Originally published at ictlms.net
On 16 September the Florida State Board of Education approved statewide rules on AI in education, with institutions given until 1 July 2027 to adopt and implement policies. Buried in the requirements is a list that reads less like education policy and more like a procurement questionnaire for anyone selling assessment software.
The requirement that matters to assessment
The policies have to address academic integrity, and the reporting is specific about what that means in practice: plagiarism, when students are permitted to use AI, how teachers can verify student work, and when students must disclose AI use.
Read those as four questions somebody will eventually put to you in writing.
The first one breaks most setups immediately. "When are students permitted to use AI" is almost never a single answer for an institution. It varies by course, by assessment type, sometimes by individual question. If your platform expresses that as a paragraph in a handbook rather than a setting attached to the paper, you cannot demonstrate compliance and you cannot enforce it either.
The second is the one we have written about before and will keep writing about. The rules ask how a teacher verifies work. They do not ask for a detector score, and that is the right instinct, because a probability output is not evidence anybody can reason about. Drafting history, an oral follow-up, a supervised tier for high-stakes papers: these produce something a human can examine and a student can contest.
The third sounds trivial and is not. Disclosure only functions if there is somewhere to record it at the moment of submission, and if that record stays attached to the attempt when it goes into an appeal eighteen months later.
About whether this applies to you
Here we should be straight with you, because the coverage is not consistent. Headlines describe the rules as covering K-12 schools and colleges. Body copy in the same reporting describes the obligations landing on school districts and charter school boards. We could not open the state's own release to settle it, so we are not going to assert a scope we have not verified.
What is not in dispute: the rules were approved on 16 September, the deadline is 1 July 2027, and the academic integrity requirements above are part of them. If you run assessment in Florida, read the rule text rather than the coverage, including ours.
The scope question matters less than it appears, for two reasons. Dual enrolment already pushes school district policy into college courses, so the boundary is porous in practice. And more importantly, requirements like these do not stay in one state. They become the template the next procurement committee works from, wherever that committee sits.
What an assessment platform should be able to produce
If you strip the policy language away, all four questions reduce to one: can you show your working, months later, to somebody who is unhappy?
Concretely, for any given attempt on a smart online exam, that means being able to answer:
What was permitted on this paper. Not what the institution's policy said in general, but the rule that applied to this assessment, on this date, as the student saw it.
What the student declared. Captured at submission, not reconstructed afterwards from memory or email.
What evidence exists beyond a score. If AI assisted the marking, what the model produced, what a human changed, and who signed it off.
Who saw what, and when. The access trail, because an appeal frequently turns into an argument about process rather than about the answer itself.
None of that is exotic. It is ordinary record keeping applied to a process that has historically been casual about it, on the reasonable assumption that nobody would ever ask.
The part vendors should sit with
The broader theme across the reporting is a distrust of opacity. Rules about undisclosed monitoring and about grading students without anyone being able to see how, aimed at exactly the kind of system that produces a number and declines to explain it.
For anyone building AI into assessment, including us, that sets the bar clearly enough. A grading model that cannot show its reasoning, cannot be overridden on the record, and cannot be explained to a student is going to fail this kind of policy test regardless of how accurate it is. Accuracy was never the thing being asked about.
Frequently asked questions
Is this law now?
The rules were approved by the State Board of Education on 16 September 2026, with a compliance deadline of 1 July 2027. That is an adopted rule rather than a proposal, but check the rule text for exactly who it binds before acting on a summary.
Does it ban AI grading?
Nothing in the reporting suggests a ban. The direction of travel is against grading that nobody can inspect. Disclosed, reviewable, overridable AI assistance is a different proposition from an unexplained score.
We are not in Florida. Why should we care?
Because these requirements are the ones your own committee will copy. It is far cheaper to be able to answer the four questions now than to retrofit the records once somebody asks.
What if our policy already covers AI use?
Most do, at the level of a written statement. The question these rules raise is whether the statement is connected to anything the system enforces or records. Those are very different things.
How long should assessment records be kept?
Long enough to outlast your appeals window, which is usually longer than the default retention in the platform. Work backwards from the longest appeal route a student actually has.
Top comments (0)