AI Detector Accuracy, False Positives & Limitations
An AI detector result can help you decide where to look more closely. It cannot prove who wrote a document, confirm misconduct or tell you whether an author's use of AI was permitted.
Here is how to read AIScan results and assess an accuracy claim before relying on it.
How accurate is AIScan?
AIScan does not currently publish a validated overall text-detection accuracy or false-positive rate. App Store ratings and product usage are not substitutes for that evidence.
Look for an evaluation of the kind of writing you intend to check. Results on long English articles do not establish reliability on short emails, translated text or another language. When you see a headline percentage, ask what was tested and under what conditions.
False positives and false negatives
A false positive happens when human-written text is flagged as AI-generated. A false negative happens when AI-generated text is missed. Both matter, but they lead to different mistakes.
Consider an editor reviewing a contributor's article. A false positive could create an unfair suspicion about the writer. A false negative could leave undisclosed AI use unexamined. The editor still needs to check the sources, factual accuracy and agreed writing policy.
What does the AI score mean?
Read the score as an assessment of the passage you submitted. It does not verify a count of AI-written words, record the drafting process or establish the writer's intent.
Read highlighted passages in their original context. A quotation, a standard definition or a missing section may change what you need to investigate. If you submit an excerpt, keep a copy of that exact excerpt alongside the result.
Can a source estimate identify the AI model?
An optional source estimate cannot confirm which model, company or app produced the text. Do not use a source label as proof that a writer used a particular service. Several suggestions do not establish that several authors or tools were involved.
Detection and source attribution are separate questions. Our guide to model estimates explains the difference without treating a suggested source as proof.
Why text length and context matter
A short message gives a reviewer less context than a complete article. Mixed authorship, quotations, editing and translation can also complicate a review. A result for one kind of document should not be assumed to apply to another.
Copying from a PDF or an image can introduce missing spaces, repeated headings or broken lines. Check the extracted text before submitting it. Changing the input means you are checking a different passage, so keep the original available.
What independent research can tell us
Liang and colleagues found disparities in how the detectors they tested treated non-native English writing. Sadasivan and colleagues found that paraphrasing reduced detection rates in tested systems. These findings show why context and a fair review process matter. Neither study establishes an AIScan error rate.
Questions to ask about an accuracy claim
Look for a test date and detector version, sample counts, known origins for the documents, and the languages and lengths tested. Results should distinguish human writing, generated writing and edited or mixed writing.
Look for false-positive rates on human writing and detection rates on generated writing separately. One blended percentage can hide those differences. Small samples need care too: no errors in a handful of examples does not mean errors never occur.
A fair way to review a result
- Check that the submitted passage matches the original document.
- Read any highlighted section with the surrounding argument and sources.
- Review available drafts, notes and version history.
- Give the writer a chance to explain their process.
- Apply the relevant policy and consider the evidence together.
An image score also cannot prove ownership or establish whether a caption is true. Review the original source and supporting evidence.
Try the text detector, follow the teacher workflow, or contact support if a result needs attention. Avoid including confidential writing in a support message unless it is needed and you have permission to share it.