Four Clauses an AI Detector RFP Needs Before a Bid Can Be Scored

A vendor-questions checklist produces answers a district can't enforce. An RFP produces terms. The conversion is mechanical: each question becomes a requirement with a named deliverable, a format and a consequence for a blank. Four families carry most of the weight: evidence for accuracy claims, data handling for student writing, the student's route to the report, and the thresholds the product arrives with. The table below turns six evaluation questions into clauses inside those four families.

The questions to ask an AI detection vendor belong to the evaluation stage. This is the document stage, where an answer has to land in a form a procurement officer can score and a general counsel can sign.

A bid that answers five of the six is not a partial win. The blank one is the term that gets argued about in March.

Require the set, not the number. A bid that quotes a false positive rate says how many documents produced it, who wrote them and how long they were, and it dates the measurement. Nature's 2026 review of detection tools found that only the accuracy rates firms announce from internal testing can be current, and that those are not externally verified ¹.

A rate with no set behind it is marketing copy.

The condition attached to a figure matters more than the figure. Turnitin puts its document-level false positive rate under 1%, and says in the same post that the number holds for documents its model reads as 20% or more AI writing ². That is a real measurement with a real scope. A district that quotes the first half without the second has put the wrong sentence in a board memo.

Whose writing sat in the set is the question that decides equity. Seven detectors tested on 91 TOEFL essays by non-native English writers wrongly flagged more than half of them, an average false positive rate of 61.3% ³. No headline accuracy figure exposes that. A clause requiring results broken out by writer population does.

The requirement is answerable, which is worth saying to a bidder who calls it unreasonable. Pangram Labs publishes a per-domain table in its Pangram 4 model card, with academic writing in English at a 0.019% false positive rate across 62,971 documents ⁴. Domain, rate, set size, in one row. A bid that can't produce a table of that shape hasn't measured itself in the way a district needs.

Require the date beside the figure as well. Models change. A model released in July is not the model the case study was written about.

A district already owes FERPA's terms on every paper it keeps: student papers are education records, directly related to a student and maintained by the district or a party acting for it ¹². Those obligations travel to a detector. FERPA treats an outside provider as a school official only where it performs a service the district would otherwise use employees for, stays under its direct control over those records, and is bound by FERPA's limits on passing the records on ⁵.

That test is contract language. The RFP writes it rather than assuming it.

The full argument sits in FERPA and student papers sent to a detector.

Department of Education guidance on online educational services adds the part that gets skipped. The provider has to meet the criteria the district's annual FERPA notice sets out for a school official with a legitimate educational interest ⁶. The guidance says a signed contract is the common way a district establishes the direct control the exception demands, and that a Terms of Service the district agreed to can carry the same provisions ⁶.

Three clauses then do the work:

• Retention. How long a submission is kept, in what form, and whether the clock starts at upload or at the close of an integrity case.
• Deletion. A route the district can trigger on request, with a completion deadline rather than a best effort, and confirmation in writing.
• Training. Whether student writing enters a training set at all, written as a prohibition rather than described as a current practice.

Where a vendor's service collects personal information from students under 13, COPPA can apply to the operator. The Federal Trade Commission places that compliance on the operator rather than the school, and says an operator should not write the opposite into its terms of service ⁷. An RFP can reject that shift by name. A bid that carries it inside a linked online agreement has answered the question without meaning to. The K-12 specifics are in COPPA and student writing.

Illinois already folds the federal test into state contract law: an operator agreement must state on its face that the operator acts as a FERPA school official with a legitimate educational interest ⁸. A district in another state can borrow the sentence.

A parent or eligible student, through the district, inside 45 days. FERPA gives a parent or eligible student the right to inspect and review education records. The school must answer within a reasonable period, and no more than 45 days after receiving the request ⁹. Where the detector report sits in a vendor system the district can't export from, the deadline still belongs to the district.

Department of Education guidance says it directly: where a provider holds education records, the school and district must be able to provide the requesting parent or eligible student with access to those records ⁶. The RFP turns that into a deliverable rather than an assurance: the full report, per document, exported in a readable format, on request, at no additional charge.

Three things belong in that export before a district can hand it to a family. The verdict or score, with the scale it sits on. Every passage the product marked, in place. And the model version and date that produced the reading, so the same document can be re-run in six months and the difference explained.

A screenshot of a dashboard is not the record a hearing needs. A number with no version attached is hard to defend at one.

Charging per export is the quiet failure here. A per-report fee turns a records request into a budget line, and a budget line is how a district ends up hesitating over something it's obliged to produce.

As a number. The bid names the default and names who owns the setting. A detector converts a continuous score into a verdict somewhere, and where that line sits decides how many papers a district argues about. Turnitin marked AI scores below 20% with an asterisk in May 2023 because the score is less reliable in that band, and raised its minimum document length from 150 words to 300 ¹⁰. Since July 2024 it attributes no score there ¹³.

Run the arithmetic on district volume before the award rather than after the first complaint. Vanderbilt ran the sum in public: 1% of the 75,000 papers it sent Turnitin in 2022, around 750 papers that could have been incorrectly labeled ¹¹. The sum assumes a rate measured on Turnitin's own set holds for Vanderbilt's papers. It also counts every submission rather than only the human-written ones a false positive can land on, so it is an upper bound on documents.

The rate reads as small. The count is 750 documents somebody has to read again.

The same arithmetic at a district's own numbers takes five minutes and belongs in the board packet.

An RFP can require a bidder to list every operating point the product supports and say which one arrives by default. One vendor's answer to that clause already sits on a public page: Human lists three points: accusation-safe flags a paper above 15% machine content, standard above 6% (the default), sensitive above 2%. Its false-flag figure, measured as of September 2026, is 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Both figures are Human's own in-house measurements on its own held-out set rather than an independent audit, which is the distinction the evidence clause exists to draw.

A threshold the district can't change is a policy the vendor wrote. A threshold any teacher can change per assignment is a policy nobody wrote. The RFP should say which of the two the product is. A setting with no named owner is the one that gets changed without anyone deciding to change it.

Require a pilot on papers the district already holds, de-identified, from its own students. Vendor benchmarks run on corpora the vendor picked. A district that scores several hundred of its own passing papers through each finalist learns the one number no vendor benchmark can supply, which is how many of its own students' submissions come back flagged, broken out by course, grade level and language program.

Two conditions keep it honest. The papers predate the product's public release wherever the archive allows, so nothing in the set was written against it. And the same set goes through every finalist at each product's default setting, with no bidder told which papers are in it.

Write the consequence in as well. A bidder that won't support a pilot on district data has told the evaluation team something about the product. So has one that supports a pilot only after a term sheet is signed.

The clause procurement forgets is the one about what happens after a flag. A detector contract silent on appeals hands the design of the appeals process to whoever happens to be in the room the first time a student disputes a verdict. The RFP can require the deliverables an appeal needs: the report, the model version, and an export the student can read without a license.

None of these clauses asks a bidder for a promise. Each asks for a document that either exists or doesn't, which is why the envelopes sort themselves.