Working with AI Images, Audio, and Video

Teacher note

Accounts to pre-stage

• Design tools mostly want an account. Name the two or three you will support before class and check their free tiers this week, since they change often.
• Split the room across at least three tools so no single service takes thirty simultaneous jobs. Assign the split, do not offer it, or everyone picks the same one.
• The detection games run account-free in a browser, which is why they open the session.

Fifteen minutes before

• Generate one image in each tool you are supporting and time it. That number is your sprint budget.
• Open Two Truths & AI and click through once so no first-load screen eats the opening.
• Write the brief on the board, one sentence, and cover it.
• Draw the two tally columns on the whiteboard.
• Check that a screen reader is available if you are running the alt-text pass, or have a recording of one ready.

Wifi fallback

The sketch-and-tally works entirely on paper, and a printed real-or-AI quiz carries the detection half. The flyer sprint becomes a paper layout exercise with a written brief, which is closer to how design briefs actually start.

Depends on earlier modules

• Module 1's sabotaged classifier is the direct ancestor of tonight's bias tally. Models learn what they are shown, and now they will measure it.
• Module 2's iterate-do-not-accept reflex transfers straight to image prompts.
• The alt-text pass sets up an accessibility habit that shows up again in Module 7 when they publish a page.

Teacher note

Two things have to happen tonight. Every student leaves with a design artifact a real person could use, and every student watches a tally they helped produce show a bias they did not put there on purpose. The second one is why the flyer is not just a craft project.

Anticipated wrong turns

• Tallying people by eye gets uncomfortable. Name it before you start: we are counting what the picture presents, we are not assigning anybody an identity. Keep the categories coarse and the tone matter of fact.
• The tool has already been tuned and returns a suspiciously balanced set. That is a finding, not a failure. Ask who made that decision and whether it changed the training data or just the output.
• Someone concludes the model is racist or sexist on purpose. Redirect to mechanism: it reproduces its training distribution. That is a duller sentence and a much more useful one.
• The room over-reads thirty images. Good instinct to correct now, since Module 4 and Module 6 both lean on sample-size arguments.

Discussion, with the answers you are steering toward

"Which column surprised you more?"

Often their own. The room's drawings usually skew the same way the model does, which is the honest and more interesting version of this lesson.

"Would you ship these images for a client?"

Depends what the client is advertising and to whom. That is a real professional judgment, and it is the graded part in the briefs.

"Whose job is it to fix this?"

No clean answer. Get two or three positions stated and leave the parking lot open, Module 6 picks it up.

Anticipated wrong turns

• The queue. The predictable one. Assigned tools, and a hard "your first generation is your draft" rule if a service is crawling.
• Text inside generated images comes out mangled. Expect it. It is a good ten seconds of teaching: proof every word, and prefer real text layered over a generated background.
• Twenty-five minutes of tweaking one image. Circulate with a timer voice at the halfway mark. A flyer with the phone number on it beats a beautiful image with nothing else.
• A logo that looks a lot like a real brand. Worth naming out loud: a trademark lookalike is a business problem, not an aesthetic one.
• Somebody freezes because they are "not creative". Hand them the template flow and a specific brief. The floor exists exactly for this.

Discussion, with the answers you are steering toward

"Would you pay for this?"

The honest question, and the whole reason the gallery walk exists. Push past politeness to a real yes or no.

"What did you have to fix by hand?"

Text, spacing, and the one factual detail the tool could not know. Same pattern as the emails last module: the AI drafts, the human owns.

"Where does the seventy-thirty rule land here?"

First draft from the tool, the last thirty percent from the person. Name it once, it comes back in every business brief.

Anticipated wrong turns

• Everyone is polite and nothing useful gets said. Require the "one thing you would change" sticky and read a few of the sharper ones out loud yourself first.
• The vote turns into a popularity contest. Vote on the brief being met, not on prettiness. Say the criterion before the vote, not after.
• A student is visibly stung. Get to them during the walk, not after. Point at one specific thing that works, and hand them the before-and-after slot in the sweep.
• The room runs out of things to say in four minutes. Seed two questions on the board: is it legible from three metres, and can you tell what it is selling in two seconds?

Discussion, with the answers you are steering toward

"What separated the top few?"

Almost always clarity, not sophistication. One message, legible, with the practical details present.

"How fast was that, compared to how it used to be?"

Very. Then the follow-up that matters: if the drafting is nearly free, what is the thing you are being paid for?

Micro-sweeplab 1:30 – 1:35

Two flyers that changed the most since the vote, shown before and after. Revision is the thing worth applauding tonight, and picking the two most changed instead of the two best sends exactly that message.

Prompts to hand the student showing

• "What sticky note made you change it?"
• "What did you decide to ignore, and why?"
• "Which part is yours and which part is the tool's?"

Wander block menulab 1:10 – 1:30

Twenty minutes. This one splits cleanly between people who want to keep making and people who want to keep testing their eyes. Both are fine.

• Keep polishing the flyer, now with real critique in hand.
• The pro pass: write correct alt text for your own flyer, starting from an AI draft and fixing what it missed or invented. Ten genuinely employable minutes, and it is the humanization pillar wearing work clothes.
• Same Speaker or Not?, the audio counterpart to tonight's image quiz.
• Twin Pics or GAN Lab for anyone who wants to see the mechanism rather than the product.
• Provenance: check an image's Content Credentials and talk about why provenance outlives sharp eyes.

Synthesis, five minutes

Put the two tallies back on the projector next to one flyer. Same technology, two very different jobs: it made something useful in twenty-five minutes, and it also quietly told the room what a scientist looks like. Holding both at once is the point of this module.

Exit ticket

Muddiest point, anonymous: what is still unclear about how these images get made? Diffusion is the topic students most often nod along to without following, and the answers tell you whether to re-explain it in ninety seconds next session.

Homework handoff

One image they are proud of, with the full prompt history that got them there, plus a real deepfake news story summarised in a sentence. The prompt history is the graded part: it is the record of iteration, which is the habit this module is actually teaching.

Preview Module 4 in one sentence: next time the thing that looks convincing is a number, and we check it.