A federal AI report told schools to keep humans in the loop

The classroom note

A district technology director weighing an adaptive maths tool in 2023 had a new federal document to read first. In May 2023 the US Department of Education's Office of Educational Technology published Artificial Intelligence and the Future of Teaching and Learning, drawn from listening sessions with teachers, families, researchers and policymakers. Its most quoted line is a boundary, not a feature list: the report says it does not intend to imply that AI can replace a teacher, a guardian or a school leader as custodian of a student's learning. That sentence reframes procurement from "what can this tool do" to "what must a person still decide."

What the evidence says

This is a policy report, not a study: no trial, no measured outcome, only recommendations from stakeholder input and cited literature. It names five — keep humans in the loop; choose models aligned to an explicit vision of learning; make AI inspectable, explainable and overridable; use assessment expertise against algorithmic bias; and target research at specific classroom contexts rather than generic claims. A companion page on the department's educational-technology site, retrieved as it stood in 2026, describes the report as guiding educators while "evaluating and limiting key risks" and places it beside later guides for developers and leaders. That second document adds scope, not new evidence: the report reads as the first in a growing series, not a one-off release.

The implementation question

The explainability recommendation has a concrete cost: a vendor must show a teacher why a tool recommended one resource over another, not merely that it did. The override recommendation names a second mechanism — teachers need protection against adverse consequences when they override an automated decision, a staffing and policy question, not a technical one. On bias, the report calls for "assessment expertise" against feedback-loop bias, which implies specialist review time few districts budget for. None of this is funded or required by the report; it is a checklist a buyer can put to a vendor before signing.

What holds and what fails

The report's front matter states that its contents "do not have the force and effect of law and are not meant to bind the public." That holds as a limit: no school, vendor or state must follow the five recommendations, and the document creates no enforcement path. It fails wherever a leader treats the recommendations as already satisfied by a vendor's marketing rather than a demonstrated answer. This is an editorial reading: the report works as a procurement question list, not as evidence a purchased tool already meets its own standard.

• Can the vendor show, for a specific student and recommendation, why the tool suggested that resource rather than another?
• Who is authorised to override an automated instructional decision, and what protects them if they do?
• What would demonstrate the tool works as well for the students least represented in its training data?

Later national documents in this batch turn parts of this same territory into binding rules or funding conditions; this one did not, and comparing the two is the more useful reading than judging either in isolation.