Two districts bought Artificial intelligence (AI) with almost nothing to go on, a state wrote rules, and a company wrote contracts. Underneath all of it sits one question nobody had answered: what does “working” actually mean here?
Dayton Public Schools paid the ed tech company SchoolAI $78,650 this year for a set of AI tools, up from $72,600 the year before. A district in Salamanca, New York bought a humanoid robot named Sally for nearly $58,000, then shelved it after community objection, before it ever reached a classroom.
Both decisions were made locally, by people trying to do right by their students, with very little evidence available to guide either one (NPR, 28 September). That isn’t a knock on the buyers. A Stanford review of more than 800 papers found only 20 producing strong causal evidence, and as of October 2025 not one of those looked at students using AI in US K-12 classrooms (Fesler et al., 2026). Robin Lake of the Centre on Reinventing Public Education gave the week its best line: “It feels very Wild West in terms of the randomness.”
The sentence I keep coming back to is quieter. Katelyn Schoenhofer, an AI specialist in Wichita, told NPR: “We wrestle with what does success really mean when it comes to AI use?”
That’s the question under all of this, and it isn’t only a district problem. If you’re configuring a tutor for a course next term, it’s your question too.
Key takeaways
- Districts are buying AI with very little evidence to go on. Dayton paid SchoolAI $78,650 this year, and Salamanca, New York shelved a robot that cost nearly $58,000.
- A union-backed company standard, Florida’s statewide rules and vendor contracts all cover data, consent and conduct. None of them defines what it would look like for a student to have learned something.
- An explicit finish line in an AI tutor’s setup came with a smaller shortfall in thinking level. The finish lines here are mine, untested, and labelled that way.
Three institutions answered it inside three weeks
A company answered it with contracts. Announced on 9 September and effective 1 November, Microsoft is making a set of protections, agreed with the American Federation of Teachers and the United Federation of Teachers, available to every school district in the United States. Ten of them, contractually enforceable, addable to existing agreements with no renegotiation. Student and educator data can’t be used to train models, apart from a narrow safety-and-security exception, and can’t be sold or used for advertising. Tracking, profiling and biometrics carry limits and review. AI companions are prohibited outright, and families get plain-language explanations of what a product does (Microsoft and AFT fact sheet, 2026). What makes this one worth reading is the enforcement. Plenty of AI commitments in education arrive as a blog post you can’t hold anybody to. This is a term a district can point at.
A state answered it with rules. On 16 September Florida’s Board of Education adopted statewide requirements for K-12, charters and its 28 public colleges. Parents opt their child in or out of interacting with an AI tool, schools have to offer a non-AI alternative, tools for pre-K through grade five face extra review, and tools built to meet a student’s social or emotional needs are barred outright. Districts have until 1 July 2027 to write local policy (WUSF, 17 September). The clause I’d watch is the non-AI alternative. It reads as administrative housekeeping, and in a room where half the parents opted in, it quietly decides whether a teacher can use the tool at all.
And a training company answered it by pointing the question at its own teachers. Multiverse, the UK firm, runs a system called Compass that transcribes online lessons and scores instructors against a behavioural rubric, flagging filler phrases, hesitation, self-correction and pauses, and producing a risk status for each one. Instructors told The Guardian they were losing sleep, that some had sought therapy, and that managers questioned them over suspected digressions that were really answers to a student’s question. The firm says only humans write performance reviews, and that the system points managers at where attention is needed (reported by The Guardian, 28 September, read here through syndication). That is reported journalism; I have not seen the underlying documents.
Three very different institutions, inside three weeks, all doing the same thing: deciding in advance what good use looks like and writing it down somewhere that binds them.
| Who | What it binds | How it is enforced | What it says about learning |
|---|---|---|---|
| Microsoft and the AFT (effective 1 November) | Data and training use, tracking, profiling and biometrics; AI companions prohibited | Ten contract terms a district can add to an existing agreement | Nothing |
| Florida Board of Education (adopted 16 September) | Parent opt-in and opt-out, a non-AI alternative, extra review for pre-K to grade five, tools aimed at social or emotional needs barred | Statewide requirements; districts write local policy by 1 July 2027 | Nothing |
| Multiverse (Compass) | How an instructor sounds on a recording: filler, hesitation, self-correction, pauses | A risk status per instructor, reviewed by managers | Nothing about what students learned |
Privacy, consent and conduct are easier to write into a contract than learning is.
The one nobody wrote down
Notice what’s missing from all three. Microsoft’s standard is about data and safety, Florida’s rules are about consent and what’s appropriate for a seven-year-old, and Compass is about how a teacher sounds on a recording.
None of them defines what it would look like for a student to have learned something.
Privacy, consent and conduct are real, and they’re far easier to write into a contract than learning is. Which leaves the one that matters most to anyone who teaches exactly where it has always been: with you, in the setup of whatever tool you hand a student. That holds just as firmly in a university course shell as in a Grade 7 classroom. Worth knowing where Florida’s rules stop, incidentally: they reach the 28 colleges of the Florida College System, and the state’s universities answer to a different board that was still drafting its own version.
The version you can write this week
There’s a study I keep coming back to, and this is the practical half of it. Twenty teachers wrote the setup prompts behind 1,479 student conversations across 39 classrooms, and researchers scored every conversation for depth of thinking against the depth the teacher had been aiming for. Most landed shallower. Where the setup carried an explicit finish line, a plain statement of what “done” looks like, the shortfall was 0.22 smaller on the Depth of Knowledge scale, which is the four-level scale a lot of curriculum work already uses. Finish lines turned up in 64.1 percent of the prompts, so most of these teachers were already writing one, which makes the gap between them and the rest more interesting (Liu et al., 2026, a preprint, and observational: nobody randomised anything).
The paper gives one example of the shape, paraphrased from a real teacher’s prompt to protect their privacy: stop after the student produces two evidence-backed discussion points. That’s a finish line somebody actually used, and it’s shorter than you’d expect.
It’s the difference between telling a contractor to tidy up and telling them the job’s done when the floor’s swept and the offcuts are gone. One of those gets you a swept floor.
So here’s my attempt at that, concretely. I’ll be straight about what these are: the three below are my own, and untested. Edit them to fit your subject. Each is written so the tool knows when to stop helping:
- For practice: “This exercise is done when the student can explain, in their own words, why the method works, not only that it gave the right answer. Until then, ask questions. Don’t supply the explanation.”
- For a first draft: “Done means the student has a thesis they can defend and three reasons they chose it. Do not write sentences for them. If they ask you to, say no and ask what they’re trying to argue.”
- For revision: “Done means the student has found two problems in their own work without being told where to look. Point at the paragraph, never at the error.”
Honestly, which words you choose matters less than having a finish line at all. Presence is what the study measured, not phrasing. Write yours, then watch what your students actually do with it.
What to watch for
What I’d take as a sign it’s working is small and slightly annoying. The tool starts refusing things, and a student complains that it won’t just tell them. In my own material, that complaint has been the sound of the design holding.
If you get the opposite, a tutor cheerfully finishing the work while technically satisfying your wording, the finish line was measuring the artifact, when it needed to measure the understanding. Rewrite it around what the student should be able to do unaided.
Next edition: what the evidence actually says about AI feedback on student writing, which is the most common classroom use and one of the least examined.
These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.
One question I’d like answered, since I’ve only tested this on my own material. If you’ve written a finish line into a tutor’s setup, what wording actually held up when a student pushed back on it? The failure mode I’d expect is a clever student talking the tool past its own instruction, and I don’t know how often that happens in practice.
How this was made. Full disclosure, and I mean it as a courtesy, not a confession. AI tools help me research, check my claims, tidy my grammar, and stay on top of far more reading than one person can. The thinking and the judgment are mine, I edit every word into my own voice by hand, and then I put my name on it. Telling you that is the whole point of keeping it human.
References
- Gaines, L. V. (2026, September 28). Schools are experimenting with AI with little evidence or policy to guide them. NPR. https://www.npr.org/2026/09/28/nx-s1-5759718/ai-schools-experiment-research (This report independently cites the Stanford SCALE review of 800+ papers that edition 8 covered.)
- Microsoft and the American Federation of Teachers. (2026). AI Safety & Privacy Standard for Schools [Fact sheet]. https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/AI-Safety-and-Privacy-Standard-Fact-Sheet.pdf
- WUSF. (2026, September 17). Florida approves statewide AI rules for K-12 schools and colleges. https://www.wusf.org/education/2026-09-17/florida-approves-statewide-ai-rules-for-k-12-schools-and-colleges
- The Guardian. (2026, September 28). Teachers at Euan Blair firm report ‘horrendous stress’ after AI used to rate their work. Originating report; read here through syndication at IBTimes UK, https://www.ibtimes.co.uk/multiverse-instructors-ai-monitoring-compass-1822478 (Journalism, not a document examined at source.)
- Fesler, L., Martinez Claeys, J. P., Agnew, C., & Loeb, S. (2026). The evidence base on AI in K-12: A 2026 review. Stanford SCALE Initiative. https://scale.stanford.edu/sites/default/files/The%20Evidence%20Base%20on%20AI%20in%20K-12%20Report.pdf (Covered in edition 8.)
- Liu, A., Sun, M., Esbenshade, L., Tian, V., Zhang, Z., & He, K. (2026, April 17). Teacher-authored prompts for configuring student-AI dialogue: K-12 classroom implementation [Preprint]. arXiv:2604.16738. https://arxiv.org/abs/2604.16738 (The finish-line finding is this edition’s starting point.)