Source-linked AI summary
Method, Mind, and Morality: How People Make Sense of Artificial Intelligence
Jacy Reese Anthis, Erik Brynjolfsson, James Evans
TL;DR
The paper asks how people can make sense of AI amid many competing social meanings. Using computational analysis and a framework of semantic frames, it identifies cognitive challenges and three debates that structure collective sensemaking, with implications for design, policy, and research.
Problem
Existing research examines particular social meanings of AI individually, but lacks a framework for how people can navigate the multitude of meanings surrounding AI.
Method
The study computationally constructs and analyzes topic models of AI-related newspaper articles and social media posts, interpreting frames as semantic building blocks that actors combine.
Results
The paper identifies four cognitive challenges and three framing debates concerning AI's method, mind, and morality.
Takeaways & Limitations
The framework offers implications for AI design, policymaking, and research by highlighting how framing dynamics shape collective cognition and social responses to AI.
Takeaways & Limitations
The exploratory, interpretive methodology is not suited to representative sampling or strong causal claims, and the predominantly English and U.S.-centric corpora limit applicability elsewhere.
Abstract
from arXiv · showhide
How can humans make sense of the rapid takeoff of artificial intelligence (AI)? We studied the sensemaking dynamics of AI through an open-ended, mixed-methods study with computational text analysis of millions of AI-related newspaper articles and social media posts grounded in 57 semi-structured interviews with AI professionals in 2021 and 2023--before and after the recent surge of public interest. We identify a range of sociological frames (interpretive schemas that structure collective cognition) and show how AI professionals use frames to address significant cognitive challenges, such as assigning responsibility for societal impacts. We develop a framework of three primary debates across which frames are adopted and contested: (i) the $\textit{method}$ of AI development, between frames of top-down expert systems and bottom-up emergent capabilities, (ii) the $\textit{mind}$ of an AI system, ranging from a passive tool to a humanlike "digital mind," and (iii) the $\textit{morality}$ of how AI is used, particularly the decision of whether to slow down or speed up AI development. As humanity enters the era of transformative AI, technologists and policymakers must account for the framing dynamics that will circumscribe our beliefs, values, and actions.
1 Introduction
AI’s rapid spread across social contexts creates a sensemaking challenge because its meanings vary across technologies, groups, and situations. The study addresses this challenge by combining large-scale discourse analysis with interviews to identify frames and the debates they organize.
- Motivation: AI carries different technical and social meanings across contexts, from algorithmic feeds and bots to threats, opportunities, and social actors.These meanings also vary between groups, such as technologists, investors, educators, and parents.
- Motivation: The study asks how people can remain cognitively oriented as AI’s meanings proliferate during its rapid societal expansion.
- Approach: The researchers combine computational analysis of newspaper and social-media discourse with qualitative analysis of 57 semi-structured interviews with AI professionals.Interviews were conducted in two waves: 30 in 2021 and 27 in 2023.
- Approach: Topic models map the field and generate interview topics, while interviews explain how individuals navigate and negotiate circulating frames.The analysis iterates between computational results, interview coding, theory development, and literature review.
- Contributions: The study identifies four cognitive challenges and three framing debates concerning AI’s method, mind, and morality.The challenges involve information overload, cross-group communication, responsibility attribution, and trade-offs; the debates range from top-down to bottom-up development, tool to humanlike mind, and slowing to speeding development.
- Contributions: The framework has implications for design, policymaking, and research, including the difficulty of promoting nuanced positions when people default to particular frames.The paper also shows how computational models of technological change can be enriched by participants’ lived experience.
2 Human-AI interaction
Human-AI interaction is shaped by users’ goals, characteristics, trust, and mental models across both salient and embedded AI systems. Prior research documents diverse benefits, risks, social meanings, and perceptions of AI mind and morality, but offers less account of collective cognitive management.
- Motivations: AI adoption is shaped by goals including productivity, while human-AI collaboration has shown improved task performance in science and creative writing.
- Motivations: AI’s economic benefits may be delayed, reflecting a possible productivity paradox analogous to delays documented for earlier general-purpose technologies.
- Outcomes: AI systems can support mental-health outcomes while also creating risks such as dependency and inappropriate content when users expect humanlike professional responses.
- Context: Task performance depends on trust, perceived intelligence and competence, transparency, explainability, and user characteristics including age, culture, gender, personality, and AI literacy.
- Context: Embedded AI in search and shopping can create feedback loops in which online behavior generates data used to train future AI systems.
- Mental models: Users readily attribute humanlike features, moral agency, responsibility, and moral standing to AI, with significant variation in perceived mind and moral status.
- Mental models: People use varied mental models of AI, including tools, servants, assistants, mediators, creators, curators, conversers, and co-authors.
- Research gap: Existing frameworks focus mainly on individual attitudes or AI behavior rather than how social groups manage the cognitive demands of AI technology.
3 Framing theory
Framing theory treats frames as shared schemas that organize meaning, action, and collective sensemaking across groups and institutions. Applied to AI, it highlights how frames arise, gain resonance, circulate through opportunity structures, and compete in contests over interpretation.
- Framing theory: Frames are schemas of interpretation that render meaning and organize experience, making framing useful for studying collective rather than merely private meaning-making.
- Framing theory: Frames connect communicators and receivers across institutional sites and circulate among social groups, including journalists, audiences, elites, and the public.
- Framing processes: Framing dynamics include alignment, discursive opportunity structures, and framing contests in which actors deliberately pit frames against one another.
- Framing processes: A frame’s reach and longevity depend on resonance with the interests and everyday experiences of particular individuals or groups.
- Frame types: Strategic frames are deliberately adopted by firms or industries to advance their interests, while technological frames describe shared interpretations of science and technology.
- Opportunity structures: In emerging fields such as AI, frames may draw on adjacent technologies or abstract master frames because preexisting discursive opportunity structures are limited.
- Application to AI: The paper uses framing theory descriptively to characterize contested AI frames and the dynamics through which actors advocate and resist them.
4 Methodology
The study used a computational grounded-theory pipeline that iterated among large-scale text analysis, interviews, literature review, and theory development. It combined newspaper and Verified Twitter discourse with 57 interviews conducted in 2021 and 2023, while acknowledging nonrepresentative corpora and participants.
- Research design: The research iterated between text analysis, interviews, literature review, and theory development rather than testing preexisting hypotheses.This followed a computational grounded-theory approach that began with empirical data.
- Research design: The study examined AI discourse from 2018 to mid-2024, focusing on changes between interviews conducted in mid-2021 and mid-2023.The interval included releases of AlphaFold 2, InstructGPT, DALL-E 2, ChatGPT, Gemini, and Claude.
- Corpora: The two corpora covered global English-language newspaper articles and Verified Twitter posts, selected to capture traditional and social-media discourse without contextual topic filters.The newspaper corpus contained 530,445 articles; Twitter coverage ended on April 20, 2023 because the academic API became unavailable later that year.
- Computational analysis: Topic models used word embeddings and discourse atoms rather than bag-of-words LDA, capturing semantic information through high-dimensional vectors and clustering-based decomposition.Discourse atoms were constructed through k-means clustering and singular value decomposition of word embeddings.
- Interviews: The interview sample comprised 57 AI professionals: 30 interviewed in 2021 and 27 in 2023, with 13 participating in both periods.Interviews were recruited through LinkedIn and structured around topic-model frames while allowing participants to introduce additional perspectives.
- Limitations: The study did not seek a representative sample; participants were English-speaking and likely more experienced and more engaged with the AI community than average practitioners.The corpora likewise represented nonrepresentative subsets of the human population.
5 Findings
The findings organize AI sensemaking around frames, the cognitive challenges they address, and three debates in which they are contested. The section therefore moves from identified frames to their social functions and central disputes.
- Findings: The findings are presented in three stages: identified AI frames, four challenges that motivate framing, and three primary debates where frames are contested.This structure links the descriptive frame inventory to interview-based sensemaking processes.
5.1 Frames
Discourse-atom topic models identified numerous AI frames, grouped into components, application contexts, system dynamics, and issues of AI and society. These atomic frames function as semantic building blocks that actors can combine for different interests.
- Frame categories: The frame inventory groups AI topics into four categories: sociotechnical components, application contexts, system dynamics, and AI-and-society issues.The categories organize a wide variety of atomic frames identified through discourse-atom topic models.
- Application contexts: Application frames span art, banking, cancer, companionship, COVID-19, criminal justice, customer service, cryptocurrency, cybersecurity, education, military uses, surgery, surveillance, and virtual assistants.The table also includes self-driving cars, sci-fi media, Tesla, vacuums, and wearable sensors.
- Dynamics and components: Examples include frames about acceleration, automation, change, efficiency, human input, humanlikeness, innovation, international agreements, labor costs, and robot interaction.The listed frames draw from both news and Twitter corpora.
- Issues of AI and society: The issues category includes frames concerning bias, dystopia, ethical guidelines, European law, existential risk, fear, hate speech, job replacement, misinformation, and political debate.Additional issue frames address reskilling, risks, social engineering, and campaigns concerning existential risk.
- Atomic frames: The displayed frames are atomic semantic building blocks rather than comprehensive social constructs, and actors can combine them to suit their interests.For example, GPUs can be combined with the 4th Industrial Revolution or Fear frames.
5.2 Challenges of understanding
AI professionals described four recurring challenges: keeping pace with change, communicating across groups, assigning responsibility for societal impacts, and weighing competing goals. Their accounts show both uncertainty and divergent attributions about AI’s social meaning.
- Keeping up with change: Rapid technical and social change challenged participants’ understanding, competency, and agency, with concern increasing after DALL-E 2, ChatGPT, Gemini, and other advanced systems.Senior professionals described expertise becoming outdated quickly and needing broader mental models of new AI systems.
- Communicating across groups: Participants struggled to explain AI work across families, organizations, clients, investors, trainees, and professional groups with different levels of exposure to AI.Some used familiar occupations, software engineering, or laptop use to make their work understandable.
- Responsibility for societal impacts: Responsibility was difficult to assign because layered AI systems diffused accountability and understanding across the development process.Participants variously attributed responsibility to data, developers, or other actors; some developers argued peers should accept responsibility rather than deflecting it onto data.
- Responsibility for societal impacts: Participants’ responsibility disputes also shaped ethics deliberation, including a committee stalemate over whether AI could be a moral agent or whether responsibility belonged to the person giving orders.The account identifies refusal to attribute responsibility as blocking agreement on basic standards.
- Trade-offs: Participants reported trade-offs between fairness and predictive accuracy, personalized data and surveillance, and technology advancement and preserving social values.Some denied trade-offs or emphasized synergies, while others framed ethical development as compatible with monetization.
5.3 Framing debates
Three framing debates organize how AI professionals make sense of AI: its development method, the system’s mind, and the morality of advancing or slowing development. Across these debates, participants contested how AI should be understood, guided, and governed.
- Three debates structure AI framing: method contrasts top-down expert systems with bottom-up emergence, mind ranges from tool to humanlike system, and morality weighs caution against acceleration.
- Method: Top-down versus bottom-up: Bottom-up development frames emphasize capabilities emerging through scaling, while participants also described human guidance through process design, prompting, and feature engineering.
- Mind: Tool versus coworker: Participants framed AI’s mind between an assistant or tool and a humanlike coworker, with humanlikeness viewed as useful for education, interaction, and emotional understanding.
- Mind: Tool versus coworker: Human-AI relationships also raised questions about agency, dependence, and whether advanced systems should be treated as tools, coworkers, or entities whose subordination is morally troubling.
- Morality: Slow down versus speed up: Moral frames divided between foregrounding AI’s risks and problems, favoring caution or slowdown, and foregrounding innovation’s benefits, favoring continued development or acceleration.
- Morality: Slow down versus speed up: Participants connected AI morality to harms including bias, environmental consumption, concentrated power, and the regulation of harmful applications.
6 Discussion
The discussion interprets framing as a way people simplify AI’s complexity while warning that simplification can erase nuance and shape technological trajectories. It highlights implications for communication, research, group tensions, and economic change, while acknowledging that causal effects were not measured.
- The study does not empirically measure whether frames causally affect AI development trajectories; its implications are interpretive mechanisms grounded in observed empirical patterns.
- People use frames to make AI’s diverse applications and contexts socially and cognitively tractable, but simplifying complex ideas can discard information.
- The study’s mind axis primarily runs from tool to coworker, potentially overlooking distinct dimensions of user and machine autonomy identified in prior frameworks.
- Nuanced messages may be sheared as audiences broaden, and policy warnings about slowing AI can be simplified into messages that instead excite acceleration.
- Strategic frames can quickly communicate multifaceted critiques, including concerns about bias, discrimination, colonialism, environmentalism, and animal rights.
- Frames around emergent abilities draw on familiar scientific concepts such as phase transitions and emergence, helping make bottom-up AI development resonant.
- AI’s humanlike framing remains unsettled, with systems described as assistants, proto-entities, invisible friends, or role-play systems that avoid anthropomorphism.
- Frames can emerge simultaneously across groups, come into tension during exchanges, or be modified within groups before re-entering broader discourse.
7 Limitations
The study is exploratory and interpretive, with important limits on causal inference, representativeness, language coverage, regional applicability, and the scope of its data sources.
- The exploratory, interpretive methodology was designed for theoretical saturation rather than representative sampling or quantitative hypothesis testing.
- The methodology cannot support strong causal claims about factors shaping AI discourse or its general consequences.
- English-language corpora and interviews, predominantly U.S.-centric materials and participants, limit applicability to non-English discourse and other regions.
- The sampling excludes or underrepresents less mainstream media, less-online professionals, and populations beyond WEIRD contexts.
- The evidence is temporally bounded by documents from 2018–2024 and interviews conducted in 2021 and 2023.
- Word embeddings and dictionary learning do not capture full semantic meaning, while interviewer perspectives affected the highly interactive interviews.
8 Conclusion
The study identifies cognitive challenges that drive AI framing and organizes contested social meanings around the method, mind, and morality of AI. This framework connects fast-changing discourse with debates over development, AI agency, and whether to accelerate or slow technological progress.
- People use AI frames to address fast information, cross-group communication, responsibility for societal impacts, and trade-offs among priorities.
- AI’s social meaning is debated through three dimensions: development method, the system’s mind, and the morality of accelerating or slowing technology.
- Method: The method debate ranges from top-down AI development to bottom-up emergence through large-scale computational resources.
- Mind: The mind debate ranges from treating AI as a tool to treating it as an agent.
- Morality: The morality debate ranges from rapidly speeding up AI to aggressively slowing it down.
A Appendix
The appendix documents the study’s term-selection conventions, participant characteristics, and interview timing, alongside a temporal visualization of selected Twitter frames.
- Table A1: Table A1 marks wildcard characters and includes terms beginning with hashtags or containing no spaces, such as “#artificialintelligence.”
- Table A2: Table A2 reports participant characteristics using pseudonyms, with some demographic details generalized to preserve anonymity.
- Table A2: The participant table indicates interview participation in 2021 and 2023 and notes that many managers and entrepreneurs had experience in other listed occupations.
- Table A2: The listed participants span occupations including designers, diplomats, entrepreneurs, lawyers, managers, and software engineers across multiple countries.
- Publication record: The study was received in May 2025, revised in January 2026, and accepted in March 2026.
- Figure A1: Figure A1 tracks relative frequencies over time for up to 50 associated words across 36 selected Twitter frames.