AI-detection research needs tools that report honestly rather than defensively. We offer free API access for non-commercial academic work, sentence-level data for analysis, and a standing invitation to evaluate us independently — including publishing findings that do not flatter us.
Free API access for non-commercial academic research, with limits raised beyond the standard free tier for the duration of a study. Requests are handled individually rather than through a self-serve form, so we can size the allowance to the work.
During development we tested ScribeLens against published, peer-reviewed journal articles and traced every case where a genuine paper was misclassified back to its cause — real academic writing, not a synthetic benchmark. That work is what drives the exclusion rules and the sentence-level weighting described on this page, and we can walk a research team through the methodology in more detail than the public documentation carries.
What we don't publish is a headline accuracy figure: no single number describes detection honestly across text length, subject, and register. Instead, every verdict ships with the sentence-level evidence behind it — reason keywords and a graded confidence tier — so a reviewer can audit the result directly rather than take a number on faith. Detection is probabilistic pattern analysis, and we say so plainly: never proof of authorship, in either direction.
Disciplined academic writing shares surface features with machine text: consistent structure, measured transitions, and even sentence rhythm. ScribeLens is built with that in mind.
Two design decisions address this directly. Reference lists, in-text citation apparatus, publisher boilerplate, and declaration sections are detected and excluded before scoring, because citation formatting is the most formulaic text in any paper by construction. And sentence-level evidence is weighed against document-level style signals, so a paper whose sentences read human is not convicted on register alone. Both behaviours are observable in the API response, not just asserted here.
If you are benchmarking detectors, we would rather be measured than marketed. We will support studies that include ScribeLens on the same terms as any other tool, and we do not ask for review rights over findings.
We are particularly interested in studies of non-native English writing and formal academic registers. Detector v4.0 was retrained on real student essays and on essays by non-native English writers, and in our own testing it is markedly better at recognising such writing as human. Independent measurement of that is welcome.
Email support@scribelensai.com from an institutional address with a short description of the study, the approximate volume of text involved, and where results are expected to appear. We ask only that published work cites ScribeLens and states the version used, which appears on every report.
Access is granted for non-commercial academic research. Commercial benchmarking, model training on our outputs, and resale of API results are not covered.
Yes, for non-commercial academic research. We raise the allowance to suit the study rather than applying the standard free-tier limit. Commercial benchmarking and resale are not included.
Because no single figure can describe detection honestly — results depend on text length, subject, and register. During development we tested ScribeLens on published, peer-reviewed journal articles, and we'd rather show the sentence-level evidence behind each verdict and let independent evaluation speak for itself.
Yes. We do not ask for review rights, approval, or advance sight of findings. Studies of academic and non-native English writing are especially useful to us.
A document-level verdict with its confidence tier, per-sentence classifications with scores, the reason keywords behind flagged sentences, and the exclusion record showing which reference and declaration sections were set aside before scoring.
Yes. A report can be given a verification link, which resolves to the stored copy of that exact analysis, so a result cited in a paper can be checked by a reader. Stored copies expire after 180 days by default; tell us if a study needs them kept longer.
The fastest way to judge a detector is to test it on writing you already know the provenance of. Without an account a scan covers 2,000 words; a free account scans a whole paper.