[ GLOSSARY ]
School AI glossary
Plain definitions of the AI, privacy and learning-science terms that come up when a school approves AI for students, each with why it matters and a source.
[ GLOSSARY ]
How to use this glossary
Short, practical definitions of the AI, privacy and learning-science terms that come up when a school approves AI for students. Each entry says why the term matters for schools, with sources.
Learning science
Cognitive offloading
Cognitive offloading is handing mental work to a tool. Why it raises performance but lowers memory, and what that means for students using AI.
ReadFormative feedback
Formative feedback guides a student's next attempt. What the research says about feedback (d = 0.48) and about AI feedback compared with teachers'.
ReadHint-first tutoring
Hint-first tutoring gives hints before answers. Evidence shows it removed the harm of answer-giving AI but did not beat no AI on its own.
ReadHuman in the loop
Human in the loop means a person reviews or acts on AI output before it affects someone. Why the best school AI results keep a teacher or tutor in it.
ReadProductive failure
Productive failure means trying a problem before being taught. The evidence (g = 0.36, reversed for grades 2-5) and why it supports attempt-first AI help.
ReadRetrieval practice
Retrieval practice means learning by recalling from memory. The evidence (g = 0.50 across 222 studies) and why AI-assisted work is not retrieval.
ReadSpaced practice
Spaced practice spreads study over time instead of cramming. The evidence (g = 0.74 vs massed study) and how to build it into AI-era homework.
ReadWorked examples
A worked example is a fully solved problem students study before solving similar ones. The math evidence (g = 0.48) and how it applies to AI answers.
ReadNexus
AI concepts
AI detector
An AI detector estimates whether text was AI-written. Why detectors misjudge non-native English writers and shouldn't be the basis for discipline.
ReadContext window
A context window is how much text an AI model can consider at once. Why free tiers' smaller windows matter for long readings and assignments.
ReadFine-tuning
Fine-tuning trains a model further on a dataset. Why training on student data raises consent and retention questions, and what to ask vendors.
ReadFrontier model
Frontier models are the most capable AI models available. Why they err less, cost more, and matter less for learning than the rules around them.
ReadGuardrails
AI guardrails limit what an AI will do. Why many are defaults students can leave, how often rules leak, and what guardrails can and can't achieve.
ReadHallucination
An AI hallucination is a confident false statement, such as an invented quote or source. How often AI errs on schoolwork and how students should check.
ReadLarge language model
A large language model is the AI behind chatbots like ChatGPT, Claude and Gemini. What LLMs do well, where they fail, and why model quality varies.
ReadModel routing
Model routing sends each request to one of several AI models by rule, such as cost, task or the user's age. How it shapes cost, quality and provider terms.
ReadPrompt caching
Prompt caching reuses the start of a repeated prompt at a discount. Why it cut total AI cost 27-40% in our model, not 90%, and what breaks it.
ReadPrompt injection
Prompt injection is text written to make an AI ignore its rules. How students get around chatbot limits, and why school rules must live outside the chat.
ReadReasoning model
A reasoning model thinks in steps before answering. Why that helps on hard problems, why it raises cost, and what teachers see in a chat review.
ReadRetrieval-augmented generation
Retrieval-augmented generation lets an AI answer from approved documents, often with citations. What it fixes, what it doesn't, and what to check.
ReadSystem prompt
A system prompt sets an AI's role and rules before a student types. Why teacher instructions are a kind of system prompt, and why rules leak.
ReadToken
A token is the unit AI models read and write, a word or part of a word. How tokens drive AI prices and what that means for a school budget.
ReadPolicy and practice
Acceptable use policy
An acceptable use policy sets the rules for using technology at school. What an AI policy adds, and why per-assignment rules matter.
ReadAI disclosure
AI disclosure means telling users they're talking to an AI, and students saying how they used AI. What laws and provider rules require.
ReadAI literacy
AI literacy is knowing how AI works, where it fails and how to judge its output. What the OECD-EC framework asks and Maryland's 2027 deadline.
ReadApproved app list
An approved app list names the tools a school has vetted for use. How common vetting is, how many tools districts use, and what to check for AI.
ReadCompanion chatbot
A companion chatbot offers ongoing personal conversation. Which state laws regulate them, and why school AI should be a study assistant instead.
ReadSafety escalation
A safety escalation route is how a concerning student message reaches a person. What a good route includes and what schools should decide in advance.
ReadShadow AI
Shadow AI is use of AI tools an organization hasn't approved. How it happened at work first, and why school AI blocks push it onto personal phones.
ReadWeb filter
School web filters sort AI sites into categories admins block or allow. How filters label AI, what CIPA requires, and how to allow one approved AI.
ReadLaw and privacy
Age assurance
Age assurance means estimating or verifying a user's age. New California rules, the FTC's 2026 policy and why schools should take age from the roster.
ReadCIPA
CIPA ties E-rate discounts to internet filtering. What it requires, why it does not require blocking AI chatbots, and what's under review.
ReadCOPPA
COPPA protects children under 13 online. What the amended rule says about retention and AI training, and how school consent works.
ReadData minimization
Data minimization means collecting and keeping only what a purpose needs. What Maryland law asks and how it applies to student AI chats.
ReadData processing agreement
A data processing agreement sets how a vendor handles a school's personal data. What a strong DPA for an AI tool should include.
ReadData retention policy
A data retention policy sets how long data is kept and when it's deleted. What COPPA and the FTC require, and what to set for student AI chats.
ReadEducation record
Under FERPA, an education record is directly related to a student and kept by the school or its vendor. Why stored AI chats likely count.
ReadFERPA
FERPA is the federal law on student education records. How it applies when schools share records with AI vendors, and what parents can ask to see.
ReadGuardian consent
Some AI providers require a parent's or guardian's consent before minors use their models. What OpenAI requires and how COPPA consent works in schools.
ReadMandated reporting
Mandated reporters must report suspected abuse. Why a disclosure in an AI chat a teacher can read may trigger that duty, using Maryland as an example.
ReadNDPA
The NDPA is the Student Data Privacy Consortium's standard student data privacy contract. How vendors sign it and how other districts adopt it.
ReadPPRA
PPRA bars analyses of protected areas, such as mental health, without written parental consent. Why AI that reads chats for mood may count.
ReadSchool official exception
FERPA's school official exception lets schools share records with vendors without consent, under conditions. What those conditions mean for AI vendors.
ReadSOPIPA
SOPIPA is California's law limiting how ed-tech vendors use student data: no targeted ads, sale or non-school profiles. What it means for AI.
ReadSubprocessor
A subprocessor is a company a vendor uses to process customer data, such as hosting or AI providers. Why schools should see the list and get notice.
ReadZero data retention
Zero data retention means an AI provider doesn't store prompts after answering. Why it has exceptions, and how it differs from 'no data collection'.
ReadTechnology
Audit log
An audit log records who accessed what, and when. Why a log of every staff read of student AI chats matters for trust and accountability.
ReadLTI
LTI is the 1EdTech standard for connecting learning tools to an LMS. What it does, and how Nexus's Canvas connection differs from an LTI tool.
ReadOAuth
OAuth lets an app get limited access to another service without a password. How school AI connections to an LMS use it and what IT should check.
ReadRostering
Rostering syncs students, classes and enrollments from the student information system into apps. The OneRoster standard and what Nexus supports.
ReadSingle sign-on
Single sign-on lets people use one school account across apps. Common SSO methods for schools, and what Nexus supports today.
ReadEvidence
Early-warning system
Early-warning systems flag students at risk from data. What trials found, how Wisconsin's labels misfired, and the lesson for AI tools that flag students.
ReadEffect size
An effect size measures how big a difference is, in standard deviations. Benchmarks from tutoring, quizzing and AI trials to compare claims against.
ReadESSA evidence tiers
ESSA defines evidence-based as strong, moderate or promising evidence, or a research-based rationale. What each requires, and how to read vendor claims.
ReadKnowledge tracing
Knowledge tracing estimates what a student has mastered from their answers. Why chat-based tracing is weaker, and how to read mastery claims.
ReadRandomized controlled trial
A randomized controlled trial assigns groups by chance to isolate cause and effect. How RCTs differ from correlations and working papers in AI research.
Read[ FAQ ]
Questions and answers
Who is this glossary for?
School leaders, teachers, IT staff and parents who need plain definitions of AI, privacy and learning terms.
Are the legal definitions legal advice?
No. They summarize public rules with links to the text.
Can I suggest a term?
Yes. Email support@narb.cc.
[ AI FOR SCHOOLS ]
The AI your school can approve.