Productivity implications for generative AI role-based prompts as a networked hermeneutic (Commentary)

Research brief

By Sean Rintel.

Takeaway: Prompts that ask AI to play a role draw on everyday, shared ways of sorting people into roles, and AI replies reflect that shared understanding back. Reading replies this way could help explain AI's mistakes, support simulations of role-based conduct, and help teams understand each other's stances.

Why this matters

What was known

Housley and Dahl (2024) showed that role-based prompts for generative AI draw on everyday ways of categorising people and describing actions. (p. 1)

What this paper adds

A Human-Computer Interaction researcher's view of what that finding means for designing generative AI for productivity, with three implications for explainability, research and team collaboration. (p. 1)

Who it matters to

Designers of generative AI systems for productivity, and the people who must craft prompts and judge what comes back. (p. 1)

Research question

• What does understanding role-based prompts as a networked hermeneutic of lay and professional sociology imply for designing generative AI systems for productivity?

Method and data

Design

Invited commentary on Housley and Dahl (2024), 'Membership categorisation, sociological description and role prompt engineering with ChatGPT'.

Setting

Discourse & Communication, 2024.

Participants / data

No empirical sample. An argument from the perspective of an industry Human-Computer Interaction researcher.

Analysis

Conceptual argument drawing on research on generative AI metacognition, co-auditing and technomethodology.

Findings

• AI reflects our own understanding back. I argue that generative AI's responses, like users' role-based prompts, reflect a shared everyday understanding of roles and actions. Treating responses as enacting that understanding, and knowing the training data, may help explain AI behaviour, from simple mistakes to doing the opposite of what was asked. #

  “Just as users' role-based prompting is a networked hermeneutic, so too are GenAI's reflections of ourselves.”

• Explainability supports metacognition. Explaining AI behaviour in this way matters because using generative AI makes metacognitive demands on people, and because people need to double-check its output (co-auditing). #

  “calibrating one's ability to craft prompts and evaluate outputs”

• Simulated roles could scale research. Because generative AI reflects shared understandings of roles, it could simulate how people in a given role behave, and those simulations could be compared with real people's responses, opening new lines of 'technomethodological' inquiry. #

  “there is potential for scaling research into human action that did not exist before.”

• Personas are not people. Simulated behaviour cannot simply stand in for real people as evidence of how people in those roles act. #

  “Such inquiries should not naïvely substitute synthetic personas for real people and assume that the results are either relevant or generalizable.”

• Coaching for team collaboration. Understanding role-based prompts could help teams work together, beginning with AI systems that coach users to state the interactional stances that will get them relevant work information. #

  (p. 2)

Questions this paper answers

Why does generative AI sometimes do the opposite of what it is asked?

One possible explanation is that its responses reflect the everyday understanding of roles and actions that role-based prompts draw on (a networked hermeneutic). Treating responses as enacting that understanding, and knowing the training data, may help explain such behaviour. (See p. 1.)

Can AI personas replace human participants in research?

Not as a simple substitute. Generative AI could simulate role-based behaviour to compare with responses from people in those roles, which could scale research into human action. But synthetic personas should not naively substitute for real people or be assumed relevant or generalisable. (See p. 2; p. 2.)

Why does explaining AI behaviour matter to the people using it?

Because using generative AI is metacognitively demanding. People must judge how well they can craft prompts and evaluate outputs, and double-check what comes back. Understanding responses as reflections of shared knowledge of roles could support that judgement beyond mechanised methods of explanation. (See p. 1; p. 1.)

Contributions and implications

• A more complex version would help unpack the stances of absent team members, and synthesise a team's stances to support alignment, onboarding and mutual understanding. (p. 2)
• Explainability could come from understanding the networked hermeneutic, not only from mechanised methods. (p. 1)

Limitations

• This is a short commentary, so the three implications are proposals, not tested results.

How to cite

Sean Rintel. 2024. Productivity implications for generative AI role-based prompts as a networked hermeneutic (Commentary). Discourse & Communication. https://doi.org/10.1177/17504813241271444

This brief was prepared with AI assistance and published by Sean Rintel. It is not the authors' published text. Page numbers refer to this PDF.