Call for benchmarks
A detector should be judged by people who don't sell it. We fund independent evaluations of our text, image and video detection with free API credits.
We publish our own numbers, with the data behind them. But we run those tests ourselves, on sets we chose, and new models ship every few weeks. A detector that looked strong in spring can fall behind by autumn without anyone noticing.
A simple AI-or-human score also hides most of what matters in practice: who gets wrongly flagged, how much editing it takes before a text stops being caught, what survives a WhatsApp forward. Teachers, editors, fraud teams and investigators make decisions on these results. They deserve evidence from someone other than us.
A study can take on one of these or several. Proposals outside the list are welcome too.
Text
How well do detectors catch text from the newest models, and what happens when that text comes out of a deep-research run, an agent harness or a long multi-turn session rather than a single prompt?
Text
Real documents are rarely all human or all AI. We want studies across the whole range: a grammar pass, a rewritten paragraph, an outline filled in by a model, full co-writing, and text run through humanizers.
Text
Sliced results by language, non-native writing, region, dialect, education level or discipline. A study that finds no bias is as useful to us as one that finds it.
Image & video
Detection of the latest image and video models after what real distribution does to a file: messenger recompression, screenshots, crops, re-uploads and deliberate laundering.
Image & video
How often genuine photos and clips get flagged: phone cameras, professional and edited photography, scans, old footage, faces of every background.
Image & video
Face swaps, identity-preserving generation and synthetic ID or profile photos, the cases behind impersonation and fraud.
Send a short email. A few paragraphs are enough; we answer within a week. Tell us: