A System Prompt for K-12

While this piece is critical of some aspects of The STUDENT FIRST Act, I want to be clear that I admire the insight and efficiency of the students participating. They had three days to create a document to rein in AI use in traditional K-12 that would be published internationally. I hope that readers will see that my critique focuses primarily on the larger systems that constrain such policy work. Read most sympathetically, this Act should be understood as a prompt to change educational systems writ large. And we should not only be listening; we should be taking concrete actions to make the change.

A few weeks ago, one hundred students came together at the Edward M. Kennedy Institute in Boston to draft a comprehensive policy document on artificial intelligence in K-12 education. Over three days of debate and revision, they produced “The Safe Technology Use, Data Privacy, and Ethical Norms for Teaching—Fostering Integrity and Responsible Systems in Teaching STUDENTS FIRST Act of 2026.” The first letters of the long title spell STUDENT FIRST, singular. The bill itself calls the acronym “STUDENT FIRST” in one place and “STUDENTS FIRST” in another. You have to wonder if AI played some role in the naming. The document itself contains no AI transparency statement, though the students wrote a disclosure requirement into both the student and teacher sections.

Section 2.6 requires that any student who uses AI “must cite how and where they used it or else it will be considered plagiarism.” Section 4.4 states that “just as students are expected to be transparent in their AI usage, teachers should be held to the same standard.” And yet, the document arrives without any indication of how it was drafted. But I’d argue that the absence matches the genre they are inhabiting. Federal legislation does not usually carry a colophon, and Senate bills do not disclose their tools.

The naming encapsulates the argument. TECH FIRST is the orientation the students are writing against. As I’ve been writing about since the spring, districts across the country have been pushing AI tools into classrooms with a speed that has outrun the capacity of educators, families, and students to make sense of what is happening. Parent groups are organizing for opt-outs and analog spaces. Teachers are being handed instructional materials generated by tools they did not choose. Students at the assembly who assembled in Boston operated inside these pressures. Read closely, the document is a set of rules written to counter defaults built into traditional K-12.This is the shape of a system prompt.

Anthropic’s constitutional AI approach uses a written document of principles to steer the behavior of a language model whose underlying capabilities are already fixed. The constitution changes what the model is permitted to output within a given interaction while leaving its architecture intact. The rules operate at a layer above the base system. When people call this a constitutional approach, they mean a text that constrains a system that is already set in its ways.

The STUDENTS FIRST Act is this kind of text. Every provision reallocates permissions inside the existing operating logic of school. Grading, seat time, standardized assessment, discipline pathways, and credentialing all remain in place. The Act adjusts what students, teachers, parents, and administrators are permitted to do with AI inside those defaults. Section 2.12 sounds radical on first read: “Verification of Mastery: Students must be prepared to demonstrate independent mastery of learning objectives through in-person discussions, handwritten assessments, or oral defenses if AI was utilized during the drafting phase.” But look again. The oral defense is a new instrument built onto the same old premise, that mastery must be verified against a graded artifact.

The students worked at the ceiling of what the exercise allowed. A three-day Senate simulation with parliamentary procedure and a mandate to produce federal-style legislation sets a specific ceiling. The students got a lot of analytical work done under it.

Consider the transparency provisions. The teacher section requires disclosure symmetrical with what the student section requires of students. Students are watching teachers cede knowledge and agency to AI tools in ways that erode instruction and connection. They are receiving AI slop as instructional material. They saw a specific failure mode in the system they inhabit and wrote a mitigation into the prompt. If you have spent any time in the adult AI-in-education conversation over the past two years, you know this observation has been relatively absent from the policy documents produced by adults.

The detection provisions are the same shape. Section 2.1 guarantees that students suspected of AI use receive “a thorough examination of their work by two school officials, as well as an AI checker software.” Students have been on the receiving end of false positives, and they designed procedural protection against them.

These are the parts of the document that could do the most work if districts engage with it.

The question is which parts of the prompt actually survive after runtime.

System prompts get outranked. Later instructions override them. Base model defaults override them. Tools with their own permissions override them. The Act will land in districts that have to reconcile it with state law, counsel review, curriculum vendor contracts, IEP procedures, teacher union agreements, and daily classroom improvisation. Which provisions will hold?

The detector and lockdown browser requirements in Section 4.2 will survive. Districts already own the surveillance infrastructure. Section 5.5’s mandated bi-annual AI literacy training for teachers will survive in some form because state departments of education can fund it and audit compliance if that remains a state imperative. The opt-out provisions in Section 5.11 will live or die based on parent activism, surviving in districts where it is organized and quietly disappearing where it is not.

Other provisions will not survive. Section 5.10 states that “administrators must consult with students, staff, and families before drafting AI rules,” and that they “must publish their guidelines at least two weeks prior to the beginning of the school year.” No district has a process for genuine consultation, and none has an incentive to build one. Most districts finalize policy in August. Section 3.9 gives parents the right to request explanation of AI classroom rules and requires the teacher to respond “within one week”; where teachers have time this will happen, and where they do not it will not. The student consultation provisions are the most ambitious parts of the document and the most likely to get overwritten.

You can predict this without grading the students. The survival of particular provisions depends on how districts operate after the policy documents are drafted.

It is worth noting at this point that Constitutional AI is a loop. Critiques of the model’s outputs feed back into training, the model updates, and the document becomes an ongoing revision of surface behaviors of the system in question.

But the STUDENTS FIRST Act does not offer a clear pathway for creating such a loop. Interested districts will read it, engage with selected parts, ignore others, and the students who wrote it will have no way to learn what districts actually did or to press for revisions. The Kennedy Institute convened a legislative simulation and then adjourned it. When the students went home, the document became an artifact to be reviewed by adults with their own objectives and constraints.

The students were assigned to write a system prompt and they did the work with great insight and efficiency. It remains to be seen if the adults who set the assignment will build the revision loop that would let the prompt keep working.

The students proved they can do the analytical work. Whether we will build the mechanism that sets the prompt in motion is the open question.

Mike Kentz’s AI EduPathways: Insights from one of our most insightful, creative, and eloquent AI educators in the business!!!

Terry Underwood’s Learning to Read, Reading to Learn: The most penetrating investigation of the intersections between compositional theory, literacy studies, and AI on the internet!!!

Alejandro Piad Morffis’s The Computerist Journal: Unmatched investigations into coding, machine learning, computational theory, and practical AI applications

Michael Woudenberg’s Polymathic Being: Polymathic wisdom brought to you every Sunday morning with your first cup of coffee

Rob Nelson’s AI Log: Incredibly deep and insightful essay about AI’s impact on higher ed, society, and culture.

Michael Spencer’s AI Supremacy: The most comprehensive and current analysis of AI news and trends, featuring numerous intriguing guest posts

Daniel Nest’s Why Try AI?: The most amazing updates on AI tools and techniques

Jason Gulya’s The AI Edventure: An important exploration of cutting-edge innovations in AI-responsive curriculum and pedagogy