Statistical assessment at time of scan — not proof of authorship. That line is printed on the face of every certification, and it belongs here too, before you read anything else. Our error rates are published on this site, so some of the certifications we hand out are going to be wrong, and you should know that going in rather than find it out later. It is also part of what makes a certification worth anything to the person who reads it, since it comes with a published record of how often the software behind it gets things wrong.
Scan a text now, for free Join the hosted-badge waitlist
Why you would bother
Your readers have got wary, and you can give them something to look at for themselves. On a feed that is full of machine-written articles every byline starts to look a little suspect, and yours does too, even if you wrote all of it yourself. A Cobalynx certification shows the reader what was scanned, when it was scanned, which version of the software did the work, and how often that version had got it wrong on our test set, so they do not have to take your word for it, or ours.
Nothing here makes money when a detector gets fooled. Cobalynx does one thing, which is to look at a text and say how likely it is to have come from AI, and then tell you how often that judgment was wrong when we ran it against a set of documents we had held back. We do not sell a humanizer, a paraphraser or a rewriting tool alongside it, and we are not in the business of producing the content these scanners are pointed at, so there is no side of this we win on when a scanner is beaten, ours or anybody else’s.
What Google has said about this
Google’s published guidance is that its ranking systems aim at “rewarding high-quality content, however it is produced”, and that “using automation—including AI—to generate content with the primary purpose of manipulating ranking in search results is a violation of our spam policies” (Google Search Central, Feb 2023). People do hope otherwise, but a Cobalynx certification does not change your rankings, since Google does not reward a page for having been written by a person as such. It is there for your readers, your editors and your clients.
How it works today
- You scan the text. You paste your final version, which can be anything from 150 to 10,000 words, into the free scan. There is no signup. An ordinary scan keeps nothing, as the text gets processed in memory and then it is thrown away, and the privacy page sets out how we prove that.
- You certify it, which is free and works right now. If the verdict came back likely human, then one click will turn that result into a certification. What you get is a report that stands on its own and carries the certificate id, the score, the date of the scan, the detector version and the SHA-256 fingerprint of the text, and along with it a PDF you can download and a watermarked preview of the badge with an embed snippet you can copy. On the free tier you can make 10 certification requests per day per IP, and we store nothing. If the verdict was anything else you get a plain explanation of why and the PDF scan report, and no badge, since a badge handed out on evidence we were not sure about is worth nothing to the reader it is meant for.
- What a free certification cannot do. There is nothing behind it on our side, which is what makes it free and is also what limits it. Nothing of yours lives at a Cobalynx address, so there is no page a reader can look the certificate up on, the badge you embed is the watermarked preview rather than one we serve live, and if you edit the article you scan it again from scratch. Hosting any of that would mean keeping your text on our servers, and keeping other people’s writing is not something to start doing casually, so it is a thing we would build deliberately or not at all. There is a waitlist further down if you want to hear about it.
The badge
There is one badge, which is a small SVG mark with the Cobalynx lynx on it, the verdict “Likely human”, the score and the date of the scan, and the alt text carries the score, the date and the never-proof line, so that a reader on a screen reader or reading a text-only page gets the same information you would. There is no badge at all for a result that came back inconclusive or likely AI. The score it shows is a calibrated human-likelihood figure out of 100, which is a different thing from a “percent-human” label on your authorship, and it is not a claim of proof. The preview you can embed today carries a “PREVIEW” watermark, and it is a plain image file, so once it is on your page it stays exactly as it is until you change it yourself.
The embed snippet is one image and one link, with no scripts in it. This one is an example, so you can see what it looks like:
<a href="https://cobalynx.com/badge" rel="nofollow"><img
src="https://cobalynx.com/badge/preview.svg?score=97&date=2026-08-25"
width="200" height="44"
alt="Checked by Cobalynx — likely human (score 97/100, 2026-08-25)"></a>
The snippet ships with rel="nofollow" on it by default, since
Google has asked for widget links to be nofollowed, and the badge is there
for the people who read you rather than for passing links around.
What it costs
Everything on this page is free today and no money has changed hands for any of it, and that covers the scans, the certifications, the PDF reports and the watermarked badge preview. The one thing we expect to charge for is the hosted certificate, since that is the one somebody has to keep online and answer for over years, and that costs real money to do properly. It does not exist yet, so there is no price list here. We would rather quote you nothing than quote you a number we might not be able to keep to. No payment is taken today.
How a certification is put together
- It is tied to a fingerprint. A certification covers one text and only one, which is the one whose SHA-256 fingerprint it has on it, and the recipe for working that fingerprint out is printed on the report as well. If the sum your reader works out from the text in front of them matches the one on the report, the report is talking about that exact text, down to the character. If it comes out different, the text has changed since it was certified and the report no longer covers what they are reading. That is the whole claim, and it is worth being exact about it, since a fingerprint sounds like it proves more than it does.
- It does not vouch for itself. We do not sign these, so nothing stops somebody editing the score or the date on a copy of one and leaving the fingerprint alone, and a reader could not tell from the document. The fingerprint tells your reader which text the report is about, and nothing about the report. The certificate id is worked out from the same fingerprint, so you can reproduce that yourself too, and it was never meant to be a secret. And none of it says anything about who wrote the text, which is something we never learn and never claim to.
- It is dated and it says which version did the work. A certification carries the date of the scan and the version of the detector that ran that day, so a text assessed in 2026 stays an assessment from 2026. When we change the engine in some way that matters, the old certifications go on saying which version looked at them, and we do not go back and rewrite them behind your back.
- If you edit, the coverage ends. The certification never covers a URL, only the text itself. If the article you published turned out to be different from the text you certified, the certification does not apply to the changed version, and the fingerprint is how a reader would find that out. Scanning the new version again is free and takes about as long as the first one did.
- You always see the same things on every artifact, which are the score, the date, the detector version, and the line “Statistical assessment at time of scan — not proof of authorship.”, with our published error rates one link away.
What certification stores
Ordinary scans are zero-retention, and so are free certifications: the certification report is self-contained, carrying the text’s SHA-256 fingerprint, score, date, and detector version, and nothing about it is stored on our servers. Hosted certificates, when they open from the waitlist, will be the one deliberate exception: creating one will mean asking us to store that exact text, its fingerprint, and its verdict so a public verification page can exist, and deleting the certificate will delete the stored text with it. That exception ships only after this policy states it in the same release.
How we prove that we keep nothing is written up on the privacy page.
Hear about hosted certificates
Hosted certificates are not open, and we are not putting a date on them, since we would rather tell you nothing than tell you a date we cannot keep. If you want to hear when there is something to hear, send us an email. There is no signup form and no mailing list, so it is a plain message from your own mail app. We write back when there is news, we do not use your address for anything else, and there is nothing to unsubscribe from because there is no list you were added to.
Common questions
What does a certification claim?
It claims that on the day of the scan, this text and no other was given a likely-human verdict by the Cobalynx scan, with the score that is shown on it. It is a statistical assessment that comes with published error rates. It is not proof that you wrote the text, we will never present it as proof, and every artifact we issue says that on its face.
Can a wrong certification happen?
Yes, it can, and we will tell you how often it happened when we tested it. We made our frozen evaluation set difficult on purpose, and on that set 16.5% of the texts we called likely human turned out to have come from AI, which was 126 of 764. If you think about what people put into the scanner in real life, where most of them did write the text themselves, a wrong certification will be a good deal rarer than that, but we are not going to tell you it will never happen, as some of them will be wrong. That is the reason every certification carries its score, its date and the detector version, links to these numbers, and claims nothing that the detector could not back up.
What if my certified text is edited later?
The certification covers the text it was issued for and it never covers a URL. So if you fix a typo or you rewrite a paragraph, the version you published will no longer match the fingerprint on the certification, and a reader who did the hash again would see that. The certification has not been invalidated by that, it just does not cover the new version. Scanning the edited text again is free and gives you a fresh certification for it, which takes about as long as the first one did.
Do you store my article when I certify it?
No, we do not. A free certification keeps nothing, the same as an ordinary scan, since the report stands on its own and there is nothing left on our servers once it has been made. Hosted certificates, which are on a waitlist, will be different, as a public verification page could only exist if we kept your text, and we would do that only when you had told us to. The whole storage story is written up on the privacy page.
Will the badge change my Google rankings?
No, it will not, and we would be suspicious of a salesman who promised you that it would. What Google has said is that it rewards helpful content however it was produced and that it penalizes scaled content abuse. The badge was made for the people who read you, and not for the ranking systems, which is why the embed link carries rel="nofollow" on purpose.
What if my text comes back inconclusive or likely AI?
Then you do not get a badge, and that is true of every result other than likely human. You can still download the PDF scan report if you want a record of what the scan said. On our evaluation set we came back with inconclusive on 5.8% of the documents rather than force a verdict we could not stand behind, and a badge that could have said “we were not sure” would take a little meaning out of every other badge.