuracy figure is incomplete without the test set, decision threshold and both error rates.
Consider a hypothetical review of a class in which very few students used prohibited assistance. Even a relatively small false-positive rate can produce disputed flags among the much larger group who wrote their own work. This is a reasoning example, not a Cortex benchmark or a prediction about your class.
Why do language, passage length and editing matter?
Stanford researchers documented bias against non-native English writing in a study of detectors available at the time. That study is evidence of a failure mode, not a current error-rate estimate for every product. GPTZero now describes its own bias-reduction work; vendor improvements still need evaluation on writing similar to yours.
A short quotation, a lab procedure and a long reflective essay supply different kinds of text. Models and thresholds also change. A result from a short excerpt on one date should not be treated as interchangeable with an institution’s full-document analysis on another date.
How should I judge an accuracy claim?
Read the methodology before comparing percentages. Tests that use different source material, AI models or scoring rules are not a fair leaderboard. Check whether the evaluator sells a competing product, whether the test includes edited and mixed-authorship writing, and whether the tested model version is disclosed.
Cortex has not published an independently validated detector benchmark. We do not claim an accuracy rate, a winning score against other detectors or a guaranteed result after rewriting. A provider’s published rate cannot be transferred to Cortex’s entire workflow.
- Identify the detector version and testing date.
- Look for human, AI and mixed-document examples.
- Separate false positives from missed AI text.
- Check language, length and subject coverage.
- Ask whether the test set and scoring method are reproducible.
What should I do if my own writing is flagged?
Keep the original draft rather than repeatedly rewriting it to chase a lower number. Gather genuine notes, source annotations and version history. Read the assignment policy and ask your instructor which passage raised a concern and what review process applies.
Explain the actual steps you took, including any assistance you used. Do not create artificial typing records or retroactive evidence. If the wording needs improvement, revise for clarity and accuracy separately from the authorship discussion. A lower detector score does not resolve a policy violation; a higher one does not establish it.
- Save the unchanged draft and the complete result.
- Collect authentic notes and earlier versions.
- Check the syllabus and ask about the review procedure.
- Explain your work honestly and disclose assistance when required.
Common questions
Can a detector prove I wrote an essay?
No. A detector evaluates the submitted text. Genuine drafting records, sources and your explanation of the work provide context it does not have.
Should I keep rewriting until the score is zero?
No. Repeated score-chasing can damage meaning and does not prove authorship. Review the argument, evidence and course policy instead.
Sources
These were read for this revision. A provider's description of its own product is an attributed claim, not independent certification. Check the current documentation before making a purchase or an academic decision.