P18: Survey Data and Session Summary

Survey Responses

Background

P18 is a Research Integrity Program Manager at a public university, with 8 to 15 years of professional experience and a master's degree. Her career spans 10 years in neuroscience before her current role, plus active contract work as a UX researcher running interviews for a company that develops niche AI chatbots. The combination is unusual in this study because she occupies two AI-adjacent positions simultaneously: she is a constrained institutional user during the day, and on the side she is one of the humans inside the loop that produces AI agents for other companies.

Her oh-wow moment with AI was a randomization spreadsheet formula. She needed to map 600 subject IDs to randomized numbers and match them across sheets, work she could not figure out how to do herself. ChatGPT solved it instantly. From that moment forward, her use has grown into a two-tier practice: she operates the university-sanctioned Google Gemini for screen-aware tasks her institution monitors, while reaching for her personal ChatGPT account for the work she actually depends on. The two-tier split is forced by the institution's HIPAA-bound contract with Google, which prohibits training the institutional LLM on user data. The institutional model is "a new user each time." The personal model accumulates context across sessions and functions, in her words, like a "best friend."

A central part of her day job is research determination: deciding whether a proposed faculty study is exempt, non-human-subjects, or limited IRB. She has tested AI on this task directly and accumulated hard evidence that AI cannot do it. The reason is structural: complete answers require combining federal regulations with institution-specific protocols, and the LLMs she has access to cannot be trained on either layer. This concrete failure anchors her broader view of organizational AI adoption and explains why she is trying to join the university's AI team alongside the VP of Research Integrity.

Key Findings

The Two-Tier AI Practice

P18 names the operational reason the institutional AI is operationally useless for her real work: the university's contract with Google prohibits using user data to train the LLM, so every session starts cold. She contrasts that experience directly with the personal ChatGPT account that does accumulate context, and she does not attempt to soften the comparison.

"That's why I love ChatGPT, because it's like my best friend. It knows everything about me. But with [university] systems, unfortunately, we just can't have that quality of training because of HIPAA compliance."

This is Corporate Tooling Gap where the gap exists not because the institution is indifferent or behind, but because the regulatory regime governing the institution makes the institutional tool unable to perform the role users need it to play. The university actively wants AI integrated. The compliance posture makes the institutional version a stranger every time, so her real work routes to the personal account.

AI Cannot Make the Compliance Determinations Her Job Requires

P18 has tested AI on the IRB research determination question directly and reports a concrete failure mode. The model defaults to "limited review" across the board, which is the wrong answer at her institution. She names the structural reason: the determinations require combining federal regulations with institution-specific protocols, and the LLM cannot be fine-tuned on either.

"We have so many human subjects research determinations at [university] because of the type of research that happens here. And our chat bots are just not good at that. They dictate everything as limited review, which is not like it sounds. It's a small review, but it actually contains the whole committee and everything, and that's incorrect. It's filling in gaps where it doesn't need to. At [university] we can't train the model, so we can't get it really fine-tuned to what our determinations actually are."

This is a specific failure surface inside Organizational AI Adoption Challenges worth tracking across future sessions: complex multi-source compliance determinations where the LLM cannot be trained on either the regulatory layer or the institutional-protocol layer. It also surfaces in her survey response, where she names the same failure mode as the reason she is trying to join the AI team: "these bots cannot make informed institutional decisions that are also founded in federal regulations in order to properly filter our faculty's research protocols."

The Recursion Vertigo

P18's side work as a contract UX researcher places her inside a recursive AI loop. The companies that hire her firm provide briefs of what they want to know about a market. Her firm interviews experts. The interviews feed the LLM. The LLM-trained chatbot goes back to the company. She is the human-in-the-loop in a process designed to remove humans from the loop. The recursion reaches into her email behavior too, where she registers that her ChatGPT-softened emails to AI-using recipients amount to "two AI bots talking to each other."

"It's like AI agents talking to AI agents. And I just start to feel like I'm going crazy because it's very Matrix-y, and then I realize when I'm emailing people, I'm like, man, it's just two AI bots talking to each other at this point."

She compares the felt experience to what her parents must have felt during the early internet: "It just feels like there's this whole world being created that's not real." This is the new theme this session contributes to the codebook: AI Recursion Vertigo. It is distinct from Synthetic Validation Recursion (P17), which describes a methodological problem (AI validating AI). AI Recursion Vertigo describes a worker's experiential concern: the felt sense that the human-purposeful layer of work is dissolving as agent-to-agent loops thicken. P18 is the seed session for this theme.

Neural Pathways and the Bike-Riding Exception

P18 brings her neuroscience background directly into her concern about skill erosion. She names the mechanism: cognitive tasks shortened down to a single AI query weaken the neural pathways that previously carried them. She offers the bike-riding pathway as the structural counterexample: motor pathways recruit so much of the body that they survive disuse, even if the rider grows unsteady.

"I spent 10 years working in neuroscience. And so, just understanding the development of neural pathways, I know for a fact we are weakening those neural pathways by shortening the process. Like, learning how to ride a bike, you don't forget, because it requires so much motor skill and functioning that that pathway never gets weakened. I mean, it can get weakened, so you're unsteady on the bike, but you still know how to ride it at the end of the day, because it requires so much from your body in each piece. So when you're taking all these complex tasks and shortening them down to just asking one question, you're shortening that path."

This is the sharpest articulation of Skill Erosion in the corpus so far. It does more than name the worry. It identifies which pathways are vulnerable (purely cognitive, single-query-shortenable) and which are not (motor-procedural, multi-system). P18's Self-Maintenance practice (pen-and-paper design intake before opening ChatGPT) is her response to this mechanism, though she is the first to acknowledge the practice fails under time pressure.

"I really try to continue brainstorming design on pen and paper. Like, starting my initial outline on pen and paper, and then maybe having a discussion with ChatGPT about how the information should go in. And even then, I still feel a little weird, because I don't want it to weaken that creativity or those skill sets. I'm sure there's other things too, but time management really is the key, and that forces me to continue using it."

The Academic Disclosure Taboo

P18 names a specific mechanism for the AI-disclosure taboo that other sessions had not articulated: intellectual elitism. In an environment where intellectual labor performs intellectual identity, ceding labor to AI reads as ceding competence. The contrast she offers is her previous institution publishing a ChatGPT-written newsletter and bragging about it. The bragging-as-counterexample is significant because it shows the norm is malleable; an institution can choose to celebrate AI-assisted work openly, and her previous one did.

"Especially in academia, when elitism is really strong, and it's an intellectual elitism. The idea of somebody else doing your work is really taboo. And I think my previous institution handled it really well. When ChatGPT was becoming a household name, they sent out a newsletter and bragged, like, 'this was written by ChatGPT, look what we can do.' And I thought that was a cool way to do it, but there's still so much resistance."

This extends Disclosure Norms with a mechanism that helps explain why disclosure resists administrative push. Administrators can ask for disclosure all they want; the underlying identity work in academia treats prompt engineering as a confession rather than as a skill. P18 thinks that should change ("I think that should be on a resume, as a matter of fact"), but she does not expect it to change quickly.

Spotting the Fresh-Grad Slop

As a former manager of fresh-grad direct reports, P18 has a concrete catalog of the signals that mark email as AI-written: em dashes, weird bolding, overvalidating language. The catalog is unusual in the corpus because it is so specific.

"In my previous job I had a few direct reports, and they were always fresh grads. This last round, like 20 to 22, was the age range. And I could see every email I was getting was written by AI. I mean, no filter. Just em dashes and weird bolding, you know, just that weird overvalidating language."

This is AI Slop Detection on the manager-receives-from-subordinate axis, which is a direction the codebook has not yet centered (most prior coding has been on the subordinate-receives-from-supervisor axis). She holds the concern alongside a tempered observation that the time these juniors save may open as many doors as it closes. The same observation reappears in her closing thoughts on the next generation, where she names her Apprenticeship Erosion concern about whether her former employees apply critical thinking to patient-facing decisions, then asks the meta-question about her own undergraduate education as the source of that critical-thinking capacity.

"If we integrate so much AI and our critical thinking skills are leaned on, like, okay, I asked this question, let me just go check with whatever chatbot. I wonder, when does that critical thinking get diminished?"

The Hope is Medical

When asked about her single biggest hope for AI, P18 frames it in family-medicine terms. A child gets a profile created at birth, the practice tracks risk signs across years (genetic markers, family history, early indicators), and AI catches signals before patterns become chronic conditions. Her examples are addiction and mental health, both rooted in her neuroscience background.

"Imagine if we could track that in this patient and address those early on, before we see any onset, before someone is deep in a repeated cycle of addiction or something. I know that's the hope in those fields, and I really pray that that outcome is what we think and want it to be, rather than a slew of incorrect diagnostics and just like the WebMD of the world."

The framing pairs concrete clinical specificity ("a profile created" at birth, "genetically") with concrete failure modes ("the WebMD of the world"). She is not coded for a separate theme on medical hope because the corpus does not yet have a second session that articulates this in this form. It is worth surfacing as a watch item.

Emerging Themes

P18 contributes one new theme to the codebook: AI Recursion Vertigo. The theme names the experiential side of the agent-economy concern that prior sessions had only approached methodologically. P17's Synthetic Validation Recursion is about validating AI with AI as a project-risk problem; P18's AI Recursion Vertigo is about the worker's felt sense that the human-purposeful layer of work is dissolving into agent-to-agent loops. Both apply at the same moment in the industry; they describe different observers' positions inside it.

P18's AI as Cognitive Prosthetic evidence is unusually clean. She names the limitation she is compensating for ("a very direct communicator"), names the alternative failure mode she also exhibits ("sometimes I get vague, like I'm too vague"), and names the mechanism by which ChatGPT helps (no personal opinion, no concern about reception, "no emotional additive"). The use is functional rather than aspirational, and it ties directly to her job: most of her emails are to principal investigators, where the cost of mis-toned communication is measurable in protocol-review delays.

"It's just a third party without that emotional additive."

P18's Corporate Tooling Gap is grounded in a specific cause (HIPAA-bound training restriction) rather than in vague institutional drag. The institution is actively pushing adoption (the Innovate Academy, multi-party president-level approval); the institutional tool simply cannot do what users need because of the compliance posture. This is a different shape than the cases in P3 and P11, where the gap was more about provisioning lag or vendor mismatch. Worth tracking whether this regulator-bound variant recurs in healthcare or finance sessions.

"Any bot we use or any agent we use can't reuse the information to train the LLM."

P18's Organizational AI Adoption Challenges evidence is centered on a specific failure mode worth a theme entry on its own if it recurs: AI cannot make complex multi-source compliance determinations because models cannot be trained on either the regulatory layer or the institution-specific protocol layer. She has the concrete test results to back the claim ("they dictate everything as limited review"), and she has the structural account of why ("regulations are so complex and they're so diverse, dependent on the institution").

"Regulations are so complex and they're so diverse, dependent on the institution."

P18's Hallucination Frustration contribution is the misattribution variant: the cited reference exists but does not contain what the model claims it does. She draws the distinction explicitly: "more of a misuse of information than actually fabricating." Worth tracking whether this is a usable refinement to the theme description in future passes.

"Like the reference will exist but won't say anything about what ChatGPT is claiming."

P18's Disclosure Norms contribution is the academic-elitism mechanism plus the previous-institution counter-example. The mechanism explains why disclosure resists administrative push (it threatens identity, not just process), and the counter-example shows that the norm is contingent (her prior institution did the opposite and it worked).

"But people are so resistant to saying and giving the idea that someone else did this work for me."

P18's Apprenticeship Erosion evidence is held alongside her AI Slop Detection catalog and her Knowledge Displacement framing. The three coexist in her observations of fresh grads as direct reports: she could see the AI use (slop detection), she worries about whether they're developing critical thinking (apprenticeship erosion), and she traces the capacity she fears they're missing to her own undergraduate education (knowledge displacement).

"I just think about my time and how much mine has been relieved because of simple things like email drafting and spreadsheet building and so on. That new generation of people who can just utilize AI as an assistant, I don't want to criticize them, because I do think that there's a huge opening of what else they can do on top of that."

P18's Skill Erosion evidence is the most mechanistically specific in the corpus so far. The bike-riding contrast is not a casual metaphor; it is a claim about which neural pathways atrophy under disuse and which do not. Worth noting that her Self-Maintenance practice acknowledges its own fragility: "time management really is the key, and that forces me to continue using it."

"When you're taking all these complex tasks and shortening them down to just asking one question, you're shortening that path."

P18's Augmentation Not Replacement evidence is in the survey response more than the transcript. The website rebuild prompts she shares are a specific instance of structured augmentation: paper outline first, then ChatGPT for structure, then explicit "DO NOT change content" guardrails repeated across iterations because the model otherwise extrapolates. The practice is operational rather than philosophical.

"I uploaded the drawn image and said 'put this information into my design. DO NOT change any content,' and then I would give it the site link again, because sometimes if I don't remind it, it will make up content."

P18's Trust Calibration evidence absorbs what an earlier coding round might have called a sounding-board behavior. She uses ChatGPT for second opinions on research determinations while knowing the answer will be unreliable, because the act of bouncing has value even when the output does not. The grad-school disaster ("I got slaughtered on that grade") is referenced as the calibration event that taught her where the limit is.

"It's not so good at that, but at least I get a second opinion to bounce off of."

Interview Transcript

00:00:00

Paul: Okay. So, I'd like to start off by asking you, what was your first "oh wow" moment with AI? What was going on that made you try AI, and what happened that made the light bulb turn on for you?

P18: That's a good question. I think the first... oh god. Probably for work. Before this current job at [university], I was at [previous university]. And when I was there, I think the first time I used AI was honestly to help write an email or something. And how quick it came, that was an oh wow. But really when it stunned me was when I had to create a spreadsheet and there was a formula I could not figure out. I could not figure out how to change the subject ID to a randomized number for like 600 fields, and I needed it to create that randomized number and then match to another sheet. And ChatGPT just did it like that, and I was like, oh, the world is endless.

00:01:13

Paul: Is there anything that you do regularly now at work or in your personal life that AI has changed massively? Walk me through what you used to do versus what you do now.

00:02:31

Paul: I'm going to go off script after only two questions and just mention that I found something today called Feynman.is, as in Richard Feynman. It's a purpose-built research AI tool, an LLM, and it runs locally, but it goes off and hits the arXiv servers. Maybe that would be useful for you.

P18: Yeah, that's it. I'm writing it down.

00:03:40

Paul: I want to talk a little bit about how your organization is handling AI adoption. Are there official policies? Are there people figuring it out on their own? I know that as a university there may be a bit more latitude, but I'd love to hear about the decision-making process, and how it's socialized, and how it's actually working out for you and others.

00:04:54

P18: There's like three different sections of it. There's the administrative section, there's the research section, and then I think technology. And it's just learning how to use [university]'s AI systems, which is predominantly Google. How to use the AI systems in this. And then there's a couple of other technologies in there that I'm unaware of yet because I haven't taken the Innovate Academy. I'm going next quarter. But there's a couple of other softwares in there. I want to call one OneNote, but it's not called OneNote. It's something Notebook.

Paul: Is it NotebookLM?

Paul: Has the university experienced AI rollouts that worked really well, or maybe that caused more challenges than expected? Can you tell me about that?

P18: Yeah. Let's use Gemini for example. I'm not sure when the rollout happened. It wasn't that long ago, within the past year or two. I haven't been here that long. I've been here for a couple of months now. But universities are slow adapters.

00:07:28

Paul: I want to ask what your biggest win has been with AI at work so far. Is it the email time savings, or is it something else?

Paul: Tell me if I'm wrong, but I'm thinking that a lot of your emails are back and forths between investigators?

P18: A lot of it is investigators. Some of it is research and departmental admin. Some of it's for sponsored projects, so financials.

00:09:48

P18: But yeah, lots of principal investigators. That's predominantly who I'm emailing with.

Paul: What do you think's been your biggest disappointment or surprise failure when you try to use AI at work?

P18: At work specifically.

Paul: It doesn't have to be. If there isn't any at work, you could switch over to personal.

P18: Okay. I have a really specific example for personal, and then it might spawn some ideas for at work. But, oh, I guess I do have one at work. I'll start with that. So when it comes to research determinations, like I said, determining if it's non-human subjects, if it's exempt, if it's limited IRB, it's really bad at that. Like, it's really bad because regulations are so complex and they're so diverse, dependent on the institution. Because we have our federal regulations, but then we have our institutional regulations that kind of mix in with those.

00:10:49

Paul: And you said you had a personal example as well.

P18: Yeah. It's embarrassing because it was in grad school. I was very overworked in grad school. I was working full-time, and I had an assignment that I really relied on ChatGPT for, and it was horrific.

00:11:59

P18: I just didn't review it enough. I threw it together. I did contrast checking and everything through ChatGPT, and it was just horrible. I got slaughtered on that grade. One of the quotes was "although I appreciate creativity, it needs to be in line with some standards." It was just embarrassing, but that's where I think human intervention is going to always have to be there.

Paul: I think we've all had the experience of being time pressured and wanting to shortcut an assignment and getting maybe a little slapped by the instructor. And with AI, it's probably even more intense, the temptation.

P18: And that was the first and last time I'll do that. I know it's a huge problem with professors right now dealing with students turning in a lot of AI-written work. And I can imagine, as a student, because I lean on it so much in my day-to-day for work, just with the time crunch.

00:13:22

Paul: Have you encountered anyone who resists using AI? What's that like?

P18: Yes.

Paul: Tell me about that.

P18: So much resistance. Like, incredible amounts of resistance. I would support it if people had more explanation for why. I think right now you have two sides of the spectrum in my field, where it's people who are really hard about AI, really excited, like it's going to change our lives, it's going to change everything. And then you have other people who are like, it is the devil. It will take all of our information and it will ruin research as a whole, which I don't totally disagree with. I know that a lot of principal investigators are dealing with chat bots responding to surveys. It happens in UX, it happens in every incentivized survey.

00:14:28

P18: And so I think a lot of principal investigators are against it because of that fear. And then also I think people are really worried about privacy, even though they use Google on the daily and they have iPhones. They're really concerned about privacy from a generalized standpoint. I haven't known anyone who's able to pinpoint AI as being this specifically dangerous asset in data collection compared to what else we have. People using Facebook and then resisting against AI is just contradictory to me.

Paul: Are you seeing unwritten rules or norms forming at the workplace about when to disclose that AI helped you with something?

Paul: I want to ask about what I'm calling skill erosion. Do you worry about losing certain skills because you're leaning on AI?

P18: Yes.

Paul: Which ones?

00:17:18

Paul: Is there anything that you still do the hard way to keep the skill sharp?

00:18:23

Paul: Right. All right. Let's zoom out and up for a minute. So, how does this increasing presence of AI in your work and personal life make you feel?

00:19:46

00:21:15

Paul: What's your single biggest fear when it comes to AI in the future?

P18: Well, I worry about human reliance on AI a lot. We've already seen [UNCLEAR] like the dissolving of the renaissance man, you know? And so I just worry that even more unique talents and aspects will be diminished because of reliance on it. But also the job market, I think, is the biggest concern, which I know is really common. But I also do believe that there is a new job market being created because of AI. And I think companies are using AI as a scapegoat for layoffs, because it does create so many new jobs.

00:22:42

P18: So I don't know how much it's founded in reality, but those are my concerns.

Paul: What's your biggest hope for what AI might deliver or enable in the future?

P18: My biggest hope is for medical research and diagnostic capabilities and patient tracking systems. I think that there is so much in the way of what can be done. Take for example, you have a family medical practice. You start going there when your child is born, and that child gets a profile created. And let's say we know all these early risk signs of, I don't know, say addiction. We know what that looks like genetically. We know what it looks like with family history. Or say early indicators of, I don't know, just mental health disorders or whatnot. Imagine if we could track that in this patient and address those early on, before we see any onset, before someone is deep in a repeated cycle of addiction or something. I know that's the hope in those fields, and I really pray that that outcome is what we think and want it to be, rather than a slew of incorrect diagnostics and just like the WebMD of the world. That's what I worry about. But I have a lot of hope for that field.

Paul: I just want to talk a little bit about the next generation of people entering the workforce now, who may never have done work without AI. What concerns you or excites you about that?

P18: This is something I think about daily.

00:25:27

P18: I can say that we're weakening all these pathways, but then think about how much that opens up for other things. I just think about my time and how much mine has been relieved because of simple things like email drafting and spreadsheet building and so on. That new generation of people who can just utilize AI as an assistant, I don't want to criticize them, because I do think that there's a huge opening of what else they can do on top of that.

Paul: That thought that you just expressed, I've heard from nearly everyone, and usually people accompany it with, "but there's always been a change that happens, whether it's typewriter to computer or computer to internet." I don't know. I don't have a strong opinion about what's going to happen, because I'm watching it all happen in real time like the rest of us. The hope that most people express is that, well, some people are naturally predisposed to being critical thinkers, and that's something we can also train and teach.

00:26:55

P18:

Paul: I appreciate you walking through these questions with me.