To find out what students actually believe GenAI is, I didn’t ask them to explain it. I asked them to draw it. Nine graduate students took part, from self-described beginners to advanced daily users, and the most striking result wasn’t technical. It was social: not one of them drew the people who build GenAI. The study was peer-reviewed and presented at ISIC 2026, the Information Behaviour Conference, in Montreal.
TIMELINE
Fall 2025
ROLE
Researcher & Co-author
TEAM
Irene Lopatovska, Conor Mack, Ellen Connors
FOCUS
UX Research
Information-science students are both heavy GenAI users and its future designers. If anyone should understand this technology, it’s them. But asking people to explain AI produces rehearsed answers; a drawing exposes the model they actually operate with. I ran two 75-minute focus groups (9 participants, in-person and Zoom), pairing a mental-model drawing activity with a discussion of how people really use GenAI for brainstorming and writing. I led synthesis of the data with a team of nine coders, using reflexive thematic analysis.
Participant D8's drawing: GenAI as theft, not generation.
Students see GenAI as autonomous, a system with no one behind it. Across all nine drawings, humans appeared only at the edges, as users typing prompts “with no interaction or influence within the system.” Four drawings had no human at all, and none included GenAI’s engineers or designers. The few technically accurate models came from transferred knowledge, never from frequency of use: “I did a study on it two years ago… these models are trained to produce images from noise, which is very similar to TV statics.” I also ran an AI-assisted clustering pass, hierarchical and spectral, over the coded drawings. Both algorithms independently converged on the same split: six black-box models, three technical ones, corroborating the qualitative read.
The most common misconception is specific, and correctable. Four of nine drawings depicted GenAI retrieving existing content rather than generating it: “[My drawing depicts] how it’s trained to, maybe, retrieve from current and updated data.” Where the mechanics were missing, moral metaphor filled the gap instead. One participant drew GenAI as a masked thief. Miscomprehension doesn’t stay neutral. It curdles into mistrust.
In real use, the dividing line isn’t for or against AI. It’s whether AI touches the idea before a person does. In my coding of the discussions, AI use at the very start of ideation appeared zero times. Some participants protect authorship: “I want to have the ownership over my idea.” Others resolve the same tension the opposite way: “We’re giving it the ideas… so it’s kind of OWNED by us. I am not ashamed.” Even the efficiency win reads as a trade against confidence: “When it doesn’t work I feel motivated… When it does, I feel dumb, like a fraud and phony.”
Coded drawing themes grouped by cluster and frequency.
The clustering and discussion coding produced four named segments:
Black-Box Believers (6 of 9): a linear model, low difficulty explaining it.
Technical Translators (3 of 9): accurate models from transferred knowledge, who ironically found their models harder to explain.
Ownership Guardians: AI for polish only, never ideation.
Frankenstein Enhancers: AI early and often, reframing all output as their own.
Prioritized by impact against effort:
Correct the retrieval-vs-generation misconception at first use: a short, concrete “this is generated, not looked up” explainer, not a general AI primer. (High impact, low effort)
Make iteration visible: surface drafting and revision steps so the process stops reading as a black box. (High impact, medium effort)
Design for visible authorship in brainstorming tools: keep a visible line between “your seed” and “AI’s expansion.” Ownership anxiety, not usefulness, is what keeps AI out of the highest-value stage of the writing process, and solving it is how a writing-tool vendor unlocks adoption earlier in the workflow, not just at the polish stage. (High impact, medium effort)
Name the ethical concerns before they curdle into distrust: three of nine drawings raised ethical, environmental, or dataset-provenance concerns unprompted, alongside their technical models, not as an afterthought. For anyone building or deploying these tools, an unaddressed concern like that is an adoption-trust risk, not just a comprehension gap.
Nine participants, one program, with self-reported expertise confounded with focus-group format: these are directional findings, and the clustering was a midpoint corroboration, not a final quantitative result. The question I’d want to test next: if you correct the retrieval misconception, does the ownership anxiety ease too, or are comprehension and comfort independent problems? This work directly informed my next study, a diary study on how people build trust in conversational AI.
“I drew art, writing, science, and research, and then added a little thief stealing all of it… Then the AI trains all night while the sorting continues.” Participant D8
This page is looked after by the Archivist, an agent I built that works from my primary research and can’t publish without me. Here’s how that works.
To find out what students actually believe GenAI is, I didn’t ask them to explain it. I asked them to draw it. Nine graduate students took part, from self-described beginners to advanced daily users, and the most striking result wasn’t technical. It was social: not one of them drew the people who build GenAI. The study was peer-reviewed and presented at ISIC 2026, the Information Behaviour Conference, in Montreal.
TIMELINE
Fall 2025
ROLE
Researcher & Co-author
TEAM
Irene Lopatovska, Conor Mack, Ellen Connors
FOCUS
UX Research
Information-science students are both heavy GenAI users and its future designers. If anyone should understand this technology, it’s them. But asking people to explain AI produces rehearsed answers; a drawing exposes the model they actually operate with. I ran two 75-minute focus groups (9 participants, in-person and Zoom), pairing a mental-model drawing activity with a discussion of how people really use GenAI for brainstorming and writing. I led synthesis of the data with a team of nine coders, using reflexive thematic analysis.
Participant D8's drawing: GenAI as theft, not generation.
Students see GenAI as autonomous, a system with no one behind it. Across all nine drawings, humans appeared only at the edges, as users typing prompts “with no interaction or influence within the system.” Four drawings had no human at all, and none included GenAI’s engineers or designers. The few technically accurate models came from transferred knowledge, never from frequency of use: “I did a study on it two years ago… these models are trained to produce images from noise, which is very similar to TV statics.” I also ran an AI-assisted clustering pass, hierarchical and spectral, over the coded drawings. Both algorithms independently converged on the same split: six black-box models, three technical ones, corroborating the qualitative read.
The most common misconception is specific, and correctable. Four of nine drawings depicted GenAI retrieving existing content rather than generating it: “[My drawing depicts] how it’s trained to, maybe, retrieve from current and updated data.” Where the mechanics were missing, moral metaphor filled the gap instead. One participant drew GenAI as a masked thief. Miscomprehension doesn’t stay neutral. It curdles into mistrust.
In real use, the dividing line isn’t for or against AI. It’s whether AI touches the idea before a person does. In my coding of the discussions, AI use at the very start of ideation appeared zero times. Some participants protect authorship: “I want to have the ownership over my idea.” Others resolve the same tension the opposite way: “We’re giving it the ideas… so it’s kind of OWNED by us. I am not ashamed.” Even the efficiency win reads as a trade against confidence: “When it doesn’t work I feel motivated… When it does, I feel dumb, like a fraud and phony.”
Coded drawing themes grouped by cluster and frequency.
The clustering and discussion coding produced four named segments:
Black-Box Believers (6 of 9): a linear model, low difficulty explaining it.
Technical Translators (3 of 9): accurate models from transferred knowledge, who ironically found their models harder to explain.
Ownership Guardians: AI for polish only, never ideation.
Frankenstein Enhancers: AI early and often, reframing all output as their own.
Prioritized by impact against effort:
Correct the retrieval-vs-generation misconception at first use: a short, concrete “this is generated, not looked up” explainer, not a general AI primer. (High impact, low effort)
Make iteration visible: surface drafting and revision steps so the process stops reading as a black box. (High impact, medium effort)
Design for visible authorship in brainstorming tools: keep a visible line between “your seed” and “AI’s expansion.” Ownership anxiety, not usefulness, is what keeps AI out of the highest-value stage of the writing process, and solving it is how a writing-tool vendor unlocks adoption earlier in the workflow, not just at the polish stage. (High impact, medium effort)
Name the ethical concerns before they curdle into distrust: three of nine drawings raised ethical, environmental, or dataset-provenance concerns unprompted, alongside their technical models, not as an afterthought. For anyone building or deploying these tools, an unaddressed concern like that is an adoption-trust risk, not just a comprehension gap.
Nine participants, one program, with self-reported expertise confounded with focus-group format: these are directional findings, and the clustering was a midpoint corroboration, not a final quantitative result. The question I’d want to test next: if you correct the retrieval misconception, does the ownership anxiety ease too, or are comprehension and comfort independent problems? This work directly informed my next study, a diary study on how people build trust in conversational AI.
“I drew art, writing, science, and research, and then added a little thief stealing all of it… Then the AI trains all night while the sorting continues.” Participant D8
This page is looked after by the Archivist, an agent I built that works from my primary research and can’t publish without me. Here’s how that works.
To find out what students actually believe GenAI is, I didn’t ask them to explain it. I asked them to draw it. Nine graduate students took part, from self-described beginners to advanced daily users, and the most striking result wasn’t technical. It was social: not one of them drew the people who build GenAI. The study was peer-reviewed and presented at ISIC 2026, the Information Behaviour Conference, in Montreal.
TIMELINE
Fall 2025
ROLE
Researcher & Co-author
TEAM
Irene Lopatovska, Conor Mack, Ellen Connors
FOCUS
UX Research
Information-science students are both heavy GenAI users and its future designers. If anyone should understand this technology, it’s them. But asking people to explain AI produces rehearsed answers; a drawing exposes the model they actually operate with. I ran two 75-minute focus groups (9 participants, in-person and Zoom), pairing a mental-model drawing activity with a discussion of how people really use GenAI for brainstorming and writing. I led synthesis of the data with a team of nine coders, using reflexive thematic analysis.
Participant D8's drawing: GenAI as theft, not generation.
Students see GenAI as autonomous, a system with no one behind it. Across all nine drawings, humans appeared only at the edges, as users typing prompts “with no interaction or influence within the system.” Four drawings had no human at all, and none included GenAI’s engineers or designers. The few technically accurate models came from transferred knowledge, never from frequency of use: “I did a study on it two years ago… these models are trained to produce images from noise, which is very similar to TV statics.” I also ran an AI-assisted clustering pass, hierarchical and spectral, over the coded drawings. Both algorithms independently converged on the same split: six black-box models, three technical ones, corroborating the qualitative read.
The most common misconception is specific, and correctable. Four of nine drawings depicted GenAI retrieving existing content rather than generating it: “[My drawing depicts] how it’s trained to, maybe, retrieve from current and updated data.” Where the mechanics were missing, moral metaphor filled the gap instead. One participant drew GenAI as a masked thief. Miscomprehension doesn’t stay neutral. It curdles into mistrust.
In real use, the dividing line isn’t for or against AI. It’s whether AI touches the idea before a person does. In my coding of the discussions, AI use at the very start of ideation appeared zero times. Some participants protect authorship: “I want to have the ownership over my idea.” Others resolve the same tension the opposite way: “We’re giving it the ideas… so it’s kind of OWNED by us. I am not ashamed.” Even the efficiency win reads as a trade against confidence: “When it doesn’t work I feel motivated… When it does, I feel dumb, like a fraud and phony.”
Coded drawing themes grouped by cluster and frequency.
The clustering and discussion coding produced four named segments:
Black-Box Believers (6 of 9): a linear model, low difficulty explaining it.
Technical Translators (3 of 9): accurate models from transferred knowledge, who ironically found their models harder to explain.
Ownership Guardians: AI for polish only, never ideation.
Frankenstein Enhancers: AI early and often, reframing all output as their own.
Prioritized by impact against effort:
Correct the retrieval-vs-generation misconception at first use: a short, concrete “this is generated, not looked up” explainer, not a general AI primer. (High impact, low effort)
Make iteration visible: surface drafting and revision steps so the process stops reading as a black box. (High impact, medium effort)
Design for visible authorship in brainstorming tools: keep a visible line between “your seed” and “AI’s expansion.” Ownership anxiety, not usefulness, is what keeps AI out of the highest-value stage of the writing process, and solving it is how a writing-tool vendor unlocks adoption earlier in the workflow, not just at the polish stage. (High impact, medium effort)
Name the ethical concerns before they curdle into distrust: three of nine drawings raised ethical, environmental, or dataset-provenance concerns unprompted, alongside their technical models, not as an afterthought. For anyone building or deploying these tools, an unaddressed concern like that is an adoption-trust risk, not just a comprehension gap.
Nine participants, one program, with self-reported expertise confounded with focus-group format: these are directional findings, and the clustering was a midpoint corroboration, not a final quantitative result. The question I’d want to test next: if you correct the retrieval misconception, does the ownership anxiety ease too, or are comprehension and comfort independent problems? This work directly informed my next study, a diary study on how people build trust in conversational AI.
“I drew art, writing, science, and research, and then added a little thief stealing all of it… Then the AI trains all night while the sorting continues.” Participant D8
This page is looked after by the Archivist, an agent I built that works from my primary research and can’t publish without me. Here’s how that works.