My Presentation: the 10-Minute Fellowship Voices Slot

One full-length video, and the number is still tighter than it looks. Every second of video is a second I cannot speak.

The times below are measured, not estimated. They are counted from the actual word count of each section at 145 words per minute. At 130 wpm, which is likelier in a room this size, add about 11 percent to every number.

Twenty years of teaching, and I have always weighted the iterative process and the holistic approach over the final artifact. That part is not new.

What is new is that once AI can produce good enough work in seconds, a student who relies on it alone cannot prove how they got there. The work looks finished. The thinking is invisible, and none of my rubrics could tell me whether it happened at all.

So I started building a way to grade the thinking directly, instead of reading it off the work. Cognitive-Process Assessment. CPA.

Before I built anything, I wanted to know whether the premise held up at my own institution. James Pearce, our Technical Manager for Faculty Research Space and Emerging Technology, confirmed that FIT has no AI detection tools in place and no plan to build any.

If your campus has no detection tool either, and most don't, you are in the same position. When you can't verify AI use from the outside, the only honest path left is to grade the process itself. That's what CPA is built on.

Research starts human led. Gemini Notebook synthesizes it. Gemini and Gemini Gems direct the UX writing and strategy, with the student acting as copywriter and strategist. Students craft three art directions in Figma and choose one, and only then does Stitch generate the first UI, from a direction the student already owns. Then create visual image and photo assets with Nano Banana. Finalize the design and the craft in Figma.

Domain-specific knowledge is applied at every step, and that is what makes each decision. Students work with AI from the first day. A human makes the decision at every handoff.

Eight dimensions, and the one to name first is error catching. Did the student catch an Attribution Error, a claim the AI simply made up? Or a Logical Error, a step that doesn't follow even from true premises? That language is Google's own, from the REVEAL paper, mapped into the rubric after Session 1.

I have since added a third, from my own R&D and from an interview with one of my students. A Representational Error: when the model returns a demographically skewed default, over and over, against an explicit instruction. Not a fabricated fact, and not bad reasoning. A different failure, and it needs its own name.

So where do I check all this? In critique. In every class we discuss the what and the why of the work. Design education has run on critique for a century.

And it cannot be only written. That was the objection I worried about most, so I put it to Thalia Merino, a student of mine who had never seen my rubric.

A written log can be generated in seconds. A spoken one cannot, because I can stop a student and ask why they rejected the other three directions, and we find out in real time whether they know.

That conversation is what timestamps the record.

▶ Roll it cold. No setup, no narration. Say nothing over it. Let it finish, and let the room sit with it for a beat before speaking.

Think Gemini. A semester-long project, answering the Young Ones Google brief. Before they generated anything they ran empathy interviews with students and instructors, and built the whole thing from what they found.

Let me be clear about what this is. My students, my course, my pilot.

I guided them through the process. I did not show them my grading rubric. That's teaching.

I asked Thalia to walk me through the judgment calls in her own words.

The teacher character. The team wanted a racially diverse cast, and text prompts would not produce one. Ten to twenty generations, and she was not getting there by writing better sentences.

That is the part worth naming. Prompting is not just words and description. When words stop working you have other moves, and she used them. She stopped rewriting the prompt and changed what she fed the model: a second reference image, on top of the team's art direction reference. Style from reference one, likeness from reference two.

She diagnosed a model limitation, separated it from her own prompting, and did not accuse anyone of intent. That is critical curation, one of the eight, and that is the Representational Error I named earlier. I had already seen the pattern in my own R&D. Hearing a student describe it unprompted is what convinced me it needed a category of its own.

They wrote four KPIs to measure whether their idea worked. One of them was whether a student could explain their reasoning without looking back at what the AI had said. That is almost exactly my seventh CPA dimension. Reflective Reasoning.

So I asked her how she arrived at it. She told me she would throw it out. Too abstract, written in a time crunch while they were preparing the submission. Then she gave me four better ones, all traced back to the pain points from her own interviews.

I asked what made it obvious months later that was not obvious then. She had presented the project many times, people kept asking questions, and somewhere in answering them she realized there was a better way to think about it.

Nobody was grading her. She reopened her own decision, named why it was weak, and rebuilt it out of her research. That is reflective reasoning. It happened because people kept asking her questions, in portfolio reviews and job interviews, which is exactly why I want a rubric that asks on purpose, in every critique.

Think Gemini is one of many. And CPA is not only retrospective now: CT302, Digital Product Design I, runs its full fifteen-week semester on the eight dimensions.

That is what the workflow produces: designers empowered with AI.

CPA doesn't depend on a licensing tier or on detection software, so it can work for any program facing this gap. So I'm sharing all of it, as open source.

The CPA rubric, all eight dimensions. There's a printed copy for everyone in this room, and it's on the site. Use it and tailor it for your own class. Reweight the dimensions your course actually cares about, drop the ones it doesn't, and it doesn't have to stay in design.

[pick up the book]

And this is the book it came from. Empowering Design Education with AI, with my co-author C.J. Yeh. CPA is Chapter 2. Much of the rest belongs to other people: our students, and the industry professionals who talked honestly with us about where this is going.

If you're wrestling with the same question, how do you know a student actually did the thinking, I'd love to compare notes over the next two days. Thank you.

• Bunny Quest, by Kokona Akatsu and Haein Noh. CT&D minors personified as bunny mascots, extended into posters, stickers, social personas, and an AI-powered quiz. Gold Cube in AI Process, not AI Output
• The Socratic Problem Breakdown and the full Vulnerability Map, including the parts that are unresolved. See the Institute page
• The CT302 gradebook, all 8 dimensions scored 0 to 3, auto-converting to the 40/30/20/10 breakdown. This is the answer to "does it actually work in practice"
• WEF data: nearly half of entry-level tasks expected to be automated by 2027; design and UX sit in the moderate-substitution zone, so the cognitive layer is what's protected
• Both empathy interview write-ups
• Empowering Design Education with AI, the book behind the pedagogy

• Ten minutes still rewards one clear argument. The temptation is to fill the extra four minutes with what the six-minute version dropped. Don't. Spend all four on the case study
• Credit all four Think Gemini team members by name, on the slide and out loud. This is a team project and the room will notice which way it is handled
• Be precise about what is retrospective and what is prospective. Think Gemini was not scored under CPA, it predates the CT302 rollout. Say so. The awards prove the underlying skill is real and valued; the periodic-presentation verification layer is what CPA adds going forward
• Written in the book's voice, not a keynote voice. Where the approach is contrasted with others, stay descriptive. The room should feel invited