UNDERSTAND THE TECHNOLOGY
How does AI
actually work?
Training shapes the model. Context shapes the current task. Generation produces a response that still needs evaluation.
Three ideas that make the behaviour easier to understand
A language model processes language as sequences of tokens and learns relationships from training examples. Those learned relationships can support explanation, comparison, planning, and code generation. Understanding this process helps faculty specify a task and judge its result without treating fluent wording as proof that a claim has been checked.
Same topic. More useful context.
A conceptual walkthrough with authored examples. It is not a live model or a display of measured probabilities.
Working context
“Create an activity about campus sustainability.”
An illustrative continuation
Discuss ways a campus could become more sustainable. Share your ideas with the class.
The task gives a topic, but leaves the learner level, expected evidence, and decision criteria open.
Tokenize
The input is represented as tokens: words, parts, and punctuation.
Use context
The model combines learned patterns with the current sequence.
Continue
A next token is selected; the sequence grows and the process repeats.
Check
Sources, tools, and human review help evaluate the resulting answer.
Training develops learned patterns
Tokens can be whole words, word parts, or punctuation; the exact split depends on the tokenizer. During training, numerical parameters are adjusted in response to examples. Further training can shape instruction following and responses people find useful. This helps explain why a model can draft something that resembles a rubric before it knows your course. It has learned patterns associated with that kind of writing. The generated rubric still needs the actual learning outcome and assessment requirements. A familiar format can conceal criteria that are irrelevant to what your students were asked to do.
Context supplies the current working material
Your prompt, examples, selected readings, and available conversation can influence the current response. Transformer attention helps combine information across tokens in that context. Supplying a reading usually changes the material available for this task without retraining the model. A product may also preserve files or project history outside the model and bring them into later work. Make important requirements explicit rather than assuming they will appear automatically. In the conceptual walkthrough, notice how adding an audience and outcome changes which continuation would be useful. The illustration does not show measured probabilities from a live model.
Generation and verification do different work
The model estimates possible next tokens, a continuation is selected, and the process repeats with the growing sequence. Selection does not always mean choosing the single most likely token. This mechanism can support useful reasoning, yet a convincing answer may contain unsupported claims. Retrieval is a separate operation that brings external material into context. Tools can also open files, calculate a mean, or test a page. Those actions provide additional observations, but we still need to examine whether the resulting conclusion follows. Ask where a claim came from and what check would establish its support.
PUT THE IDEA TO WORK