The trust problem this page answers
Tools in this market commonly advertise “99% accuracy” while independent tests measure far less; some publish no error rates at all; a few even sell, on the same site, the paraphrasing tools that defeat their own detector. Meanwhile real people — students, job applicants, writers — get punished on the strength of a single unexplained percentage.
Cobalynx was built as the opposite bet: that the most useful detector is the one that tells you exactly how often it is wrong. We measured our system once, on a frozen evaluation set, and committed the results — the flattering numbers and the unflattering ones. Every claim below links to the page where you can check it.
Designed to minimize false accusations, not maximize accusations.
Twelve questions to ask any detector
“A typical detector” below is not one product — it is the pattern across the market's best-known tools as of August 2026. Where we describe ourselves, we link the evidence.
| The question to ask | A typical detector | Cobalynx | Check it |
|---|---|---|---|
| What accuracy is claimed? | A single hero number, often “99%” | No single accuracy number. Error rates with 95% confidence intervals, per situation | /evidence |
| How often is human writing falsely flagged? | Rarely stated up front | Measured: 0.5% of human texts that received a verdict (3 of 609; CI 0.2%–1.4%); 0.2% on native-speaker writing (1 of 491) | /evidence |
| What about non-native English writers? | A documented industry bias, rarely self-measured | Measured and published: 1.7% (2 of 118) — higher than for native writing, and we say so | /evidence |
| What happens on borderline text? | A verdict anyway | “Inconclusive” — 16.5% of our evaluation set (251 of 1522). We abstain rather than guess | /methodology |
| What does the score mean? | An unexplained percentage | A population-scoped, measured score: on our frozen evaluation population, texts scoring like yours were AI about that often — with the base-rate arithmetic done for you | /methodology |
| Does “likely human” prove anything? | Often implied | No. On our evaluation set, 26.9% of texts we called likely human were actually AI (223 of 829) — and we publish that number | /evidence |
| Paraphrased (“humanized”) AI text? | Silence | A published weakness: 33.5% recall on paraphrase-attacked text (73 of 218 with a verdict) | /evidence |
| Texts too short to judge? | Scored anyway | Refused below 150 words — short texts are where detectors fail worst | /methodology |
| Evidence beyond writing style? | Style analysis only | File provenance too: Content Credentials (C2PA) and document metadata — evidence carried by the file, not inferred from prose | /check-image-content-credentials |
| Is a bypass/paraphrasing tool sold next door? | Sometimes | Never. No humanizer, no paraphraser, no bypass tooling — ever | this page |
| Is your text stored? | Often unclear; signup walls are common | No signup. Zero retention: scan text is processed in memory and discarded | /privacy |
| Who verified these numbers? | “Verified by studies” (selectively cited) | Honestly: self-measured and reproducible; no third party has audited our numbers yet. Standing invitation: contact@cobalynx.com | /evidence |
What we ask of you
Don't trust this page — check it. Every number above traces to our evidence page, which documents the frozen evaluation set (1,522 documents, August 16, 2026), the thresholds we operate at, and the commands that reproduce the run. And whatever detector you use, ours included: never treat a score as sole evidence to punish someone. If you've been flagged, start here. How the scoring works, end to end, is on the methodology page.
Who builds this
Cobalynx is built and maintained by Mario Federico, CEO of Roviant S.r.l. — working in AI since 2020, with over 15 years of software engineering experience.
One person is accountable for every number on this site, and every number traces to a versioned, reproducible evaluation run rather than to a marketing department. When something doesn't hold up, the correction is published under the same name that made the claim.