Method, Mind, and Morality: How People Make Sense of Artificial Intelligence
Jacy Reese Anthis, Erik Brynjolfsson, James Evans
cs.CY, cs.AI, cs.CL, cs.LG, stat.ML
2026-08-25
Chicago and Stanford pair 371k news articles, 1.39M verified tweets, and 57 interviews to map how people make sense of AI along method, mind, and morality.
The word "AI" now covers recommender feeds, chatbots, driverless cars, and the autocomplete sitting in an editor. The same stack looks like a growth engine to investors, a mental-health risk to teachers, and a threat or a tool to engineers. CSCW already has local studies: how data scientists treat AutoML, how organizations hand decisions to algorithms. What has been missing is a map of how people stay cognitively afloat as digital minds enter everyday life.
Jacy Reese Anthis and James Evans at the University of Chicago, with Erik Brynjolfsson at Stanford, do not score models or referee the ethics debate. They ask how people assign social meaning to AI.
The design follows computational grounded theory: large-scale text first, then interviews that walk the resulting terrain. Timing is the ChatGPT hinge. Thirty interviews in 2021, twenty-seven in 2023, thirteen people in both waves. Documents run from 2018 through mid-2024.
News came from ProQuest English newspapers worldwide: 530,445 articles raw, 371,312 after dropping non-English, too-short, and OCR-garbled pieces, 789 words on average. Because a news story often spans several domains, the models used only the two sentences around a search term, 84 words on average. Social media came from Twitter's academic API: 1,391,195 tweets from verified accounts, January 2018 to 20 April 2023, 24 words on average. Search terms included artificial intelligence, chatbot, large language model, neural net, and a long appendix list.
Topic modeling skipped LDA. LDA treats a document as a bag of words and falls apart past ten to twenty topics. The authors used discourse atoms: k-means plus SVD dictionary learning over word embeddings, which stay coherent at 100 to 200 topics.
Recruitment went through LinkedIn attendees of AI events in July 2021, with no pay. After three people declined demographics, mean age was 41, 77% male, currently in the US (50%), India (14%), and the UK (9%), with a mean of nine years in AI and a median of five. Occupations clustered around managers (20%), data scientists (18%), researchers (16%), and software engineers (14%). Forty-nine hours of audio produced 841 first-order codes and 213 higher-order codes, open-coded and axially coded by one researcher, aiming for theoretical saturation rather than inter-coder reliability.
The topic models produced atomic frames, grouped by hand into four buckets: system components (chips, GPUs, GPT-4, OpenAI), application contexts (cancer, customer service, self-driving, surveillance), dynamics over time (acceleration, a fourth industrial revolution, job replacement), and social issues (bias, existential risk, EU law, the Hollywood strike). News leaned toward components; Twitter leaned toward issues. Atoms are building blocks. Someone can laminate GPUs onto "fourth industrial revolution" to talk about economic foundations, or onto "fear" to talk about loss of control or environmental cost.
Interviews pinned those blocks onto four cognitive jobs.
Keeping up. In 2023, policy professional Marco said it was hard to think "beyond the weekend." Manager Diana said a skill set "becomes antiquated overnight," that the new paradigm is "models feeding into models," and that it takes "an art to the science." Younger participants were more excited and less anxious.
Talking across groups. Meera told family she "sells data" for money. Deepak called himself a software engineer because that is what his uncles do. Salesperson Simon ran into a bank IT team that insisted on building a "God bot"; it could not handle "My laptop is broken."
Assigning responsibility. Some described systems as "bits on bits on bits," with no one owning the whole pipeline. Some said only developers can scrub bias. The most common move was to blame the data: a computer is "only as biased as the data that's inputted to it." An AI ethics committee at a large international organization walked away after it could not pin responsibility on either the human or the system.
Making trade-offs. The literature has fairness versus accuracy and personalization versus privacy. Most interviewees denied trade-offs or pushed the choice onto someone else. Marco, in an AI-policy role, said his job was to deliver information, not to hold anyone to a choice.
Above those four jobs, contestation collapsed onto three axes.
Method: top-down expert systems versus bottom-up scale. Deep learning has been dominant, but the argument continues. Critics used COVID as a counterexample: correlation-based models collapsed, and A/B tests plus causal modeling were crowded out by "predict, predict, predict, predict." Others treated further scaling as inevitable. Diana called prompt engineering and feature engineering "an art." Eoin said the field needs "a change in culture, and an investment in people."
Mind: tool versus coworker. Robot vacuums and chatbots do not sit at the same pole. Participants called systems assistants, early "proto-entities," invisible friends. Others held that coworker talk is a marketing ploy and the thing is still a tool. Anthropomorphism felt like a default. Mark, at an affective-AI startup, said a chatbot without empathy creates "pissed off customers." Some reached for slavery: Charles contrasted Star Trek mastery with WALL-E servitude; Raymond asked how well enslaving something smarter than us has gone historically. Ryan, a researcher at a leading lab, went the other way: a system with no subjectivity is a "cacophony of 10 million different personalities" with no overarching goal, so you cannot put yourself in its shoes.
Morality: slow down versus speed up. One side stacked bias, environmental cost, and concentrated corporate power, and asked regulators to catch up. Rebecca in 2023 said she was "very disappointed with the regulators." The other side treated ethics as already settled. People who stressed benefits often framed AI as an external force that could pull humans out of a downward spiral. Few mixed the two.
The discussion adds a warning. Nuanced positions get sheared into simple frames by the discursive opportunity structure. "Superintelligence is coming, so regulate" can be heard as "superintelligence is coming," which excites capital and speeds things up. The paper notes Sam Altman's remark that Eliezer Yudkowsky was critical to the decision to start OpenAI. Frames are also resources: "stochastic parrots" to deflate hype, "emergent abilities" to analogize a phase transition. Anthropomorphism looks hard to remove; whether the thing is an assistant or a proto-entity is still unsettled, and that is where contestation has room. The authors connect framing contests to the productivity paradox: electric motors paid off after factory floors were redesigned; AI may be waiting on a sociocognitive redesign.
Product, policy, and comms people get a map they can actually retell. A carefully layered risk message can land as "it is coming." Design choices that treat a system as a tool versus a coworker change whether users expect empathy and whether they ask if it counts as an entity. Slow-down versus speed-up is a moral frame, not a technical score. For researchers, Goffman's frames and Kaplan's framing contests get wired into AI discourse, which offers a social-cognition reading of why GPT looks strong while macro productivity has not moved.
This is a descriptive map, not a playbook. There is no causal identification, and no instruction on which pole to occupy.
The authors are clear: exploratory and interpretive, not representative, not causal. All text and interviews are in English. The sample leans US, leans toward English-speaking professionals who attend LinkedIn events, and is 77% male. Verified Twitter is not all social media. Word embeddings plus dictionary learning miss full semantics. Interviews were highly interactive, so the interviewer's stance leaked into answers. The world changed between 2021 and 2023; thirteen overlapping participants are not enough for a before-after statistic.
A few claims sit loosely. The three axes were distilled after the fact, not pre-registered; another team could cut the same material along different lines. Atomic-frame labels are researcher-invented; five nearest words do not prove the public uses that frame. In the responsibility and trade-off sections, "most common" is a reading of quotes, not a count. The productivity-paradox link is a proposed mechanism; the paper says it did not measure causal impact. "God bot" and a walked-away ethics committee make vivid stories. Generalizing them to the industry is a further step the evidence does not take.