Empathy Interviews

The two conversations do different jobs. James Pearce gives me the institutional constraint: what FIT can and cannot do about AI, from someone who sits on the policy committee. Thalia Merino gives me the student view, plus something I did not go looking for, a documented self-correction of her own work months after the course ended.

A third and fourth interview were on the plan and I cut them on Aug 11. August availability is unreliable, and two solid write-ups beat four thin ones. If anyone asks why only two, that is the answer.

Interviewed: James Pearce, Technical Manager, Faculty Research Space & Emerging Technology, Academic Services (IT), FIT

Date: June 17, 2026 · ~28 minutes · Conducted by Christie Shin

Purpose: Kick-off Institutional Check-In for the Google Higher Ed Faculty AI Fellowship — identify FIT's primary friction point around AI: policy, academic integrity, or budget.

As part of the Google Higher Ed Faculty AI Fellowship, I was asked to speak with someone at the institutional level who works closely with emerging technology. James Pearce came to mind immediately. He sits at the intersection of academic affairs and IT at FIT, which made him exactly the right person to ground my fellowship work in the institution's current reality.

This report summarizes what I learned across four topic areas: institutional AI policy, budget and tool access, academic integrity infrastructure, and the path to scaling new tools in the classroom.

"James, I'm honored to share that I've been selected as part of the inaugural cohort of the Google Higher Education Faculty AI Fellowship. One of our first assignments is to speak with someone at the institutional level who works closely with emerging tech — you came to mind immediately. I just have a few quick questions about where FIT currently stands on AI."

Three friction areas I came in ready to probe: Policy (approved tools, data privacy, FERPA), Budget & Access (licensing, API access vs. free tiers), and Academic Integrity (AI detection ownership) — with Scaling as a bonus question if time allowed.

FIT has a comprehensive AI policy, developed in part by a committee the president tasked specifically with this work. James is on that committee, and he deliberately positioned himself on the academic side rather than the IT/compliance side, focusing on how policy shapes teaching and learning rather than just cybersecurity and data management.

Relative to other SUNY campuses, FIT is ahead. SUNY recently released framework-level guidance recommending where campuses should be, and FIT is already beyond that baseline.

The current posture leans conservative. James was direct about why: feedback from students, not about AI-generated content, but about faculty using AI in ways that affect how they communicate with and grade students. That concern has pushed institutions, including FIT, toward caution rather than openness. His word was "conservative," though he also acknowledged it could just as accurately be called cautious or careful.

This was the most concrete and practically useful part of the conversation.

Google: FIT has a SUNY-wide Gemini agreement. For most of the 2025–26 academic year, Gemini was available to faculty only; starting fall semester, it became available to students too. However, the agreement covers only a baseline tier — Gemini video generation isn't included, and NotebookLM and other Google Labs tools aren't accessible through FIT credentials. I currently pay for full access out of pocket, using my personal account.

One workaround I developed for CT380: Google offers a one-year free program for students (personal Gmail, not FIT credentials). My students signed up through that program and got access to a broader set of tools — that's how we've been doing image generation and other AI work, outside the official FIT umbrella entirely.

Adobe Creative Cloud: The SUNY agreement includes monthly AI credits that reset each month, covering a fairly wide range of Firefly features, though not every tool.

The pattern: FIT's officially supported access consistently lags behind what the tools can actually do. Faculty who want to work at the frontier are funding it themselves or building workarounds.

FIT does not currently have AI detection tools in place. There is no system to check whether student assignments were produced using AI in ways that violate course policy. James was clear and unambiguous about this.

What does exist is a review process for any new tool a faculty member wants to introduce into the classroom, covering cybersecurity and intellectual property. That process is about vetting tools before official adoption, not monitoring how students use tools once they have them.

This is a significant gap. James acknowledged it plainly: the AI policy exists, the detection infrastructure does not.

James identified the absence of a faculty sandbox as one of the most important gaps FIT needs to address. Right now, introducing any new tool into the classroom officially requires running it through legal counsel, cybersecurity review, and accessibility compliance (EITA) — the right processes for full adoption, but wrong for rapid experimentation. There is no fast lane, no protected space where faculty can test tools without triggering the full review workflow.

James used the phrase "nimble sandbox" himself — something FIT needs to build, and he was honest that they aren't there yet.

For full classroom adoption, the pathway is annual: faculty submit requests through their chairpeople every fall, and if approved, the tool is implemented the following fall — roughly a one-year minimum cycle, increasingly mismatched with tools that update weekly.

The BYOD question (Bring Your Own Device) adds another layer: the president's office is exploring whether FIT can rely less on campus lab computers, which would make version consistency even harder to control.

He is not critical of this — he understands why it happens. But the gap between what the policy says and what is actually occurring in classrooms is wide, and the institution does not yet have a framework to manage it. This is the exact problem my fellowship work addresses: CPA closes that gap through documentation, not enforcement.

• Review FIT's public AI policy documentation (James will share the link)
• Reference James's framing of the "nimble sandbox" gap in the fellowship Digital Workbook
• Use the access constraints as evidence for why CPA must be tool-agnostic and process-centered
• Cite SUNY policy alignment as institutional context when pitching CPA at FIT scale

Interviewed: Thalia Merino, Motion Design & Art Direction on Think Gemini. Team: Thalia Merino, Vashtie Persaud, Elias Cherrier-Vickers, Hailey Ho-Sang. Young Ones One Show brief for Gemini, 12 weeks.

Date: August 11, 2026 · 26 minutes · recorded · conducted by Christie Shin

Purpose: Process interview for the In-Person Institute. Think Gemini is the project carrying my presentation, and I wanted the judgment calls underneath it in Thalia's own words.

What I did not do is show them the rubric. That is the line that matters. The team never saw the eight dimensions, and they defined their own success metrics anyway. One of their four landed almost exactly on my seventh dimension. Then months after the course ended, with nobody grading her, Thalia threw those metrics out and rebuilt them from her own research. Neither of those was designed by me, and both survive the disclosure.

The external result is the measurable outcome of that R&D. Three of my students won D&AD New Blood: The Portfolios 2026 in Digital Design, Thalia among them, and her winning portfolio includes Think Gemini. Think Gemini did not win an award on its own. And to be precise about what this supports: an outside jury with no stake in my framework judged the students' work to be excellent. That is evidence the workflow produces good designers. It is not evidence that the rubric works, because no one outside my own classroom has ever scored anything with it. Those are two different claims and I only get to make the first one here.

The team documented their process and submitted it with the work. What I wanted from this conversation was the layer underneath that: what she looked at, what she threw away, and why she chose as she did. She also described a failure mode the rubric had no category for, independently of my own R&D, which is a better outcome than agreement.

I told Thalia at the top that I needed it in her words, not mine, because the whole argument of CPA is that the reasoning belongs to the student. Quotes are used with her verbal consent, and she has the write-up and the deck to review. Where the recording is unclear I have paraphrased rather than guessed, and said so.

The clearest rejection story is also the most uncomfortable one, and it is better for being uncomfortable.

The team built 3D doll-like characters. Leo, the student, came first and came easily. The teacher character did not. They had set out to build a racially diverse cast, and text prompts alone would not produce it. The failures came in two flavors: a character who read as the wrong race, or one who read far too young to be a teacher. Both were rejections against a standard the team set before they opened the tool.

Her diagnosis is the part I want in front of a room:

The fix was structural, not verbal. She stopped rewriting the prompt and changed the input architecture instead. She had been feeding the model one reference, the team's art direction. She added a second: a photograph of a specific person the character should resemble. Style from reference one, likeness from reference two. That is the moment the tool started producing usable output, and even then she kept editing by hand. Change the hair, drop the glasses. Her phrase was "a little mix and match."

Volume: 10 to 20 generations before a usable frame. Not because the prompt was bad, but because in a chat interface every small edit spawns a whole new image rather than modifying the one in front of you.

Where she stopped re-prompting. She tried the prompt first, every time. But the frames that made the final film all went into Photoshop for aspect ratio, crop, position, and removal. She framed the crop work as direction rather than cleanup: "It's more like the directing goes, maybe we want this to be a close-up shot, we want this to be more of a wide shot." Consistency across frames was never a problem, and she was specific about why. One locked visual reference, reused throughout. The consistency was a decision, not luck.

Two criteria, applied in order.

First, fidelity to a picture that did not exist yet. The team had established the narrative before generating anything: Leo at his desk, interacting with the teacher. So the question was never "which of these is good." It was:

"What is the closest thing that we could get to this visual that doesn't exist yet?"

Second, error count. She called this the negative criterion herself: "If we felt like the generation got kind of warped or looked strange, automatic, no. But if we're like, okay, maybe we could work with this, it's not there yet, then okay, this is better, we can at least edit it."

So the winning option was not the prettiest one. It was the one closest to a predetermined art direction that was also cheapest to repair. That is a working definition of Critical Curation, arrived at by a student with no rubric in front of her.

She used the image model in a chat interface first, then moved to node-based canvases. Her reason for the move is the reason I have been trying to articulate in the Orchestration Logic dimension: "I could add as many references, both visual and written, and it would take account into everything."

Chat forces sequential correction. The canvas lets her hold multiple constraints at once, which is what killed the 10-to-20-generation loop. In her words, it stopped being a constant back and forth of verbal corrections, and the results got much more accurate.

She moved between two node tools deliberately, not randomly. One could be asked to describe what it saw in an image, and she used that description as the basis for the next generation. The other handled straightforward start-frame and end-frame video work.

And when she needed a video prompt, she went outside both tools: "I didn't want to be wasteful of my generation. So sometimes I will go to ChatGPT or Gemini and say, hey, I have this in mind, help me write this in a way that keeps the image consistent or doesn't distort it in a scary way." Her own summary of the workflow was "It's a lot of using AI to speak to other AI."

I expected her to name the edit. She named the product.

"Using existing Google and Gemini's UI, our team created an ecosystem and a dashboard for our concept idea that was all created by our team."

Beyond the dashboard: custom storyboards and high-fidelity style frames for the case study, plus the vectors, tags, and gradients built to sit inside Google and Gemini brand standards. Her assessment of the ratio was that the only AI-generated aspects were the small moments where the 3D dolls kept the narrative flowing. The generated material is the connective tissue. The designed system is human.

The research underneath it. This is the part that matters most for the framework, and she needed prompting to claim it, which tells its own story about how students value their own thinking: 8 to 12 empathy interviews across instructors and students at high school and college level, affinity mapping, personas built to identify real pain points, and MVP definition with feature prioritization, all before any tool was opened.

The sequence is the argument. Research, definition, and solution came first. AI entered at production. A student who cannot do the first part cannot direct the second, and no artifact reveals which one you are looking at.

Their published takeaway was "understanding the product is as important as understanding the user," and I asked where that came from. It came from not knowing. To design for Gemini, the team first had to learn what Google's AI ecosystem actually contained, and they were surprised by the size of it: "We didn't realize that Google had so many products. We were just trying to wrap our heads around all these products and services." Tool selection as a research finding rather than a preference. That is a good answer to any faculty member who thinks tool fluency is not intellectual work.

Months after submission, unprompted, Thalia disavowed the team's original success metrics and re-derived four better ones. I asked her how she got there. The original metrics were written under submission pressure: "We were doing it in a bit of a time crunch, simply because we were trying to get submissions ready. We did pay mind to it, but not as much as we could, and likely also had AI help us a little bit to put our ideas and feelings into words."

Two mechanisms surfaced the weakness later, and both are worth naming because both are things a course can be designed to produce.

Repeated public explanation. "Since I've presented this project a million times, sometimes people ask things, or when I'm saying it, I realize maybe there's a better way to think about it or to explain it."

Re-grounding in the research. The revised metrics work because she traced them back to the pain points from the original interviews rather than writing new abstractions: "I felt that these were easier to understand, but were also relating back to those original pain points discovered back in the user research."

She did not revise because she reread the document. She revised because she had to say it out loud to people who asked questions, in portfolio reviews and job interviews after the course had ended. That is the mechanism CPA installs on purpose, in every critique, rather than leaving it to whatever happens after the course ends.

Her four revised KPIs, written to me by email on August 11, 2026, and quoted here as she sent them. This is the only place they exist in full.

1. Student adoption. "Weekly class gem utilization, or consider the percentage of learning tasks that students can complete within the integrated 'Think Gemini' environment. This is probably the most important aspect to test if our proposed learning loop would work: Will students use this to solve their work vs regular AI that gives an easy response outside of the instructor's material and control?"
2. Quality of AI responses inside the class Gems. "Such as instructor approved/aligned response rate."
3. Trust between instructors and students regarding AI.
4. Rated quality or usefulness. "The actual quality or usefulness that both students and instructors may rate the class gem and ecosystem."

Her answer was not a clean endorsement, and it is more useful for that.

Support, with one condition. From a recent graduate's vantage point, explaining decisions on a record is already the professional norm: "Sometimes when you do interviews you have to do case studies, examples or design exercises, and sometimes they ask you to record yourself going over your presentation. That's kind of the norm."

She was clear that she learned the value of this after leaving school, through portfolio reviews and job interviews, not during it. A student in the middle of the semester has no reason to see it that way yet. She also named the resistance honestly: "I can also see some people feeling not exactly willing to do it."

Her design condition: verbal, not written.

This is convergence, not correction. My rubric already verifies through periodic live presentations rather than a written log, and Thalia had never seen the rubric. She arrived at the same design decision from the student's side, for a reason I had not articulated as sharply as she did: a written trail can be generated, a spoken one cannot. Two people reasoning from opposite ends of the same problem and landing in the same place is better evidence than either of us agreeing with ourselves.

"Can instructors trust that students will do their work without using AI in an unethical way, and vice versa. That was such a big issue, that teachers were distrustful of students, and students felt, in a sense, reprimanded or falsely accused."

Her proposed mechanism is bounded access rather than prohibition: "You have access to it, but it's not free-formed. By giving them access in a slightly more controlled manner, it will help reduce the tension and create trust."

Note the symmetry in how she frames it. Transparency on both ends. The faculty member declares what is permitted, the student works inside it, and neither side is guessing about the other. Her framework and mine solve the same problem from opposite directions. She constrains the tool, I document the thinking.

Closed since this report was written: her four revised KPIs are recorded in full above, from her email of August 11, and the rejected teacher-character frame is now paired with the final on the deck. It proves the argument faster than any paragraph on this page.

One asset still outstanding:

1. A node-canvas screenshot. Shows the multi-reference structure visually and makes Orchestration Logic concrete for faculty who have never used one.

The tool stack, resolved August 20. Stills were Nano Banana. The video footage was generated with Veo, Google's own video model, and the final was assembled in After Effects. Node-based canvases sat alongside Nano Banana in the motion pipeline. That closes the item where the auto-transcript garbled her answer past reconstruction. It also matters for the argument: the generation stack is Google end to end, and the non-Google tools are craft tools rather than model substitutes. Which Veo version is still unverified, so no version number goes on a slide.

The remaining items are unconfirmed and I have listed them rather than smoothing over them. Two are places where the auto-transcript is unreliable on a specific product name, so the paraphrases above are what I am using in the meantime. The other two are the interview count and the teammate name spellings. The core argument does not rest on any of them, though the Orchestration Logic rating below does lean on one, and I have said so in the table. I would rather publish less than put a sentence in a student's mouth that the transcript cannot support.