← Back
Prisma for Academic Research

AI conversation experiments — control, observation, analysis

Design the protocol, assign participant codes, observe live, and export clean data — without building the infrastructure.

Attitude change study — Participant P-047
AI
When you think about AI in hiring decisions over the next decade, would you say you're mostly optimistic or skeptical?
P
Skeptical, mostly. The efficiency gains are real, but I think the bias risks are underestimated — especially for candidates from non-traditional backgrounds.
AI
That's a well-grounded concern. What would need to change — technically or legally — for your view to shift?

1 · Design a simulation

A protocol that holds its character, every time

Validity starts with the agent staying in character. Prisma lets you write theoretical grounding and persona constraints directly into the agent's instructions, so it doesn't drift, improvise, or break the manipulation partway through a session — the same protocol runs identically for participant one and participant three hundred.

Example

A study on persuasive messaging defines its agent with the exact rhetorical strategy from the theoretical model being tested — say, a fear-appeal condition grounded in protection motivation theory — with explicit instructions never to soften the appeal or slip into a neutral tone, even if the participant pushes back.

2 · Control groups

Every condition, held apart cleanly

Participant codes are what keep conditions clean. Each code is tied to a specific session and configuration, so you decide who gets which agent, which persona, which feedback rule — or no feedback at all — without touching a single line of anyone's data.

Example

A between-subjects design assigns half the participant codes to a treatment session where the agent delivers personalised feedback at the end, and the other half to a control session where it doesn't — both run at the same time, under the same study, with no risk of cross-contamination.

3 · Analyze

Coded variables, timing, and cross-model checks

Prisma codes your variables directly from transcripts — categorical, numeric, or free-text — following the same academic conventions you'd use for manual coding, and records how long each conversation took down to the turn. To support reliability, you can run the same transcripts through more than one model and compare the coded output before committing to a final analysis.

Example

A study measuring attitude change codes each transcript for a pre/post sentiment score and a count of counter-arguments raised, then re-runs the coding pass with a second model to check inter-model agreement — the same logic as a second human coder, at a fraction of the time.

Try Prisma for free

One simulation, no credit card, no setup. See for yourself before choosing a plan.

As a teacher Build classroom role-plays, interviews, and case simulations for your students. Try for free
As a researcher Design conversation experiments with participant codes and live observation. Try for free