EDPU5003 Chap.4 Generative AI and Large Language Models
Generative AI and Large Language Models
Generative AI creates new content from learned patterns, while a large language model generates sequences of language by estimating plausible continuations. Educational analysis must identify the particular capability, the learner action it changes and the outcome evidence.
Generation, support and analysis can expand practice and access, but the same capabilities can introduce confident error, dependency, bias, privacy risks and false confidence. A useful evaluation begins by specifying the counterfactual: compared with what existing practice, for which learners and under which conditions? A rapid explanation may outperform no support while underperforming a skilled dialogue.
Average gains may conceal that novices accept errors while experienced learners benefit from variation. Access also includes language, disability, cost and reliable connectivity, so the same system can widen one route while closing another. Task design determines whether assistance becomes a scaffold or a substitute. A scaffold is faded, prompts active reconstruction and makes the learner's decisions visible.
A substitute performs the target act while leaving little opportunity to build it. This difference cannot be inferred from the tool name; it must be analysed at the level of learner activity. Evaluation should include quality, transfer, time, experience and distribution of effects. Privacy, intellectual property and the consequences of confident error belong in the main judgement, not an appendix.
When writing about opportunity and peril, pair them around the same capability. Generation can increase variety and also normalise shallow patterns. Personalisation can make support timely and also intensify profiling. Automation can reduce routine workload and also weaken professional oversight.
Finish by assigning roles: what the system may propose, what the learner must do, what the educator verifies and what evidence controls continuation.
What this chapter covers
- 01
Artificial intelligence and nested categories
- 02
Generative models and LLMs
- 03
Generation, support and analysis
- 04
Educational opportunities and perils
- 05
Learner agency and productive struggle
- 06
Accuracy, bias and privacy
- 07
Access versus understanding
- 08
Transfer as outcome evidence
Audit AI-generated feedback
- +1Define alignment among goals, activities and evidence as the learning purpose.
- +1Let the model propose questions rather than final judgements.
- +1Require learners to accept, reject or revise suggestions with reasons.
- +1Remove personal data and check confident inaccuracies and bias.
- +1Evaluate revision quality and delayed transfer.
Key terms
- Generative AI
- Systems that produce new content from patterns learned in training data.
- Large language model
- A model that generates language by estimating probable continuations from encoded patterns.
- Productive struggle
- Difficulty that contributes to constructing knowledge or judgement rather than merely blocking access.
- Automation bias
- A tendency to trust a system recommendation more than the available evidence warrants.
- Transfer task
- A task that requires a learner to apply an idea in a meaningfully different context.
Generative AI and Large Language Models FAQ
Why is the label AI too broad for analysis?
Different systems perform different functions and change learner activity in different ways. Name the capability, task, learner action and evidence instead of attributing a general effect to AI.
How can an opportunity also create a risk?
Immediate feedback can expand practice but may create dependency; rapid generation can broaden examples but also spread plausible error. Analyse both effects through the same capability and design choices.
What evidence goes beyond student satisfaction?
Use quality of explanation, revision decisions, delayed reconstruction and transfer to a new case. Satisfaction can inform experience, but it does not establish durable understanding.
When should friction be preserved?
Preserve difficulty when it is the act that develops the target capacity, such as forming a problem representation or judging evidence. Reduce barriers that do not serve the educational purpose.
Assessment move
Sort examples into generation, support and analysis, then state the learner verb and one risk for each. Practise turning a broad claim into a design audit: purpose, tool role, learner role, risk, evidence and boundary. Add a delayed transfer task to every evaluation plan. Check current task rules before using any generative system and protect personal or confidential information.
Choose one educational use and write a counterfactual, learner verb, target outcome, risk and transfer test. Repeat for a novice and an experienced learner to expose distributional effects. State which friction the tool should reduce and which must remain. Finish by assigning verification and accountability roles before you decide whether the design should continue.
Build a compact risk register for accuracy, dependency, privacy, bias and unequal access. For each risk, name one design control and one indicator that would show the control failed. Pair the register with a learning measure, because a safe tool that does not advance the educational purpose is still a poor design. Practise expressing the final decision as continue, adapt, restrict or stop.
Complete the Generative AI and Large Language Models review by teaching the central distinction to a peer, answering one challenge without notes and recording the exact evidence that would make you change your recommendation.
Working through Generative AI and Large Language Models in EDPU5003? Sia is AskSia’s AI Education tutor — ask any EDPU5003 Generative AI and Large Language Models question and get a clear, step-by-step explanation grounded in how EDPU5003 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
