For teachers who are tired of adjudicating probability scores.

Detector accuracy gets most of the attention, and it deserves some. A 2023 study from Stanford found AI detection tools flag 61% of writing by non-native English speakers as AI-generated, against roughly 3% for native speakers. If a chunk of your class writes in a second language, a detector-led policy is not applying the same standard to everyone in the room.

But suppose the tool were excellent. You still receive a number with no reasoning attached. You cannot see which passage triggered it. You cannot rerun it under conditions you choose. The student cannot rebut it except by saying they wrote the thing, which is what anyone would say. You end up weighing a classifier's confidence against a person's word, and there is no version of that conversation you are equipped to win.

The result is a role most teachers did not sign up for. Somebody has to open the meeting, and it is you, and the only evidence in the room is a percentage neither party can interrogate.

The strongest detectors publish false positive rates around 1 in 10,000, backed by outside evaluation. Taken on its own that is impressive. The trouble starts when you multiply it by usage.

Arvind Narayanan ran the arithmetic. A student turns in something like 500 to 1,000 pieces of written work over four years. Scan all of them at a 1-in-10,000 error rate and 5 to 10 percent of the student body gets falsely accused at some point in their degree. Nothing is wrong with the tool. The volume does the damage.

Institutionally that number is small. For the individual student it is a formal integrity process built on evidence nobody can examine, at an age where the outcome follows them. Turnitin is careful about this in its own guidance: it states that it does not make a determination of misconduct and that the indicator is data for an educator's judgement, not a finding. It is worth reading what your vendor actually claims before treating their output as a verdict.

The alternative is to stop asking whether a finished document looks suspicious and start asking students to show how it was made. Version history does some of this already, which is why so many instructors have started asking for Google Docs revision logs. Certification does the same job in a form you can check in a few seconds instead of scrolling through hundreds of edit states.

In practice it is one line in the brief. Something like: drafts for this assignment must be written in Humanums and submitted with the verification link. Students write as normal, the session records keystroke cadence, pause structure, revision depth, and paste behaviour, and the finished draft gets certified. What they hand you is the paper plus a URL.

You open the URL. It shows the certificate, the signals behind the score, and whether the text has changed since certification. That takes a few seconds per submission and requires no software on your side, no LMS integration, and no administrator to approve anything.

The shift is in who does the work. Instead of screening every paper and initiating conversations about the ones that look odd, you set a requirement that applies to the whole class in advance. Nobody is singled out, and the evidence arrives with the work.

Announce it before the assignment, not after a paper looks wrong to you. A requirement introduced mid-term reads as a response to a specific student even when it is not.

Apply it to everyone. Selective monitoring recreates exactly the problem detector bias causes, except now the selection is yours to justify.

Publish an alternate route and mean it. Some students dictate, some use assistive technology, some are writing on a borrowed phone, some have connectivity that will not hold a session. An oral defence, a supervised in-class write-up, or a submitted outline plus a short conversation all serve the same purpose. Make the alternate route ordinary rather than something a student has to negotiate.

And be explicit about the negative case: a missing certificate is not evidence of misconduct. It is a missing piece of coursework, handled the way any missing component is handled. If a certificate becomes something students can be punished for lacking, you have rebuilt the accusation machine with extra steps.

Students do not need a paid plan. The free tier includes three certifications a month, public verification pages, and a portfolio page, which covers a typical course load for most students without anyone paying anything.

The Publisher plan is for the staff side. It runs $29 a month, or $24 a month billed annually at $288 a year. It includes up to five team members in total, counting the account owner, so a course team of four plus a coordinator fits. Invitations occupy a seat while they are pending and expire after seven days if unused. Everyone on the team gets Publisher benefits on their own account for as long as the owner's subscription is active.

On top of that it adds unlimited certifications, custom branded badges, bulk certification for handling a set of documents at once, and API access if your department wants to pull verification results into an existing system. Full details are on the pricing page.

If you need more than five seats, that is a conversation rather than a checkout, and it is worth having before you commit a whole faculty to a workflow.

A certificate shows that the certified text was typed by a human, with human pacing, pauses, and revisions, inside a monitored session, and that it has not changed since. That is the specific claim, and it is worth being precise with students about where it stops.

It does not establish which human. If a student pays someone to sit at the keyboard, the session is genuine and belongs to the wrong person. It does not prove no AI was consulted for research, structure, or feedback along the way. It does not speak to originality, and it is not a plagiarism check, so it does not replace whatever you already use for that.

It also cannot be applied to work that already exists. There is no way to certify a finished document after the fact, which means this only works if the requirement is set before the writing starts.

Used honestly, it is one piece of evidence that arrives with the work instead of an accusation that arrives after it. That is a smaller claim than most tools in this category make, and it is the reason it holds up.