Grades 6-8 are the first band whose default has students driving a generative system themselves, and what the teacher fixes in advance decides what that first contact is. The recommended default for the band is supervised, bounded use. Four things a teacher sets up in advance hold it in place: explicit prompt cards, a teacher-assembled source packet, disclosure of what the system contributed, and an initial attempt made before the tool opens.

UNESCO’s age-13 guidance for independent conversations with general-purpose systems runs straight through this band. Middle school sits across that line, which is part of why the band’s use stays supervised.

Bounding what the system can do before students open it

A student who is still building knowledge of a topic cannot reliably judge the answer a chatbot hands them. Recognizing that a synthesis omitted the counterevidence, or invented a citation, takes the domain knowledge the student is using the system to acquire. Open generation during initial exposure sets up a circular problem the student cannot get out of alone. Bounding the sources is the way out: when every claim has to point at a document in the packet, checking becomes locating a passage, which a student can do before they know the topic well.

A source packet is a defined collection the teacher or librarian assembles. It can hold primary sources from the Library of Congress or a state archive, the course textbook and teacher-authored notes, a set of court opinions, or several competing historical interpretations. The system runs under instructions: use only those materials, name the source behind each claim, say when the evidence is insufficient, add no outside facts, and make the student locate the relevant passage. Source-bounded systems are not infallible. What they improve is traceability, and traceability is what makes verification possible.

The initial attempt fixes the timing, which is the part teachers most often leave loose. Before the tool opens, the student produces evidence of their own first thinking, and it does not have to be long — a five-minute handwritten claim, three evidence selections, a spoken explanation. That attempt activates what they already know, makes their thinking visible to you, gives the system something specific to respond to, and preserves a baseline you can assess against later.

The prompt card and the disclosure line finish the set. The card fixes the narrow job the system is allowed to do on this task; the disclosure line records what it contributed, so authorship is on the record.

A school that refuses the technology outright has not removed AI from a student’s life. It has left the unsupervised version at home as the only version they get, and supervised purposeful use is the counterweight to that. The test to apply to any classroom technology is whether it changes what the learning can be, and AI meets that test or goes, the same as everything else.

Deciding which tasks earn any AI at all is a separate discipline, and it runs assignment by assignment. That decision has its own two pages here: the Five Tests for any proposed classroom AI use and The Four Modes of AI Assignments.

Running AI on one screen per group

Group use may be preferable to one-device-per-student use in this band, because it makes deliberation visible and reduces isolated dependence. With a device each, a student reads a fluent summary of a Library of Congress document, decides it sounds right, and moves on, and nobody — you included — sees the moment of acceptance. Around one screen, whether the summary says what the document says has to be settled out loud, and you can hear the disagreement and correct it.

Isolated dependence is the slower cost. A student working alone builds a private habit of asking the system first, and nothing interrupts a habit nobody else can see. In a group, a peer watches the reach for the tool happen and says to check the document first.

Keep the individual attempt even when the screen is shared. Each student writes their own claim before the group starts, so the group has something of its own to test the system’s answer against, and you still have a baseline for each student.

What students can start learning about the systems

Students can begin to learn how generative systems predict language, why hallucinations occur, how training data and design choices affect who gets represented, and what happens to the information they type in. The facts underneath are plain enough to state to a twelve-year-old. Fluent language is not evidence of truth; the same prompt may produce different answers; providers add hidden instructions the user never sees; and a model may not know where a claim came from.

Verification is the habit that turns those facts into practice, and its core move — lateral reading — is leaving the answer to investigate the information around it. Social studies runs the four-move version; the version that works at this age keeps every source out loud, in the group, rather than silent on one screen.

Judgment starts here too, in two specific forms. Students learn to tell appropriate delegation from inappropriate delegation: which parts of a task a system may carry, and which parts are the assignment itself. They also learn to revise using feedback without handing over authorship of the work. Neither is finished in this band. System knowledge, practical use, verification, ethics and civics, and self-regulation are the five parts of a full AI-literacy curriculum, laid out in what AI literacy covers across K-12. Social studies carries this work most naturally, because teaching AI in social studies already means teaching source verification.

Younger and older bands are shaped differently. AI in elementary school stays teacher-facing and largely unplugged, AI in high school moves to autonomy graduated by task rather than blanket permission, and the AI in K-12 hub collects the rest of this work.

Teach these facts inside the assignment students are already doing. When a student traces a claim back to the archive document it came from, they are learning where claims come from and doing the history assignment at the same time.

Start with the packet and the attempt

Next time you plan a unit where students will meet a generative system, decide the prompt card, the packet, the disclosure line, and the attempt first. Which tool they open matters less than those four. If you want more of this work as we publish it, the Kindred K-12 newsletter is where it goes. Kindred K-12 is built for teachers: there are no student accounts, no personal information about students, and the AI never chats with a student.