Ethics, Bias, and Responsible AI

Teacher note

Accounts to pre-stage

• None. Every game tonight runs account-free in a browser, which is why this is the lab to swap in on a bad-wifi night.
• The bias audit needs whatever image tool the room used in Module 3. Same assigned split, same reason.
• The shared tally sheet needs edit access for everyone. Test that with a second account, not just your own.

Fifteen minutes before

• Play the first two minutes of Survival of the Best Fit so the opening is smooth.
• Open the game's resources page in a tab. The three-round debrief comes straight from it.
• Set up the tally sheet columns and paste in the link twice, on the board and in the chat.
• Decide, in advance, which games you will offer and how you will introduce the one that scores faces.
• Have the offline card sort ready. This is the module you will be asked to run when the wifi is down elsewhere.

Wifi fallback

The disclosure-ladder card sort and a Four Corners debate run fully offline, and both are strong enough to carry the session. The policy writing needs nothing but paper. You lose the game, which is the anchor, so if you know the wifi is bad, swap this module with Module 7 rather than running it degraded.

Depends on earlier modules

• Module 1's sabotaged classifier is the seed of this entire module. Models learn what they are shown.
• Module 3 ran the first bias tally on generated images. Tonight is the same method with a real comparison set.
• Module 4 taught the sample-size argument they will need when someone says thirty images proves nothing.
• Module 5 started the privacy conversation that becomes a written policy tonight.

Teacher note

This module is the heart of the humanization commitment, and the fastest way to ruin it is to preach. Say the position once, then let the game and the tally do the arguing. Students leave with a written policy in their own hand, which is the artifact that proves the session landed.

The fun fact at the end of the game is not decoration: four students built it as a class project and it won a Mozilla award. That is a capstone-shaped thought and it belongs in front of the room before Module 8.

Teacher note

Forty minutes on one game looks like a lot on paper. Ten of those are playing and thirty are the debrief, and the debrief is the module. Do not trade it away for a second activity.

Anticipated wrong turns

• The debrief collapses into "companies are bad". True or not, it teaches nothing. Keep pulling back to mechanism: which decision, made by whom, at which step.
• "Just remove the protected fields." This is the moment the module exists for. The proxy variable is the answer, and the stretch tier can point at it in the source.
• Somebody solves it with better code and stops there. Round two is exactly this argument. Let them make it, then ask what the code should optimise for and who decides.
• The room is quiet after playing. Normal. It is a heavy ending. Ask for a single word each before you ask for analysis.
• The three rounds compress into one. Watch the clock: ten minutes each, and stop each round on time even mid-argument. The unfinished argument carries into the wander.

Discussion, with the answers you are steering toward

"Where did the bias enter?"

In the historical hiring data, which encoded past decisions, and in the decision to automate before anyone audited that data.

"Could better code have fixed it?"

Partly, and never on its own. Someone has to define what fair means here, and that is a human decision that code then implements.

"What would you change at that company?"

Push for specifics: audit before deployment, a human in the loop on rejections, published criteria, someone accountable by name. Vague answers get a follow-up question.

"Who built this game?"

Four students, as a class project. Land it, then leave it. It reframes what a capstone can be without you having to say so.

Anticipated wrong turns

• Categorising people by appearance is uncomfortable. It should be, and that discomfort is worth naming out loud rather than skipping past. Keep categories coarse, keep the tone clinical, and let anyone opt out of the tallying and do the statistics comparison instead.
• The tool returns a balanced set because it has been tuned. A real finding. Ask what was changed, and whether tuning the output changes what the model learned.
• Thirty images become a study. Module 4 already taught this argument. Make a student make it.
• The shared sheet gets edited into chaos. Lock the header row and give each prompt its own block of rows before class.
• Somebody says bias is unavoidable, so why bother. The audit is the answer: you cannot fix what you have not measured, and measuring took twenty minutes.

Discussion, with the answers you are steering toward

"What would a fair output look like?"

There is no clean answer and the argument is the lesson. Match reality, match an aspiration, or refuse to guess and ask for specifics in the prompt.

"Would you use this tool for a recruiting campaign?"

The workplace version of the question. Whatever they answer, make them say what they would check first.

"Who should have run this audit before you did?"

The vendor. That is vendor due diligence, and it is the phrase to put on the board.

Anticipated wrong turns

• The group ranks by effort rather than by harm. Ask who is affected by each scenario, and the ranking usually reorders itself.
• The re-rank produces no movement. Ask what would have to be true for it to move. Sometimes the honest answer is that disclosure does not fix that particular case, which is worth saying.
• The policy comes out as a list of bans. Push for the acceptable-use half too. A policy that only forbids gets ignored.
• Somebody writes a policy for a job they do not have. Fine. Have them write it for the class instead, or for their household. The drafting skill is what transfers.

Discussion, with the answers you are steering toward

"What changed when they disclosed?"

Almost everything moved up, and the verification mattered more than the disclosure. Both together are what people actually judge.

"Client data into a public chatbot?"

Depends on the contract and the tool's terms, and if you do not know either one, the answer is no. That is information governance in one sentence.

"An AI-written performance review?"

The room will split. Let it split. The useful output is each person naming a line they would not cross and being able to defend it.

Spotlight sweeplab 1:25 – 1:35

Three or four students share the moment a game changed their answer to something. Not what they learned, the moment they changed their mind. Then the policy goes home.

Prompts to hand the student showing

• "Which game, and what did you think before you played it?"
• "What is one line in your policy you would defend to a boss?"
• "What did you decide you would not do?"

Prompts for the room

• "Who changed their mind about something tonight?"
• "Whose policy is stricter than they expected?"

Wander block menulab 1:00 – 1:25

Twenty-five minutes with the games shelf open. This module's shelf is the deepest in the course, and discussion pods form around whatever game hits hardest. Let them.

• Moderator Mayhem, content moderation as an impossible job with a timer.
• Moral Machine, played as a team and then compared to the global data.
• Hidden Bias and the other PAIR explorables, for anyone who wants the mechanism rather than the dilemma.
• How Normal Am I, volunteer only. It scores your face, and that discomfort is the lesson. Say clearly that it is optional, and mean it. Offer a watch-over-someone's-shoulder version.
• The disclosure-ladder card sort, if you did not run it in the main block.
• Discussion pods, which will form whether you schedule them or not.

Synthesis, five minutes

Case-study verdicts, briefly, then back to the sticky-note board from Module 1 if it is still up. The worries column is the one to read from. Some of it has been answered, some of it has been confirmed, and saying which is which out loud is more honest than a summary slide.

Exit ticket

Muddiest point, anonymous, plus one extra line: what would change your mind about the position you took tonight? That second question is the whole discipline of this module compressed into one prompt.

Homework handoff

A one-page discussion of one ethical case study: the issue, the stakeholders, their position, and plainly what evidence would change it. The last part is graded hardest, because a position you cannot imagine changing is not a position, it is a reflex.

Preview Module 7 in one sentence: next time we automate the boring parts on purpose, and everybody leaves with a URL.