Two parts, with a break between

Part 1 gives us a common frame. Lunch is a proper break. Part 2 puts those ideas to work.

Our stance: critical judgement

• You are in the right room whether you use these tools daily or have never opened one.
• We are building judgement, not expertise in any one product.
• We will use the tools and interrogate their outputs, in research and teaching.
• Success is knowing what helped, what failed and where a person must remain responsible.

You bring the expertise. The tool is fast and can sound certain, but its reliability depends on the task.

A brief word with your neighbour

Take about a minute each, and there is no need to perform. 'I have not really used one' is a perfectly good answer.

Most of us already have a story, and it usually turns on a moment of friction.

The latest in a long line of tools

We have long extended ourselves with tools. Institutions that reward output keep making the next one attractive.

Its output can resemble considered judgement. That is precisely why it deserves more scrutiny than the tools before it.

Getting started, with no expertise needed

A chat assistant takes a request and writes back. Behind it sits a large language model.

Mixed groups are a strength. Newcomers ask the questions that matter, and regular users show what is possible. No one is behind.

Technical background: given the context, a model assigns probabilities to possible next tokens, then selects or samples one. A vast corpus gives breadth, and post-training can add granularity, though recipes vary and are rarely public. None of that makes an output verified, impartial or right for your context. Bender et al. (2021). Ouyang et al. (2022) describes one post-training approach.

Cognitive friction is a signal

When an output is glib, misses the point or smooths over an important distinction, that friction is information. It marks where your judgement matters.

• Survey (n = 319). Higher AI self-confidence was associated with less self-reported critical-thinking effort, while higher confidence in one's own expertise was associated with more (Lee et al., 2025). Correlation, self-report.
• Small experiment (n = 24). Brief prompts that challenge an output induced critical and metacognitive thinking in an AI-assisted shortlisting task (Drosos et al., 2025, preprint).

We treat friction as data to record, whether we are marking, writing or analysing.

In an AI-assisted decision-making study, deliberate friction, a prompt that makes you pause and decide, reduced over-reliance where explanations alone did not (Buçinca et al., 2021).

The ethical landscape

Different risks need different checks.

In teaching, add assessment integrity, disclosure to students and equitable access. Frameworks: Russell Group, European Commission/ERA Forum, UNESCO and ICO.

Stepping back, the bigger questions

Beyond any single task, three larger shifts are worth carrying into Part 2.

• Thinking and writing. We often think through writing. If a tool drafts, where does that thinking go, and what might be lost?
• Fairness. AI might narrow gaps linked to language, support or career stage, and it might widen them. For whom, and why?
• Disclosure. What use of AI is material to disclose? Would fuller disclosure expose the inequities the tool helped to ease?

There are no settled answers here. These are policy questions in the making. Notice them in your own task.

Design human oversight

Generated material can inform a decision. A named person or role remains accountable for it.

Before you start a task, decide the following in advance.

1. Where must someone verify a claim, a number, a citation or a source?
2. Who signs off, against which standard, with time, evidence and authority to pause, override or escalate?
3. What will never be delegated: the final call on correctness, ethics, authorship or impact?
4. How is oversight arranged? Interwoven means a check at every step. Staged means a check at a few points between phases.

Name the checkpoints before the system produces anything, not after a failure. This is human oversight in practice.

What Part 2 asks of you

Take one task, use a tool and examine the output. You have fifteen minutes and three things to do.

1. Use it. Keep a tangible result: a critique, a template, a chart or a translation.
2. Catch it out. Note one or two moments where the tool was fluent and wrong, and keep the output as it appeared.
3. Prepare one line for the room, which you give in up to 45 seconds at 13:33.

The spectrum guides the tool choice. The red lines govern what enters it. Caught errors reveal friction. A named checkpoint keeps oversight in place.

Your pack also holds a five-dimension rubric and a short reflection on the wider questions. Both support the debrief after the activity.

Form your groups

Move to the idea you would most like to try. Form a group of five, and the letter beside it is your track.

Note your group and track. You regroup after lunch without re-forming. A · Methodological Blind-Spot Detector, test a design. B · Accessible Executive-Function Layer, build a support layer. C · Rapid Prototyping, make an artefact. D · Public Engagement Translator, reach a public. Briefs are in your pack.

Open the workshop app

How Part 2 runs (30 minutes)

Only the fifteen minutes from 13:18 are yours. The rest belongs to the room, so protect that window.

Three jobs, shared in seconds

Fifteen minutes is a demanding window for this, so take prompt action, pun intended. Have something in the tool within two minutes, then improve it.

Five of you, one problem, one shared note. Divide the three jobs and start, combining them if your group is smaller.

1. One of you drives the tool.
2. One guards the red lines, using the check on your group sheet.
3. One keeps the note in the app and speaks for the group at 13:33.

Everyone else questions the output. Pass the keyboard around.

The five-dimension rubric in your pack supports reflection afterwards. A sharp caught error tells you more than an unexamined 5.

Work on one problem

Use the agreed problem: your seed, a member's non-confidential task or a ready example.

1. Use it and keep a tangible result, such as a critique, a template, a chart or a translation.
2. Catch it out. Write down the fluent-but-wrong moments as they happen. They show what needs checking.
3. Prepare one line for the room: the limitation you found, the checkpoint you would require before use or how your field over- or under-uses AI here.

If time runs short, write the line first. If time remains, map automated and person-led steps. Then decide whether oversight is interwoven (checks throughout) or staged (checks at set points), and reflect on how your field uses AI.

Those last three are optional extras.

Capturing your work in the app

There is no report to write. Use the app to capture the essentials.

1. The note-keeper opens genai-rt.web.app, names the group and enters the passcode. One device creates the group. Everyone else joins with its name and six-character code.
2. Record the problem, artefact, tool, caught errors and the line for the room. The app saves as you go.
3. Submit before the lightning round at 13:33 where possible. If you are still typing when the round begins, submit as soon as you have spoken.

If you have no device, or the app misbehaves, HackMD (hackmd.io) is the fallback for non-sensitive content only, with the same data boundaries and headings. A countdown on screen shows the time left, and the facilitator archives approved work afterwards.

The museum of caught errors

Keep a brief record of what went wrong and what it teaches you.

• A confident but fake citation · flattened nuance · a plausible but invalid method · a biased rewrite · a misgraded answer.
• For each, note what happened, how you caught it and what it signals about where humans must stay in charge.

These entries often provide the most useful discussion material.

Back at your desk

Choose one to try this week, in research or in teaching.

1. Write a two-line record for one AI-assisted task: what you used, what you checked and what you did not delegate.
2. Add one human checkpoint to a workflow you already use AI to assist.
3. Before pasting anything sensitive, work through the data check on your group sheet.
4. Keep your own museum of caught errors for a fortnight.

A short follow-up, with a responsible-use commitment, will reach you by email.

Thank you

Generated language can be fluent and persuasive. You decide what counts as sufficient evidence, what can be shared and where responsibility remains.