For Universities

    How do you keep the bar when they'll use AI anyway?

    Knowing whether a student used AI tells you little — most of them did. How they used it shows what they understood. Cognity records the whole session — every prompt, every rejection, every revision — so that is what you grade.

    Ban it or allow it. You still cannot tell what they know.

    Most departments have picked one and are living with the cost.

    If you ban it

    The semester turns into a detection operation.

    Detector scores, accusations you cannot prove, and appeals to hear.

    If you allow it

    You cannot tell a good student from a good prompt.

    Polish is what the model does best, so the best-reading essay tells you least.

    Both are answers to one question: did they use AI? Even a certain answer would not tell you who learned anything.

    Detectors fail — and they fail your international students first

    Accuracy figures for AI detectors are mostly vendor-reported and conditional on how much of a document gets flagged. The more useful question is who the errors fall on. When researchers tested seven commercial detectors on essays by non-native English writers, every one of them misclassified that writing as AI-generated, while classifying native-speaker writing correctly.

    A university with a large international intake is exactly where that failure concentrates on the same students, semester after semester — and where an appeal against a probability score is hardest to resolve fairly. Whatever else the ban costs, it also costs those students disproportionately.

    Liang et al., Patterns (Cell Press), 2023 — US eighth-grade essays vs TOEFL essays.

    Ask how they used it, not whether.

    A finished document does not remember how it was made. That is the only reason "how" ever looked unanswerable — and it stops being true the moment the document is not the only thing you collect.

    So collect the making. Cognity is the workspace the student writes in, and it keeps the session: every prompt, every reply, every revision, in the order it happened.

    Five things you can put in a rubric

    This is what "how" looks like on a rubric row. Not proxies for understanding — the places where understanding becomes visible in the way someone works, the same things you would look for sitting beside them while they wrote.

    01

    Did the prompt show they understood the problem?

    A student who can frame the task precisely has already done the hard part. A vague prompt and a sharp one come from different levels of grasp, and both are in the record verbatim.

    02

    Did they notice when the AI was wrong?

    The most useful thing in the transcript is often a rejection: the moment a student pushes back on a confident, wrong answer. You cannot fake that, and it is difficult to do without knowing the material.

    03

    Did they verify, or did they accept?

    You can see whether a claim was checked against a source before it entered the draft, or taken on trust. That distinction is most of what academic judgement is.

    04

    Whose reasoning shaped the argument?

    The record shows where the structure came from — where the student set direction and where they followed. An argument the student steered reads differently from one they received.

    05

    How much of the final text is their own wording?

    Measured by verbatim n-gram overlap with the AI's output, so the figure points at particular sentences you can read, not at a probability about the document. It is evidence for a conversation, not a verdict.

    And you can teach the skill they are hired for

    Working well with AI is not a loophole around competence; it is turning into a competence in its own right. Framing a problem so a model can help with it, recognising a plausible answer that is wrong, knowing which part of the job to keep — employers are already selecting for this, and few graduates arrive having been taught it, because most were told not to practise it.

    A visible process makes it teachable. You can set the AI's role per assessment, show students what good use looks like, and assess them on it — rather than leaving them to work it out privately, badly, and against the rules.

    What this does not do

    It doesn't prove a negative.

    A student can draft in another tool and paste it in. What the record shows then is a submission with no working behind it, which is a signal worth raising — but Cognity does not call it proof, and neither should a panel.

    It doesn't measure understanding directly.

    Nothing does. It gives you evidence about how the work was made, next to the work. The judgement is still an academic one, and it stays with you.

    It doesn't grade for you.

    Cognity summarises and surfaces; it does not decide. There is no dishonesty score, and no threshold that flags a student on your behalf.

    It fits the procedure and the systems you already run

    None of the above is worth much if it means a parallel system, a separate gradebook, or a new misconduct process.

    • For a misconduct hearing: the session in order, with the point at which a specific sentence entered the draft — an exchange a panel can read, rather than a percentage the student disputes
    • Per-assessment AI settings, from no assistance to full drafting support, enforced by the tutor at the point of use and recorded with the submission
    • Summaries first for a large cohort — you open the full record only where the summary gives you a reason to
    • LTI 1.3 launch from any LMS that supports it, on every plan; roster from the LMS, grades and record back with the submission
    • Institutional SSO, AWS hosting in Korea with other regions under an enterprise agreement, a DPA, and no model training on student work

    Who brings this to a department

    Course instructor

    Assessment that still means something when every student has a chatbot open — without going back to supervised handwritten exams.

    Director of learning and teaching

    One AI position the department can actually hold, across courses that reasonably want different rules.

    Academic integrity officer

    Fewer cases built on a contested percentage, and evidence a panel can read in the ones that remain.

    Curriculum and careers lead

    AI fluency taught and assessed on purpose, rather than picked up by students in spite of the rules.

    What universities ask first

    Policy, procedure, cohort size and systems — the questions that come before a pilot.

    Try it on one assessment

    One course, one cohort, inside your own LMS. By the time grades are due you will know whether "how did they use it" was the question you should have been asking all along.