The review desk · Practical guides
Why AI detectors disagree
Two results can disagree without either one establishing who wrote the text. Start by checking whether they even answer the same question.
Compare like with like
A detector applies a model to a particular input. A writing-pattern report describes features such as repetition or sentence length. In CouldItBe, GPTZero screening and local observations are separate outputs. A repetitive passage is not a second detector vote.
Before comparing tools, record the exact passage, language, date, provider and model version when available. An article with its navigation and footer included is a different input from its body text. A revised draft is a different input from the original.
GPTZero distinguishes a classification from confidence in that classification. A confidence score is not the proportion of words written by AI. See GPTZero’s result explanation. Do not average unlike scores into a new “truth percentage.”
A worked review: an edited draft
Original illustrative scenario · No measured scores
An editor receives a contributor’s revised article. One tool flags AI-like writing; another favors human writing. The local report also finds repeated phrases. The contributor says they edited an AI-generated first draft.
What should the editor conclude?
The disclosed workflow already supplies useful context. The detector disagreement does not settle whether the final article meets the publication’s rules. Check those rules, ask for the draft history, verify sources, and document the disclosed assistance. Repetition alone cannot distinguish a template, a quotation or AI output.
What would make the comparison unfair?
Scanning different revisions, using a short excerpt in one tool and the entire article in another, or comparing a sentence score with a document score. Keep the input identical and preserve uncertain or unavailable results rather than dropping them.
This exercise explains a decision process. It is not a benchmark, a real contributor case, or a claim that a particular tool failed.
Why context and sample size matter
Different detection methods can use different signals and decision rules. Research by Chakraborty and colleagues explores how the amount of text needed for detection depends on the separation between human and machine text distributions. That finding supports caution with thin evidence; it does not establish the accuracy of today’s product on your draft. Read the original research.
Our own small writing evaluation also illustrates the need for careful labels. Both assignment groups produced a mixture of classifications. Assignment instructions did not independently verify authorship, so those counts cannot honestly become an accuracy rate.
A useful next step
- Preserve the input and each completed report, including uncertainty and version details.
- Check whether the outputs refer to the same text and the same type of assessment.
- Ask for drafting history, sources and an explanation of any declared assistance.
- Apply your publication policy and fact-check the content independently.
A low AI score does not validate facts or clear plagiarism. A high score does not prove misconduct. If the evidence remains mixed, record that uncertainty.
Keep the evidence in context
These guides support a review, not a finding of misconduct. Read our methodology and actual test results, known limitations, and privacy policy. Report a correction if a source or product detail changes.
Browse all resources →