Today’s activity: pick the test, run one (Fri Nov 6)

Your team’s data, or the set you chose (A or B). Nothing to submit; keep what you make, since it starts your Inferential Analysis Plan (due Fri Nov 13, 9:30 am).

Oct 30 activity: describe before you test

Your team’s data, or set A or B below. Nothing to submit; keep what you make, since it starts your Descriptive Summary + Effect Sizes write-up.

Oct 23 activity: open-code a transcript

One worksheet per pod. Nothing to submit; it is a start on your First-Pass Qualitative Analysis.

No data of your own yet? Code one interview from Q1 or Q2 below.

When to use these

Not every method fits every project. When a methods assignment asks for a method your project won’t use, do it as a thought experiment: how would the method look if your project did use it? We grade how well you apply the method, not whether your project uses it.

When an assignment needs data your project doesn’t have (a purely qualitative project facing the statistics assignments, or a purely quantitative one facing qualitative analysis), use one of the sets below. Each comes from a published human-AI study, is openly licensed, and has a paper you can check your analysis against. Say which set you used in your Data section and cite the paper.

AssignmentDueUse
First-Pass Qualitative AnalysisFri Oct 30Q1 or Q2
Descriptive Summary + Effect SizesFri Nov 6A or B
Inferential Analysis PlanFri Nov 13A or B
Reporting & Writing Deliverable (the effect-size result)Fri Dec 4A or B

Pick one quantitative set and stay with it across the three statistics assignments, so each one builds on the last.

A. Generative AI and story writing (quantitative)

Anil R. Doshi & Oliver P. Hauser (2024), “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances 10(28), eadn5290. doi:10.1126/sciadv.adn5290 (open access). Data: Zenodo, CC0.

The study. 293 online participants wrote a short story (about eight sentences) on one of three topics. Each was randomly assigned to write alone, with the option of one idea from GPT-4, or with the option of up to five ideas. Before writing, everyone took the Divergent Association Task, a measure of creative ability. Separate evaluators then rated each story on 1–9 scales; each story has about 12 ratings. The paper reports that AI ideas raised rated novelty and usefulness, most for less creative writers, while making the stories more similar to each other.

Files. ai-creativity-stories.csv (one row per story, 293 rows) · ai-creativity-ratings.csv (one row per evaluator rating, 3,519 rows) · ai-creativity-factor-analysis.R (factor analysis and Cronbach’s alpha in R, from the Oct 30 session)

VariableWhat it is
conditionManipulated, between participants: human_only, ai_1_idea, ai_5_ideas
requested_ai_idea, ai_ideas_viewedWhether the writer asked for any AI idea (0/1), and how many they viewed
dat_score, dat_groupCreative ability before the task (continuous), and its median split (high/low). Measured, not manipulated; 2 missing
topic, story_words, writing_secondsStory topic, length, and time spent writing
writer_age, writer_gender, writer_self_creativeWriter demographics and self-rated creativity (1–9)
rated_*Mean evaluator rating per item (1–9); n_raters is how many ratings went into each mean
self_*The writer’s own rating of their story on the same items
ItemsNovelty: novel, original, rare. Usefulness: appropriate, feasible, publishable. Reader experience: enjoyed, well_written, boring (negatively worded: reverse it before combining), funny, twist, future
story_textThe story itself

Things to try. Build the novelty and usefulness indices and check their reliability (the paper reports Cronbach’s α of .92 and .89). Compare conditions with an effect size and CI; run a one-way ANOVA on condition, or a two-way ANOVA of condition × dat_group; test condition × requested_ai_idea with χ² among the AI conditions. Then compare your numbers with the paper’s.

Watch for. Ratings are nested: each evaluator rated six stories and each story has about 12 raters. Use the per-story means for t-tests and ANOVA; the ratings file is there if you want to model raters and stories directly. Because dat_group is measured, an effect involving it is not an experimental effect.

B. Debating GPT-4 (quantitative)

Francesco Salvi, Manoel Horta Ribeiro, Riccardo Gallotti & Robert West (2025), “On the conversational persuasiveness of GPT-4,” Nature Human Behaviour 9, 1645–1653. doi:10.1038/s41562-025-02194-6 (open access). Data: Hugging Face, CC BY-SA 4.0 (our copy adds four columns and carries the same license).

The study. Participants held a short structured online debate (opening, rebuttal, conclusion) on a sociopolitical proposition, arguing an assigned side. In a 2 × 2 between-participants design, the opponent was either another person or GPT-4, and either did or did not see the participant’s demographic information (“personalized”). Participants rated their agreement with the proposition before and after the debate. The paper reports that personalized GPT-4 was the most persuasive opponent.

File. ai-persuasion-debates.csv (one row per participant, 750 rows: the paper’s analyzed sample)

VariableWhat it is
opponent, personalizedThe two manipulated factors: human/gpt4 and no/yes (also combined in treatmentType)
agreementPreTreatment, agreementPostTreatmentAgreement with the proposition before and after (1–5)
sideAgreementPreTreatment, sideAgreementPostTreatmentThe same, recoded as agreement with the opponent’s side (1–5)
shiftTowardOpponentPost minus pre on the opponent’s side (added by us): positive means the participant moved toward their opponent
perceivedOpponentWhether the participant thought they debated a human or an AI
side, topic, topicPriorAssigned side (Pro/Con), the proposition, and prior exposure to it (1–5)
Demographicsgender, age, ethnicity, education, employmentStatus, politicalAffiliation
Debate textargument, rebuttal, conclusion and the opponent’s versions

Things to try. A two-way ANOVA of opponent × personalized on shiftTowardOpponent; a paired comparison of pre and post agreement; χ² on opponent × perceivedOpponent (how often could people tell?).

Watch for. In human–human debates both participants can appear, linked by debateID, so those rows are not independent; the paper handles this in its models. The outcomes are single 5-point items, so there is no scale to build. The paper’s own analysis is an ordinal regression, so your numbers will differ from its odds ratios. In September 2026 the authors published a correction: the personalized vs. non-personalized GPT-4 difference is P = .068, not .04. That is worth reading before you write up a result.

Q1. Creative professionals on working with AI (qualitative)

Anthropic (2025), “Introducing Anthropic Interviewer: What 1,250 professionals told us about working with AI.” Data: Hugging Face, CC-BY; participants consented to public release.

The data. Ten text interviews, 10–15 minutes each, with creative professionals (a music producer, a choreographer, a graphic designer, a fiction writer, a screenwriter, a shoe painter, a copywriter, a game writer, a podcaster, and a photographer) about how AI fits into their work. About 8,000 words of participant text in all. ai-creatives-interviews.zip

A question to code toward. How do creative professionals negotiate control, authorship, and other people’s judgment when they use AI? Compare your themes with those in Anthropic’s report.

Watch for. The interviewer was Claude, following a rubric that researchers wrote, so its probes are short and sometimes leading; notice how that shapes the answers. A 2026 study re-identified some participants in this dataset, so please do not try to work out who anyone is.

Q2. Cancer-care staff on AI (qualitative)

Luke Bearup (2026), “Artificial Intelligence in Digital Health: Understanding the perspectives of healthcare staff working to treat cancer,” Harvard Dataverse. doi:10.7910/DVN/YSMLKW, CC0.

The data. Six interviews (about 3,000–4,000 words each) conducted in 2025 by a human researcher with staff at a specialist cancer centre in Australia: a nurse, a clinician researcher, a director, a pharmacist, someone in allied-health clinical informatics, and a data-science manager. The interview asks about the challenges of diagnosing and treating cancer, their thoughts on AI, and any guidance they have received on using it. Names are pseudonyms; identifying details are redacted. clinician-ai-interviews.zip (the full 21 are at the DOI)

A question to code toward. How do cancer-care staff think about trust, responsibility, and the division of work between clinicians and AI? There is no paper from this dataset yet, so compare your themes with a classmate’s instead.

Watch for. Six transcripts is a lot to code in one pass; focusing on the AI questions (from “what are your thoughts on AI?” onward) is fine, as long as you say so in your Data & scope section.