An exploratory study offers a map of how people make sense of artificial intelligence, organizing the discussion around three broad debates: how AI should be developed, what kind of thing an AI system is, and whether development should speed up or slow down. The authors treat these as frames — ways of structuring a complex subject — rather than as settled answers.
The framework is interpretive: it describes possible ways that frames help people understand and debate AI. It does not measure whether those frames change AI development.
A map built from large text collections
The research combined text analysis, interviews, literature review and theory development, moving between those elements as the analysis took shape. It began with empirical data rather than testing pre-existing hypotheses, making it exploratory and data-driven.
Its final text collections contained 371,312 newspaper articles, averaging 789 words, and 1,391,195 tweets from Verified accounts, averaging 24.0 words. The computational models used “discourse atoms” derived through k-means clustering and singular value decomposition of word embeddings — numerical representations of patterns in word use.
The models identified many smaller frames and grouped them into four broad areas: components, application contexts, system dynamics and societal issues. News placed relatively more emphasis on components, while Twitter placed relatively more emphasis on issues.
What AI practitioners described
The interview material came from 57 semi-structured interviews with AI practitioners: 30 conducted in 2021 and 27 in 2023. Thirteen participants were interviewed in both periods, while 14 joined after the researchers re-invited participants in 2023.
Recordings were transcribed and coded by the first author using open and axial coding, constant comparison and theoretical saturation. The analysis covered 49 hours of interviews and produced 841 first-order codes, or initial labels, and 213 higher-order codes used to organize them.
The analysis identified four challenges in making sense of AI: dealing with rapid change, communicating across groups, assigning responsibility for effects on society, and weighing trade-offs. These are qualitative, theory-building categories, not estimates of how widespread each challenge is.
Assigning responsibility was difficult in the accounts analyzed. The paper reports that participants most commonly attributed responsibility to data.
Participants also described trade-offs between fairness and predictive accuracy, and between personalization and privacy. The study presents these as interpretive accounts rather than quantified preferences.
What the study can and cannot establish
The study was exploratory and interpretive, did not aim for representative sampling, and was not suited to strong causal claims about AI discourse.
That matters because the proposed framework is not a tested causal model. The authors describe possible mechanisms through which frames might matter, but the study did not measure causal effects on AI development.
The result is therefore a framework for organizing debates around AI, not a measured account of effects on development. Whether the proposed frames or framing contests affect AI development remains unresolved within the study.
The document is an arXiv preprint dated 25 August 2026.
Paper data and sources
Original title: Method, Mind, and Morality: How People Make Sense of Artificial Intelligence
Authors: Jacy Reese Anthis, Erik Brynjolfsson, James Evans
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: 10.1145/3816971
Original paper · Full text