Glossary

AI did not invent the academic integrity problem, but it has reshaped it. The useful question is no longer 'did they use AI' but 'does this submission honestly represent the learner's understanding.' That requires policy that names allowed and disallowed uses, assessment design that makes process visible, and conversations with students about why integrity matters beyond the grade.

Adaptive systems range from simple branching exercises to genuinely AI-driven engines that re-sequence content in real time. They work best for well-defined skill domains — early literacy, math fluency, language grammar — and weaker where learning is open-ended. The instructional model behind the system matters more than the underlying AI.

AI detectors are unreliable. False positives — flagging genuine student work as AI — are common, especially for non-native English speakers and students with formal writing styles. False negatives are trivial to produce with light editing. Major vendors have walked back accuracy claims. Most thoughtful educators treat detector output as a conversation starter, never as proof.

Watch out — Never base an academic integrity decision on a detector score alone. The legal and ethical exposure is real.

AI disclosure is the new citation convention. Norms are still settling, but the direction is clear: name the tool, describe the use, take responsibility for the final output. Good disclosure practice is teachable starting in middle school. It also flips the conversation from 'caught using AI' to 'used AI well' — a much healthier frame.

Example — A footnote: 'I used Claude to brainstorm essay structure and to check grammar. All sources and arguments are my own.'

Useful AI policy is specific, time-bounded, and revisable. It names tools where it can, names behaviors where it must, distinguishes by course or task type, and includes both students and staff. Bad policy is a single sentence banning 'AI' as if that term is well-defined. Plan to revise yours yearly — the ground is still moving.

AI proctoring is one of the most contested categories in edtech. It surfaces real privacy, equity, and accuracy concerns: students with disabilities, students in shared housing, and students of color have all been documented as more likely to be flagged. Many institutions are moving away from proctoring toward assessment redesign — see authentic assessment and performance task.

Watch out — A flag is not evidence. Treat proctoring output as a prompt for human review, never as a verdict.

AI tutors range from narrow drill-and-feedback bots to richer Socratic systems that ask questions back. The best ones explicitly refuse to just give the answer. The worst ones are dressed-up homework solvers. Quality varies enormously by subject — math tutoring tends to be more reliable than essay tutoring, where the model can confidently mislead on style and substance.

Watch out — If the tutor never says 'I don't know' or 'try again,' it is probably a homework machine in disguise.

Algorithmic bias shows up in education as essay scorers that down-rate non-standard dialects, proctoring tools that flag darker-skinned faces more often, and recommendation systems that route some groups toward less ambitious content. NIST and other standards bodies have published frameworks for thinking about and mitigating it, but no audit eliminates bias entirely. The institutional answer is human review at consequential decisions.

Watch out — 'The algorithm is neutral' is almost never true. Neutrality is a design outcome, not a default.

True anonymization is harder than it looks. Removing names and IDs is rarely enough — combinations of less obvious fields (school, grade, demographic descriptors, timestamps) can re-identify individuals, especially in small populations. The stronger practice is data minimization plus careful aggregation, with anonymization as a backstop rather than the main line of defense.

Watch out — 'Anonymized' often means 'pseudonymized.' Ask which one the vendor actually does.

Authentic assessment asks students to do something that someone in the world actually does — design, decide, defend, build, present. It is harder for AI to short-circuit because the task includes situated judgment and process visible to the teacher. Authentic assessment is the most durable answer to the take-home essay problem.

Example — Instead of a five-paragraph essay on local water quality, students sample the local creek, analyze data, and present findings to the city council.

Backwards design — popularized by Wiggins and McTighe — has three steps: identify outcomes, decide what evidence shows mastery, then plan instruction. In an AI era this discipline matters more, not less: if you cannot say what genuine learning evidence looks like, you cannot tell whether AI helped or short-circuited the learning.

Most large modern AI models are black boxes in practice. You can see what went in and what came out, but not why. For high-stakes uses — grading, placement, intervention — this is a real problem. The mitigation is procedural rather than technical: human review at consequential decisions, audit logs, the right to challenge automated outputs.

Watch out — If a vendor cannot explain why their model made a decision, you should not let that decision stand alone.

Bloom's is the most widely shared vocabulary educators have for cognitive demand. In AI work it is especially useful as a prompt scaffold: asking the model for questions at 'application' versus 'analysis' versus 'evaluation' levels produces meaningfully different output. The taxonomy is not gospel, but it is a common language worth using.

ChatGPT is the most widely recognized AI assistant in education, partly because of timing and partly because of the free tier. It offers a free model and paid tiers with newer models, image generation, file uploads, and custom GPTs. In many districts, 'ChatGPT' is shorthand for 'any chatbot' — which causes confusion in policy conversations.

Watch out — 'No ChatGPT' policies often unintentionally permit Claude, Gemini, Copilot, and local models. Name the behavior, not the brand.

Claude is widely used by educators for long-document work — reading full papers, comparing multiple student drafts, drafting curriculum across multiple sessions — because of its large context window. Its training emphasis on helpfulness paired with honesty makes it a reasonable choice for K-12 settings, though no LLM is appropriate without human review.

Working memory is finite. Cognitive load theory distinguishes intrinsic load (the task itself), extraneous load (poor design that wastes attention), and germane load (effort that builds schemas). AI can reduce extraneous load — formatting, summarizing, simplifying — but should not absorb the germane load. That is the learning.

Watch out — Removing all difficulty is not a kindness. Productive struggle is where schemas form.

Constructivism underwrites a lot of modern pedagogy: project-based learning, inquiry, productive struggle. The AI tension is real — if the model hands a learner a finished construction, no construction happened. The implication for design is that the learner's cognitive work has to remain visible and assessable, even when AI is in the loop.

Anything outside the context window is invisible to the model: it cannot remember earlier conversation turns, cannot read a document chunk that did not fit, cannot recall last week's session unless that history is re-supplied. Newer models have very large windows, but bigger is not always better — long contexts can still bury the instruction that matters most.

Watch out — A long chat does not give the AI 'memory.' Once you exceed the window, earlier content silently drops out.

Copilot in education usually means Microsoft 365 Copilot — AI features inside Word, Excel, PowerPoint, Outlook, and Teams. For institutions standardized on Microsoft, it shortcuts adoption. The name is also reused by GitHub Copilot, a coding assistant, which is a different product with different licensing — worth disambiguating in policy documents.

AI adoption in education is, underneath, a data governance project. Which student data leaves the building, in what form, to which vendor, under what contract, for what purpose, retained for how long. If those questions do not have crisp answers, AI procurement should pause until they do. This is also where most institutional risk actually lives.

Data minimization is the single most useful privacy principle in AI work. The data you never collected cannot be breached, sold, subpoenaed, or repurposed. Before any AI adoption, ask: what is the smallest dataset that lets this work, and can we strip identifiers further. The default of most vendors is to collect more — your job is to negotiate down.

Diagnostic pretests are underused. They tell you who needs the next lesson, who needs the lesson before that, and who is ready to leap. AI can generate diagnostic items quickly and grade them. The actionable part is what you change about tomorrow's lesson when the data comes in.

Differentiation is one of the most time-expensive teaching practices, which is exactly why AI is so compelling here: leveled readings, alternative explanations, multiple modalities, scaffolded prompts. The risk is silently lowering expectations for some students. AI should give more learners a path to the real outcome, not give some learners a simpler outcome.

Watch out — Differentiated input is good. Differentiated standards, hidden under personalization, is tracking.

An embedding turns a sentence, paragraph, or document into a long list of numbers — a coordinate in 'meaning space.' Two pieces of text with similar meaning end up close together, even if they share no words. Embeddings are how search inside RAG systems, plagiarism-adjacent similarity checks, and many recommendation features actually work under the hood.

FERPA governs who can access student records and under what conditions. For AI tools, the practical questions are: does student data leave the institutional perimeter, is the vendor a 'school official' under FERPA, what consent applies, and what happens to the data after the contract ends. Free consumer tools rarely satisfy FERPA for identifiable student work.

Watch out — A free tool with no data agreement is not FERPA-compliant just because you trust the brand.

Fine-tuning takes a general-purpose model and nudges it toward a specific tone, format, or domain by feeding it your curated examples. For most educators, fine-tuning is overkill — prompt engineering and retrieval (RAG) cover 95% of needs at a fraction of the cost and risk. Fine-tuning makes sense when you have hundreds of high-quality examples and a stable, repeated task.

Watch out — Fine-tuning does not 'teach' the model new facts reliably. For facts, use RAG.

Formative assessment is the daily diet of teaching: exit tickets, quick checks, peer discussion, a misconception probe. AI shines here — generating quick checks aligned to today's objective, suggesting follow-up questions, summarizing class responses. Stakes are low, the teacher closes the loop, and errors get caught fast.

GDPR applies wherever the data subject is in the EU, regardless of where the institution sits. It establishes lawful bases for processing, data subject rights (access, deletion, portability), and high penalties for violation. For AI, the awkward questions are about training data, profiling, and automated decision-making — areas the regulation explicitly addresses.

Gemini handles text, images, audio, and video natively. For schools already on Google Workspace for Education, the integration into Docs, Slides, Gmail, and Classroom is a practical advantage. Data handling for education tenants differs from the consumer product — read the admin documentation before assuming student data is safe.

Hallucinations are a structural feature of how LLMs work, not a bug to be fully eliminated. The model is optimizing for plausible-sounding language, not truth. Fabricated citations, invented historical events, made-up quotes from real people, and confidently wrong math are all common. The educator's job is to verify anything the model claims as fact before using it with students.

Example — An AI invents a peer-reviewed study with a real-sounding journal name, real-sounding authors, and a DOI that resolves to nothing.

Watch out — The more confident the tone, the more carefully you should check.

Inquiry-based learning puts the question — not the answer — at the center. AI tools fit awkwardly here unless used carefully: an LLM that hands over polished answers shuts inquiry down. Used well, AI can be a research partner that helps students refine their questions, find sources, and notice contradictions worth investigating.

LLMs power tools like ChatGPT, Claude, and Gemini. They do not understand meaning the way humans do — they predict likely next tokens given the conversation so far. That prediction is good enough to draft lesson plans, summarize papers, and explain concepts, but it also means the model has no intrinsic notion of truth. Treat its output as a confident first draft, never as a verified source.

Example — Asking an LLM to draft a unit plan on photosynthesis for grade 7 in under five seconds — fluent, structurally reasonable, and still in need of teacher review.

Watch out — Fluency is not accuracy. A grammatically perfect answer can still be wrong.

Learning analytics promised early identification of struggling students. The reality is mixed: signals are noisy, models are often built on assumptions that disadvantage some groups, and 'early warning' can become a self-fulfilling prophecy. Use these platforms to start conversations with humans, not to assign labels.

Watch out — A 'risk score' is a hypothesis, not a diagnosis.

The LMS is where AI features land first in most institutions: AI-assisted grading hints, draft feedback, auto-generated rubrics, integrated chat. Whatever you adopt has to play well with the LMS, or it stays on the margins. The LMS is also where most student data lives, so AI integrations into it are where the privacy questions get loudest.

Maturity models give leadership a shared language for honest self-assessment. Typical levels run from ad-hoc experimentation, through coordinated pilots, defined practice, measured improvement, and finally embedded culture. The point is not to rush to the top — it is to know where you actually are and to invest in the next step honestly.

Misconception probes are diagnostic gold. Built on research into how students actually think wrongly about specific concepts — Newton's laws, evolution, fractions — they show whether instruction shifted the conceptual model or only the vocabulary. AI can generate decent first drafts of probes, but the strongest ones come from teachers who know their students' typical errors.

Example — 'A heavy ball and a light ball are dropped from the same height. Which hits the ground first?' — answers reveal Aristotelian intuitions about gravity that no amount of reading the textbook erases.

Performance tasks integrate content with skill: write a paper and defend it orally, design an experiment and analyze the results, build a model and explain the engineering tradeoffs. They are harder to fake with AI because the assessment includes the doing, not just the artifact. They are also more time-expensive to grade — exactly where AI assistance, used well, helps teachers.

Good prompts specify role, audience, constraints, format, and examples. Prompt engineering is not magic incantations — it is clear technical writing applied to a peculiar reader. The same skills that make a good lesson plan brief make a good prompt: tight objectives, explicit success criteria, scaffolding, and an example of the desired output.

Example — Instead of 'Write a quiz on the French Revolution,' try: 'You are a grade 9 history teacher. Write five multiple-choice questions on the causes of the French Revolution at a Bloom comprehension level, with answer key and a one-sentence rationale for each correct answer.'

RAG connects an LLM to a searchable library — your curriculum, your policies, your student handbook — and feeds the most relevant passages into the model alongside the user's question. This dramatically reduces hallucinations on internal content and lets the model 'cite' real sources. Most enterprise AI in education today is RAG plus an LLM, not the LLM alone.

Example — An institutional chatbot that answers staff questions about leave policy by retrieving the actual HR handbook passages, then summarizing them with the LLM.

Watch out — RAG quality is bottlenecked by the quality and structure of your underlying documents. Garbage in, polished garbage out.

Rubrics make assessment legible to students before they submit, which improves both work and feedback. AI is good at drafting first-pass rubrics from learning objectives, and at applying a rubric consistently across many submissions. Teachers should still calibrate against real student samples — the model has no idea what your particular grade 8 class actually produces.

Scaffolding is meant to fade. The point is independent capability, not permanent assistance. AI can be scaffolding done right (a tutor that asks guiding questions and steps back) or scaffolding done wrong (a tool that just finishes the task). The pedagogical question is always: what is the learner doing that they could not do before, that they will be able to do alone next time?

Watch out — If the scaffold never comes down, it is not a scaffold. It is a crutch.

Summative assessment is where AI policy gets uncomfortable. Traditional take-home essays and untimed problem sets are trivial to complete with AI. The choices are: change the format (authentic assessment, performance task, oral defense), control the environment (in-class, supervised), or change what counts as evidence (portfolio, process documentation). Pretending the old format still works rarely ends well.

Surveillance pedagogy is what you get when the response to academic integrity worries is more monitoring rather than better assessment design. It tends to corrode trust, disproportionately stress already-marginalized students, and produce diminishing returns. The healthier move is usually to redesign the assessment, not surveil the assessment.

Low temperature (near 0) produces more deterministic, conservative answers — useful for grading rubrics, factual summaries, and structured output. Higher temperature produces more creative, varied responses — useful for brainstorming, generating discussion prompts, or drafting creative writing examples. Most consumer chat tools hide this control, but it shows up in API access and some education-focused platforms.

Watch out — Low temperature does not mean accurate. It only means consistent.

Models split text into tokens before processing. A common rule of thumb in English is that one token is roughly three-quarters of a word. Tokens matter because most AI tools price by token usage and limit how many tokens fit in a single conversation. Long PDFs, transcripts, and student portfolios can hit those limits quickly.

Example — A 1,000-word student essay is roughly 1,300 tokens — half a textbook chapter is more like 8,000.

Introduced in 2017, the transformer architecture lets a model weigh the relevance of every word to every other word in the input. That parallel attention is what made LLMs scalable. Educators do not need to implement transformers, but knowing the name helps when reading documentation, research, or vendor claims.

Watch out — 'Transformer' in AI has nothing to do with the toys or electrical components — it is purely a model architecture.

UDL is built on three principles: multiple means of engagement, representation, and action/expression. Instead of designing one lesson and adding accommodations, UDL designs the lesson with variability assumed. AI is a powerful UDL ally — text-to-speech, summarization, translation, alternative formats — but only if you start from outcomes, not from convenience.

Lock-in shows up as proprietary data formats, integrations that only work with the vendor's other products, training investments tied to a specific UI, and contracts that make data export painful. For AI tools, ask early: can we export student data and prompt history in open formats, and what happens to our content if the vendor is acquired or shuts down. These questions feel paranoid until they don't.