Library / Glossary / The craft glossary
AI in learning products glossary
44 terms from R16, AI in learning products — 19 defined in the guide itself and 25 more from the field around it. Every term the guide teaches links to the slide that teaches it.
R16 · How to improve
AI in learning products
Where does AI help learning, and how do we know?
Every term below is defined in the words of ai in learning products, guide R16 of craft guides for educational technologists, and opens the guide at the slide where it is taught. 25 of the 44 are the field’s vocabulary rather than the guide’s own: words a reader will meet around this subject, defined here because the guide assumes them. 1 term is also defined by another guide in the series; where the two differ, both wordings are given. The whole craft glossary holds all of them together.
| Term | Definition | Referred to in | Read further |
|---|---|---|---|
| A | |||
| Adaptive learning | Software that changes the content, difficulty or order of tasks for each student according to a running estimate of what that student knows. |
|
|
| AI agent | A system in which a model plans and carries out a multi-step task by calling tools such as search, code or other software, with limited human direction. |
|
|
| AI literacy | The knowledge and skills that let people understand what AI systems do, use them well and judge their output critically. |
|
|
| AI Risk Management Framework | The US National Institute of Standards and Technology's framework, which organises AI risk work into four functions: govern, map, measure and manage (National Institute of Standards and Technology, 2023). |
| |
| AI-writing detector | A tool that labels text as human- or AI-written; seven detectors wrongly flagged most essays by non-native English writers, so a flag is a reason for a conversation, never proof (Liang et al., 2023). |
| |
| Algorithmic bias | Systematic differences in how well an automated system works for different groups of people, such as a marking model that is less accurate for some students than others. |
|
|
| Automated essay scoring | Software that assigns a mark to a piece of extended writing, using a model trained on essays that people have already marked. |
|
|
| Automation bias | The tendency to accept an automated system's output without checking it, even when other evidence contradicts it. It weakens human review of AI decisions. |
|
|
| B | |||
| Benchmark | A public, standard test of a model's general ability. Useful for comparing models, weak evidence about a specific classroom task. |
| |
| C | |||
| Cognitive offloading | Using an action or tool to reduce the thinking a task demands (Risko & Gilbert, 2016). |
| |
| Confabulation | Confident but false content produced by a generative model, often called hallucination (National Institute of Standards and Technology, 2024). |
| |
| Context window | The most text, counted in tokens, that a model can take into account at once, including the prompt, any supplied documents and its own reply. |
| |
| D | |||
| Deep learning | Machine learning with neural networks of many layers, each layer building more abstract features from the one before. It underlies modern speech, image and language models. |
|
|
| E | |||
| Effect size | The size of a difference between groups, often in standard deviations. Comparable only when tests and comparison groups are similar. R18 The size of a difference or relationship, in raw units or standardised against the spread. R21 A difference between groups expressed in standard deviations, so results from different tests can be compared. |
| |
| Embedding | A list of numbers that represents a piece of text, an image or another item so that items with similar meaning sit close together. It is the basis of semantic search. |
|
|
| EU AI Act | The European Union's 2024 regulation laying down harmonised rules on artificial intelligence, which lists systems used to evaluate learning outcomes among high-risk uses (European Union, 2024). |
| |
| Evaluation set | A fixed collection of real tasks with agreed answers, used to test a feature each time the model or prompt changes. |
| |
| Explainability | How far a system can give reasons for an output that the people affected can understand and check. |
|
|
| F | |||
| Fine-tuning | Further training of an already-trained model on a smaller, specific dataset so that it performs better on a particular task or in a particular style. |
|
|
| Foundation model | A large model trained on broad data that can be adapted to many tasks, and on which many separate products are built. |
|
|
| G | |||
| Generative AI | AI that produces new text, images, audio or code in response to a prompt, by modelling patterns in the data it was trained on. |
|
|
| Grounding | Making a model answer from specific supplied material, so its answers can be traced and checked. |
| |
| Guardrail | A rule built into a product that limits what the model will do, such as refusing to give a full solution before an attempt. |
| |
| H | |||
| High-risk AI system | Under the EU AI Act, a system in a listed use, including evaluating learning outcomes, that carries extra obligations (European Union, 2024). |
| |
| Human in the loop | A design in which a person can see, question and override an AI system's output (U.S. Department of Education, 2023). |
| |
| I | |||
| Inference | Running a trained model to produce an output, as distinct from training it. Every response a student receives is one inference, with its own cost and delay. |
| |
| Intelligent tutoring system | Software that guides a learner step by step and adapts to their responses. It predates generative AI (VanLehn, 2011). |
| |
| K | |||
| Knowledge tracing | Estimating, from a student's sequence of right and wrong answers, the probability that they have mastered each skill. Adaptive systems use it to choose the next task. |
|
|
| L | |||
| Large language model (LLM) | A neural network trained on very large amounts of text to predict the next token, which lets it generate, summarise and transform language on request. |
|
|
| M | |||
| Machine learning | Building systems that learn patterns from examples instead of following rules written by hand. Most current AI, including generative AI, is machine learning. |
|
|
| Model card | A short document published with a model that states its intended uses, its limits and how it performed for different groups of people. |
|
|
| P | |||
| Productive failure | A design in which students attempt problems before instruction, and learn from the attempt (Kapur, 2008). | ||
| Productive struggle | Effort on the thinking a lesson exists to build, as distinct from effort wasted on confusion. |
| |
| Progressive disclosure | Giving help in increasing steps, with a full solution last and only after an attempt. |
| |
| Prompt | The instruction and material given to a generative model as input. The model's output depends on it, so wording, examples and supplied context all matter. |
|
|
| R | |||
| Red-teaming | Deliberately trying to make a system fail or behave unsafely, to find problems before users do. |
| |
| Reinforcement learning from human feedback (RLHF) | A training stage in which people rate or rank a model's outputs and the model is adjusted towards the responses they prefer. It is how chat assistants are made helpful and polite. |
|
|
| Retrieval | Finding passages in an approved library and passing them to a model with the question, a common way to ground answers. |
| |
| Retrieval-augmented generation (RAG) | A design in which the system first searches a chosen collection of documents and then passes the passages it finds to the model to answer from. |
|
|
| S | |||
| System prompt | Standing instructions set by a product's developer, usually hidden from the user, that fix the model's role, tone and limits in every conversation. |
| |
| T | |||
| Token | The unit of text a language model reads and writes, typically a word or a fragment of one. Context limits and usage prices are counted in tokens. |
|
|
| Training data | The examples a model learns from. Their coverage, quality and biases shape what the model can do and whom it serves badly. |
|
|
| Transformer | The neural network design behind current language models. It uses attention to weigh every part of the input against every other part when producing each output. |
|
|
| Two sigma problem | Bloom's (1984) challenge to find group teaching as effective as the two-standard-deviation gain he reported for tutoring. |
| |
No term matches. Try fewer letters.
Nearby