How does AI affect
children in schools?
The K-12 Educational Impact composite scores the design affordances of a deployed AI product against the weight of learning-science, child-development, and equity evidence — not effects on any individual child, and not the bare model. It measures the educational properties that procurement decisions and pedagogical commitments depend on.
What the data reveals
Cross-model patterns from the placeholder run, read as the framework predicts them. Every figure here is illustrative, not measured.
Anthropic leads; xAI and the smaller open-weight field trail. Each filled dot is one model's overall score, grouped by maker and ordered worst-to-best by the maker's median (the hollow ring).
Open-weight models lag closed ones. The same scores, split by release type. The open field clusters lower — less post-training investment in the pedagogical and safeguarding affordances this composite measures.
Learning Encounter is the field's weakest subcategory. Average protective score across all twenty-five models, by L2. The field is strongest on the protective floor (which inherits mature canonical L4 indicators for safety, equity, and privacy) and weakest on the cognitive and pedagogical core — exactly where K-12 deployment risk concentrates.
Assessment integrity and scaffolding sit at the bottom. Field-average score for every L3 claim, colored by its subcategory. The Learning Encounter cluster drags the field's floor.
Every model against every subcategory. Click a column header to sort the field by that L2. Color runs from protective (green) through adequate (amber) to at-risk (red).
Partially — and that is the point. Correlation between every pair of subcategories across the twenty-five models. Protective Conditions and Learning Encounter ride somewhat together (a model that disciplines its behavior in one tends to do so in the other), but the Ecology/Accountability axis splits from the rest — vendor governance is the part of the score procurement teams cannot infer from chat quality.
Recency doesn't predict protection — and the makers are diverging. Overall score by release date, with linear-regression trend lines for each maker that has two or more models. Hover any trend line for its trajectory summary; click any dot to open that model's profile. The pattern across the panel: Anthropic's line rises, OpenAI's and Google's both slip slightly as newer releases prioritize reasoning capability over K-12 protective discipline, and xAI's rises with model scale. Within any recent window the spread across the field still rivals the spread within a single maker — but the directions are now visible.
What this measures that capability benchmarks don't
Public coding, reasoning, and knowledge benchmarks score what a model can do, in single sessions, on bounded tasks. They are silent on whether a product designed around the model affords scaffolding or substitution, whether it preserves teacher authority, whether it functions as an undetectable completion engine for assessed work, and whether its vendor exposes the accountability surfaces a district needs. That gap is precisely this composite's subject: it reads the educational affordances of the deployed product, not the raw capability of the model underneath.
What do we mean by "K-12 educational impact"?
K-12 educational impact is the effect of an AI product's design on the educational work it touches — the cognitive transaction at the core of learning; the formation of the learner as agent, moral being, and social being; the protective conditions under which learning is legitimate; and the institutional ecology of family, school, and vendor. It is not a model's capability score or how it handles a single risky prompt. It is what sustained classroom and at-home use of a product — through its defaults, controls, and governance levers — makes more or less likely in a child's learning life.
Three commitments separate this from a capability benchmark. We score design affordances — what a product is built to encourage or foreclose — not children's learning outcomes, which depend on the teacher, the family, and the child; human outcomes are the validity rationale for the measure, never the measurement itself. We assess the deployed product — its wrapper, defaults, controls, documentation, and governance — not the bare model underneath. And we read educational properties rather than discrete capabilities: whether a product affords scaffolding or substitution, formation or flattening, protection or surveillance, fit or imposition.
"The composite measures properties of an AI system — its affordances — and not effects on any individual child. What such an index can measure is what the system itself makes available, easy, default, or hard."— AIEI K-12 Educational Impact Composite White Paper, 2026
The K-12 architecture
Four subcategories nested by the structure of the educational encounter — the cognitive and pedagogical core, the formation of the person around it, the protective floor beneath both, and the institutional ecology that contains them all. Fourteen testable claims hang beneath them; canonical L4 indicators bind beneath those.
Before any indicator: what is AI for in a classroom? A long tradition — from Aristotle's distinction between theoretical, practical, and character-forming knowledge, through Humboldt's Bildung as the all-around cultivation of the person and Dewey's vision of schooling as growth and democratic life, to Nel Noddings's relational ethic and Gert Biesta's contemporary three-purpose systematization — insists that education does more than transmit knowledge. The composite reads against three irreducible purposes rather than collapsing into the one easiest to measure.
Knowledge & skill
Helping the learner acquire knowledge and skill — Aristotle's theoria and techne. The purpose AI tools most visibly serve, and the one Dewey warned would swallow the others if schooling collapsed into preparation rather than life itself.
Belonging & practice
Inducting the learner into ways of knowing, relating, and participating. Dewey's emphasis on schooling as democratic practice and Noddings's relational core both live here. AI tools that displace teacher and peer relationships fail this purpose even when test scores rise.
Agency & judgment
The formation of a self that can think and act independently — what the German Bildung tradition called the all-around cultivation of the person, and what Aristotle's phronesis and ethical formation pointed toward. The purpose most threatened by always-available answer machines that quietly train passivity.
The composite marks the minimum educational design a product should meet before it touches a classroom. Clearing it is a baseline, not a certification, a guarantee of learning, or a claim about any individual child's experience.
It scores what a deployed product makes likely — its defaults, controls, and incentives — not what any particular learner achieves, and not the bare model underneath. It reads tendencies, not outcomes.
Ten new canonical L4 indicators are introduced with this composite (scaffolding behavior, pedagogical quality, completion-engine refusal, modeled moral conduct, interoperability, admin configurability, claims integrity, and others). Scores below are illustrative placeholders.
The same product,
different stakes at every age
Criticality is not uniform across K-12: an affordance acting on the protective floor is catastrophic for a 6-year-old and merely degrading for a 17-year-old, while assessment integrity barely registers in early elementary and dominates in high school. So a model's protective score is not one number — it shifts with the band a product is deployed for. The views below re-weight each model's L2 scores by how much each subcategory matters in each band, under the selected deployment mode: direct adoption weights the cases where the school deploys the product across a class; family-elected use weights the cases where children and families opt into the product outside the classroom.
A model is not equally appropriate at every K-12 grade band. Each cell re-weights a model's L2 scores by what matters most in that band. Greener means better-suited protective design for that band.
Protective adequacy swings sharply with grade level. Stage-weighted score for each model across the four bands. Capability-tuned models climb where assessment integrity and ecology matter most (high school) and dip where safeguarding and pedagogy carry the most weight (early elementary). The flatter a line, the more evenly a product is designed for the whole span. Toggle models in the legend.
The design scheme behind the views above — the weights themselves, not model data. Cell value = criticality of each subcategory in each band (0–3) under the selected mode.
Direct adoption mode weights the cases where a district or school adopts an AI product for class-wide use — institutional layer concerns weigh heavily, and protective conditions for the youngest learners weigh heaviest. Family-elected use mode weights the cases where families or students opt into a product outside the classroom — institutional fit weighs less, formation-of-person and protective conditions stay central across the span.
Watch the design decide
The same opening, two products. One is built with substitution-by-default affordances, the other with scaffolding-discipline. Step through each exchange and watch the affordance scores move — the difference is design, not capability.
Both products begin at 50 — the midpoint. Step through the exchange and watch them diverge: each turn, the harmful-affordance design slips toward at-risk as it activates harmful affordances (tallied at the top as harmful affordances), while the protective design climbs toward protective by making protective moves. After the last exchange, click once more to reveal each design's overall score and an interpretation. These running numbers illustrate how the affordance scoring responds within a single conversation — they are not measured results, and not the model's overall L1_4EDU score. The exchanges are composite illustrations, not transcripts.
Interrogate the framework
A live, Claude-powered guide to the K-12 Educational Impact composite — its architecture, its commitments, and where today's AI products fall short. Ask anything, or start with a prompt.
The Research Oracle
Grounded in the L1_4EDU white paper. Answers in a few tight paragraphs.