Using AI well means asking better questions

13 Oct 2026
Presenter
13 Oct 2026

The FHS EdTech Symposium brought together expert perspectives on AI in teaching, learning, assessment, and research. Rather than pointing to to one settled position, the talks prompted questions to help educators make more thoughtful, informed choices about where AI belongs and where we need to retain human judgment.

Delegating vs learning

Walji’s discussion of cognitive offloading prompted an important distinction: where is the line between convenience and developing expertise? People routinely ostensibly use AI tools to reduce effort and free attention for other work. In education, however, the step that is convenient to delegate may also be the step a learner needs to practise.

Slide titled "Friction, offloading, surrender" with three sections. The first section contains the text "Friction: effortful thinking (retrieval, error, struggle) makes learning durable". The second section states "Offloading: using tools to reduce cognitive demand, ideally freeing attention for what matters". The third section states "Surrender: handing over the very thinking the task was designed to develop"

That does not mean preserving every difficulty. Some friction supports learning; other friction comes from confusing instructions, inaccessible systems, or unnecessary repetition. More useful questions to ask are: "What capability is a task fulfilling that is intended to develop skills? Does AI assistance support or replace it?".

For educators, this suggests making the purpose of an activity explicit. Students should know what help is permitted, what they are expected to check, and what they should still be able to explain or do themselves. Importantly, they should be able to articulate why educators prohibit certain kinds of AI help.

When is a plausible answer still inadequate?

Ferlito’s presentation drew attention to the way AI can shape which explanations and perspectives become visible. An answer may contain accurate statements yet omit important context that matters to the decision. This makes checking more than simply scanning for factual accuracy. We can also ask the following questions: "What explanation does this response make seem obvious? Whose perspectives are missing? Could that missing information change the recommendation?"

Slide titled "Accuracy is not epistemic adequacy". There are two sections. The text on the left reads "Accuracy: are the statements true? Are the cited sources genuine? Do the findings support the claims?". The right section states "Epistemic adequacy: Are relevant perspectives included? Does the framing fit this content? Are uncertainties and omissions visible?"

These questions need evidence. Asking a system to “include different perspectives” may produce a convincing account without supplying the voices or sources needed to support it. When relevant evidence is unavailable, it is better to name the gap than to fill it with an invented perspective.

Slide titled "Critical engagements: three questions". The three questions are "1. Whose: which scholarships, languages, and experiences are represented?", "2. What is missing: which explanations, contexts, or methods are absent?", and "3. Who decides: who can inspect, contest, and change the system?"

Why the “honest machine” matters

Kefale’s framing encouraged a practical stance: AI is a machine that can be directed, questioned and checked. Its fluency should not be confused with authority. A useful prompt can improve the response, but it cannot guarantee that the response is correct or appropriate. 

Slide titled "One AI tutor, four objectives". The slide contains a table with four rows and two columns. The header rows read "The tutor is optimised for..." and "How it can climb, and which machine". Row 1 column 1 reads "Time on platform, sessions per week". Row 1 column 2 reads "Make the student predictable: capture". Row 2 column 1 reads "Student satisfaction ratings". Row 2 column 2 reads "Please rather than teach: a proxy". Row 3 column 1 reads "Novel, surprising student answers".

For a learner, the goal should extend beyond producing a polished answer. They should be able to explain the idea, examine the support for a claim, and apply their understanding to a new example. 

Large format slide reading "If you can climb the objective by making the person more predictable, or less grounded, it is the wrong machine.

That gives teaching a clearer way to distinguish assisted performance from developing capability.

Slide titled "Noise is more surprising than insight". A confused or destabilised learner produces unpredictable output in abundance. A machine that maximises measured surprise can harvest it: a second capture.

Implications for teaching and assessment

These ideas are informing a redesign of first-year AI literacy teaching.

Students will first interpret a shared source themselves, then use AI to draft or revise an explanation. They will check claims against the source, look for consequential gaps, explain their revisions and apply the same judgment to a new scenario. We will emphasise a process that students can describe: what did the tool contribute? What did humans change? What was hand-checked? What can we still not conclude?

A polished output alone cannot show whether a learner understands the material.

Assessment also involves practical trade-offs. Oral questioning can provide useful evidence, but it takes staff time and may be difficult to apply consistently at scale. Short written explanations, transfer tasks and targeted discussion may each be appropriate for different learning aims. The choice should begin with the capability we want to assess and the evidence that could reasonably support that claim.

A continuing conversation

The symposium has left me with a set of guiding questions rather than a universal checklist:

  • What is the human-centred objective of this task?
  • Which parts of the work can AI support?
  • What should the student be able to explain or decide?
  • What evidence would show that an AI-generated answer is well supported and contextually adequate?
  • How can we gather that evidence without creating an unsustainable burden for staff or students?

Those questions can help educators approach AI with both curiosity and care. Removing every difficulty or treating every AI use as the same is counter to our aim here; we must make visible the purpose of assistance and go forward designing teaching around the knowledge, judgment, and independence we want learners to develop.