What we found, and how we found it

How HKS Students
Are Navigating AI

This site presents the analysis of ten in-class focus groups on how HKS students are navigating AI, and uses that analysis to demonstrate what different qualitative coding methods can and cannot do. (An eleventh section ran the protocol but was not recorded.)

10 focus groups · coded utterances · last build
↓ read
a letter from Prof. McKenna
loading memo…
read the rest of the letter →← collapse
01 / methods

How this analysis was conducted.

protocol
A 30-minute in-class focus group run by student moderators. Seven sections: opening, go-around, group reaction, two-sided risk vote, professional readiness discussion, pair-share, dot voting.
data
Ten in-class focus groups, audio-recorded and transcribed (Whisper), paired with hand-written field notes from the designated note-taker in each room. One additional section was not recorded. Participant names pseudonymized.
coding
A 20-code a-priori deductive framework applied by an LLM (Claude), with human inductive coding surfaced rather than reconciled. One focus group was also coded inductively by hand as a demonstration.
positionality
The instructor wrote the protocol, taught the students in the room, and is now analyzing what they said, which shapes what gets seen.
/ votes

What the rooms voted for.

Two-sided risk split and pair-share dot totals, hand-entered from notes and flip-chart photos.

no vote tallies entered yet.
02 / method

There are at least five ways to code the data, and each approach yields slightly different findings.

Every method the site uses has a specific shortcoming.

methodwhat it's good athow it breaks
human deductive fairly fast blind to what you didn't ask
human inductive / in-vivo preserves voice, surfaces surprise doesn't scale, not reproducible across coders
LLM deductive fast and consistent when given a real codebook homogenizes, loses affect, invents plausible codes
LLM inductive generates themes from scratch in seconds generic themes, no codebook trail, no in-vivo preservation
embeddings / topic models "objective"-looking (probably) recovers structure you already knew — often a function of how the protocol was designed

Triangulating across methods is expensive but is sometimes the only defensible approach for a public artifact, peer-reviewed research, or high-stakes real-world analysis.

03 / same exchange, three codings

What the coders saw.

One six-turn exchange from FG-10 Sanchez (§4, two-sided risk vote), read three ways: by an LLM applying the a-priori codebook, by the same LLM generating themes from scratch, and by a human reading closely for what the other two missed.

← collapse the three codings expand the three codings →
Lens 1 · LLM deductive

LLM deductive coding is fast, (usually pretty) accurate, and unsurprising.

MOD"anyone wants to defend their vote?"

A"I always feel guilty when I use it… I love writing, I think writing is such a good way to learn… I just struggle with wanting to feel like I'm learning and not using a crutch."

B"I'm doing research on intellectual atrophy from AI… I think not using AI is more costly than learning how to use AI in a way that is still conducive to your learning… that's kind of like the thesis of the paper."

C"My partner spends eight hours a day using Fluid to build different things… that is something I would not be able to do right now… I think there's also a gender gap emerging in how we're using AI… everyone who is not capable might be at risk of losing their job."

A"I got to Harvard, I did it… what if I get expelled because I used AI in the wrong way… I felt very proud of things I've turned in where I haven't touched it at all."

D"I'm taking some writing course where I have to refrain myself… I'm struggling not using it for writing course because that's one of the major things I wanted to learn here."

Lens 2 · LLM inductive

LLM inductive coding summarizes rather than discovers.

MOD"anyone wants to defend their vote?"

A"I always feel guilty when I use it… I love writing… the cleaning up is so helpful… I just struggle with wanting to feel like I'm learning and not using a crutch."

B"I'm doing research on intellectual atrophy from AI… I think not using AI is more costly than learning how to use AI… that's the thesis of the paper."

C"My partner spends eight hours a day using Fluid to build different things… that is something I would not be able to do right now… I think there's also a gender gap emerging in how we're using AI."

A"I got to Harvard, I did it… what if I get expelled because I used AI in the wrong way… I felt very proud of things I've turned in where I haven't touched it at all."

D"I'm taking some writing course where I have to refrain myself… I'm struggling not using it for writing course."

LLM-generated themes look inductive but they aren't emerging from the data — they reflect patterns the model learned from vast prior texts about AI, anxiety, gender, and work. Without a codebook, the model defaults to the statistically common frames.
Which means it can miss moments where the data resists easy categorization. Those moments of friction are where real insight lives.
"Gendered Technology Access" may be technically accurate, but it overlooks how the dynamic plays out inside a specific personal relationship and everyday life.
Lens 3 · Human in-vivo

The human reader discovers what she did not expect.

MOD"anyone wants to defend their vote?"

A"I always feel guilty when I use it… I love writing, I think writing is such a good way to learn… I just struggle with wanting to feel like I'm learning and not using a crutch."

B"I'm doing research on intellectual atrophy from AI… I think not using AI is more costly than learning how to use AI in a way that is still conducive to your learning… that's the thesis of the paper."

C"My partner spends eight hours a day using Fluid to build different things… that is something I would not be able to do right now… I think there's also a gender gap emerging in how we're using AI."

A"I got to Harvard, I did it… what if I get expelled because I used AI in the wrong way? I felt very proud of things I've turned in where I haven't touched it at all."

D"I'm taking some writing course where I have to refrain myself… I'm struggling not using it for writing course because that's one of the major things I wanted to learn here."

Bonus · a code the LLM never would have written

The "English/ChatGPT accent" — or, The Little Mermaid code.

From the note-taker in FG-05 Ge:

"In the case of using AI to help write papers in English (the participant was not a native English speaker), one participant mentioned the risk of developing an 'English/ChatGPT accent'; people laugh."

Three different things you see in the same six turns, depending on what method you use to code.

/ field notes

What the transcripts didn't catch.

A note-taker in each focus group wrote down what the microphone couldn't hear: who laughed, who went quiet, who crossed their arms, who leaned in. Below are four moments (this is just a selection), paired with the exchange in the transcript.

no pairings yet — run make stage01e.
06 / executive summary

Four preliminary findings.

Grounded in 67 votes on the two-sided risk question, 90 dot-vote items aggregated across ten focus groups, and 1,565 coded utterances (out of 3,674 transcribed segments).

  1. The dominant emotional register is anxiety, not enthusiasm (but the protocol primes for this). When asked which risk feels more present, 57% of students named over-use (their own learning, writing, and judgment) over the 43% who named under-use (falling behind peers and employers). Across the corpus, anxiety codes — about lost skill, loss of authorship, dependency — appear roughly 250 times; anxiety.learning_loss alone is the fourth-most-frequent code. Students' main concern seems to be something like "I am not yet calibrated."

    However: Section 4 of the protocol asked participants to pick between an over-use concern and an under-use concern, which structurally invited concern-talk; a different opening prompt might surface a different register.

  2. The integrity worry is less salient than the learning-atrophy worry. Anxiety about getting caught or plagiarism (anxiety.integrity, 77 utterances) lands well below anxiety about losing skill (anxiety.learning_loss, 130). The dominant fear is not "what will my professor think," it's "what is this doing to me?" [see also: The Little Mermaid code]

  3. The credential feels fragile. Many students seem to be experiencing the degree itself as vulnerable to being hollowed out by AI use while they are earning it, or to being devalued by a generation-level stigma they did not choose. The learning-atrophy worry and the credential-devaluation worry are fused in ways the codebook does not capture separately.

  4. What students want is judgment, not vague honor-code policy or a new course on how to use ChatGPT. They want a framework for making their own decisions, consistent guidance across classes, and real-world case studies of how people in their intended fields are actually using these tools. They want to be treated as people making serious bets about their own formation.

React with an emoji or leave a comment on any finding.

/ more voices

More from the focus groups.

Every substantive quote from the ten focus groups, plus the ten multi-voice exchanges, grouped by the section of the protocol they came from. Lives on its own page so it doesn't swallow this one.

Open the quote bank →

08 / say something back

What do you think?

Reactions, pushbacks, questions? Anonymous unless you say otherwise.

sign in with your Harvard email

We send a one-time magic link; your email is not shown to anyone.

one quick thing

So we can show you the readership mix without showing your name.