How Knowledge Workers Think Generative AI Will (Not) Transform Their Industries

1 INTRODUCTION

• Public narratives predict generative AI could automate much knowledge work and harm labor, while evidence of task-level productivity gains leaves its future impact uncertain.The introduction frames worker expectations as a timely topic amid these competing accounts.Source: §1 · p. 1
• The study uses participatory workshops with 54 knowledge workers across seven industries and three US cities to explore anticipated effects.The industries include advertising, business communications, education, journalism, law, mental health, and software development.Source: §1 · p. 1
• The inquiry covers not only task and job changes but also whether generative AI may shift the nature and structure of work.The authors also identify five HCI research challenges at the intersection of generative AI and knowledge work.Source: §1 · p. 2
• Across industries, participants commonly envision generative AI doing menial work under human review and potentially amplifying deskilling, dehumanization, disconnection, and disinformation.The authors also report industry-specific perspectives.Source: §1 · p. 2

2 BACKGROUND

• The background positions the study at the intersection of generative AI technologies, knowledge work, prior research, and public perceptions.It also reviews narratives about generative AI’s societal and workplace effects.Source: §2 · p. 2

2.1 Generative AI Models

• Generative AI systems generate human-like content from input; transformer models and foundation models underpin many current text, image, and code-generation systems.The systems are described as generalized, capable of producing original content, and accessible through intuitive interfaces.Source: §2.1 · p. 2

2.2 Knowledge Work

• Knowledge work applies specialized theoretical and analytical knowledge to non-routine problem solving and the development of products and services.Knowledge workers may possess expertise without necessarily completing a formal education program, though some professions require credentials.Source: §2.2 · p. 2

2.3 Generative AI and the Future of Knowledge Work

• Prior analyses identify knowledge-work activities as vulnerable to automation; one estimates approximately two-thirds of US occupations may experience changes in how work is done.Research also reports task-specific productivity gains, including larger gains for novice customer-service agents.Source: §2.3 · pp. 2–3

2.3.1 Generative AI Impacts on Worker Tasks.

• HCI studies have examined writing and professional tasks, but research on impacts across professional industries remains limited, especially outside creative contexts.Prior work includes studies of creative and educational settings, as well as potential efficiency effects for knowledge workers.Source: §2.3.1 · p. 3

2.4 Sensationalized Generative AI Narratives

• Sensationalized narratives forecast sweeping disruption and labor displacement, yet often omit knowledge workers’ perspectives and can narrow public debate.Calls for more measured accounts emphasize how generative AI may interact with specific kinds of professional work.Source: §2.4 · p. 3

2.5 Public Perception of AI

• Public perceptions of AI commonly combine expectations of future impact with mixed emotions and concerns including job loss and social isolation.A Pew survey found that about one in five of the 37% of US respondents more concerned than excited about AI cited job loss as their main concern.Source: §2.5 · p. 3

3 METHODOLOGY

• The research asks how knowledge workers expect generative AI to affect their industries and how they relate it to other anticipated industry changes.The study investigates these questions through participatory workshops with 54 participants from seven industries in three US cities.Source: §3 · pp. 3–4

3.1 Participants

• Each industry group included 7–8 participants, combining central and specialized professional roles to elicit varied perspectives.Recruitment also sought a balanced mix of genders and half non-White participants in each workshop.Source: §3.1 · p. 4

3.2 Workshops

• The participatory approach centers lived experience, attends to social change, and seeks to mitigate research power asymmetries.Workshops used probes and provocations to support collaborative reflection and agenda-setting.Source: §3.2 · p. 4
• Three-hour, in-person workshops combined a shared generative AI overview with change-card discussions, policy activities, and voting on helpful and detrimental aspects.Workshops were held in July 2023 across Columbus, New York City, and Oakland.Source: §3.2 · p. 4

3.3 Analysis

• Researchers analyzed corrected transcripts and workshop artifacts through reflexive thematic analysis, using independent open coding, memo writing, and collaborative theme development.Four authors reviewed transcripts in paired and independent configurations and compared them with corresponding artifacts.Source: §3.3 · p. 5

3.4 Limitations

• The small, nonrepresentative sample—one group per industry in three US cities—supports exploring participants’ perspectives, not generalizing to entire industries.The authors also note that teaching content and activities may have influenced participants, despite efforts to minimize those effects.Source: §3.4 · p. 6

4 FINDINGS

• The findings cover a cross-group narrative about generative AI, deskilling, dehumanization, disconnection, disinformation, and industry-specific perspectives.The section proceeds from participants’ expectations to the four forces and then to perspectives in particular knowledge industries.Source: §4 · p. 6

4.1.1 Generative AI can automate menial work: “We need to just

• Participants favored generative AI for routine, formulaic tasks to free time for meaningful human work, but expected workers to review its output.They also expected human review to ensure output quality, sometimes by a qualified professional.Source: §4.1.1 · pp. 6–7

4.1.3 Existing industry governance structures are amenable to overseeing generative AI: “It would just be part of what we’ve already

• Participants expected oversight of AI-generated work to fit existing review structures, including legal and compliance approvals and software code review.One business communications participant trusted AI-generated messages if they passed existing brand-voice and consistency checks.Source: §4.1.3 · pp. 7–8

4.1.4 Certain jobs will be downsized, but do not expect broad disruption: “A lot of entry level positions will be eliminated,” but “I don’t

• Participants expected some downsizing in entry-level or rote-heavy jobs, but generally viewed industry-wide disruption claims skeptically and expected limited personal impact.Some acknowledged larger potential shifts, particularly in software development, advertising, or business, but this was not the prevailing view.Source: §4.1.4 · p. 8

4.1.5 Need to evolve and adapt with generative AI: “Let’s evolve

• Participants urged workers to adapt their skills to use generative AI, including learning prompt engineering, while preserving independent thinking and learning.Some saw adaptation as a way to stay relevant and avoid being left behind.Source: §4.1.5 · p. 9

4.2 Social Forces: Deskilling, Dehumanization, Disconnection, and Disinformation

• Participants expected generative AI to amplify four existing forces shaping knowledge work: deskilling, dehumanization, disconnection, and disinformation.They framed these as social forces that generative AI may intensify, alongside changes not driven by AI.Source: §4.2 · p. 9

4.2.1 Deskilling: The “Uberfication” of Knowledge Work.

• Participants feared AI could deepen precarious work, enable cheaper services and client substitution, devalue expertise, and eliminate entry-level pathways into senior roles.They described contract and freelance work as vulnerable to replacement, while AI-enabled services could reduce demand for trained professionals and provide poorer outcomes for clients.Source: §4.2.1 · pp. 9–10

4.2 Social Forces: Deskilling, Dehumanization, Disconnection, and Disinformation

• Participants feared AI could make work more impersonal by displacing human interaction, rapport, creative fulfillment, and worker autonomy; some proposed human-work quotas.Mental health professionals emphasized that therapy requires human rapport, while other participants valued the satisfaction of making things themselves.Source: §4.2.2 · p. 10

4.2.3 Disconnection: “You have to be there.”

• Participants worried AI could worsen disconnection by substituting chatbots for human support and further separating journalism from on-the-ground reporting and reality.They also feared journalism and other content would recycle increasingly nonhuman material, distancing audiences from firsthand experience.Source: §4.2.3 · pp. 11–12

4.2.4 Disinformation: “I think they call them hallucinations... it puts

• Participants feared AI could amplify low-quality information, accelerate harmful content production, make truth harder to discern, and compound media concentration and homogenization.They noted that AI trained on existing internet material could inherit its misinformation, while media upheaval and aggregation already threatened information quality.Source: §4.2.4 · p. 12
• J2 feared generative AI could worsen media concentration by making it cheaper for owners to replace newsrooms and advance their political agendas.The participant imagined owners buying media, firing newsroom staff, and running the publication with generative AI.Source: §4.2.4 · p. 12

4.2.5 Industry Perspectives.

• Industry composites illustrate how themes vary by industry; Table 4 orders industries by expected impact and lists each industry’s three most salient themes.The authors describe the composites as capturing the content, language, and tone of responses.Source: §4.2.5 · p. 12, Table 4 · p. 13
• Table 4 is not comprehensive: deskilling was discussed extensively in Law but was not among its top three themes.Source: Table 4 · p. 13

5.1 Research Challenges: Human-in-the-Loop

• Participants viewed human review as sufficient, but review errors, poor handoffs, and unequal accountability motivate research into HITL’s limits and better systems.Potential improvements include review approaches that scale and interventions or training that strengthen analytical review; they are not expected to resolve every concern.Source: §5.1 · pp. 13–14

5.2 Research Challenges: Knowledge Worker Expectations of Impact

• The gap between workers’ limited expectations of impact and broader disruption narratives raises questions about why some feel insulated and how to support informed input.The authors suggest exploring industry-specific demonstrations and transparency artifacts as possible ways to help workers assess potential impacts.Source: §5.2 · p. 14

5.3 Research Challenges: Deskilling

• Potential skill shifts raise questions about reskilling, worker protections, and how to value labor that helps automated systems perform work.The authors point to organization-based reskilling, critical thinking, human review, and higher-level skills; they also note that human labor can help make automation viable or train systems.Source: §5.3 · p. 14

5.4 Research Challenges: Dehumanization

• The authors ask how to protect meaningful work that AI can perform efficiently, since automation could shift devalued labor or deepen inequalities rather than liberate workers.Possible responses include systems that reserve some decisions for people and broader regulation; the discussion also raises concerns about diminishing the value of human contributions.Source: §5.4 · p. 15
• Participants’ concern that AI may reduce reasoning tasks motivates research on collaborations and systems that engage people in problem solving rather than only providing answers.The discussion connects this concern to the possibility of a “lazy society.”Source: §5.4 · p. 15

5.5 Research Challenges: Guardrails

• Responsible AI approaches should assess impacts in their social contexts, since model-output benchmarks alone are insufficient to understand harm.The discussion notes that current responsibility metrics address disinformation more readily than deskilling, dehumanization, or disconnection.Source: §5.5 · p. 15
• Responsible AI assessments should incorporate community and worker perspectives, not focus only on product harms and developer-facing interventions.The authors propose engaging communities of practice and ask how workers affected by job changes can provide input before AI solutions are developed.Source: §5.5 · pp. 15–16

• Participants positioned generative AI as extending and exacerbating existing social forces rather than as an independent disruptor of their industries.The four forces identified were deskilling, dehumanization, disconnection, and disinformation.Source: §6 · p. 16
• Participants situated generative AI’s increasing availability and visibility amid a global pandemic and economic uncertainty.Source: §6 · p. 16
• This framing raises questions and opportunities for exploring and shaping generative AI’s future impact on knowledge workers and their industries.Source: §6 · p. 16

A INDUSTRY DETAILS

• Table 5 provides a brief overview of industries, recruitment, and roles.Source: Table 5 · p. 22, A INDUSTRY DETAILS · p. 22
• The industry framing distinguishes roles considered central from secondary roles.Source: Table 5 · p. 22
• The role descriptions were shared with recruitment partners.Source: Table 5 · p. 22

B INTRODUCTION TO GENERATIVE AI

• An approximately 40-minute education session introduced generative AI, with a 20-minute presentation followed by 20 minutes for questions and discussion.One of two author-researchers led the session, and materials were lightly customized to each industry.Source: B INTRODUCTION TO GENERATIVE AI · p. 23
• The shared definition described artificial intelligence as a computer or machine’s ability to think or learn.Source: B INTRODUCTION TO GENERATIVE AI · p. 23
• The 15 concepts included bias, generic outputs, brainstorming, emergent properties, falsehoods, and grammatical writing.Source: B INTRODUCTION TO GENERATIVE AI · p. 23
• Additional concepts covered polished syntax, tacit structure, privacy breaches, mimicry, variable responses, and unclear provenance.Source: B INTRODUCTION TO GENERATIVE AI · p. 23
• The concepts also addressed remixing, fallible safety systems, scale and speed, and prompt-based influence over outputs.Source: B INTRODUCTION TO GENERATIVE AI · p. 23

C ARTIFACTS

• Participants were invited to draw a map of their industry or field during check-in.Source: Figure 1 · p. 24
• Participants completed change cards individually and then shared them in facilitated discussion; this example concerns AI image generation.Source: Figure 2 · p. 25
• Participants completed policy handouts individually and discussed them; the example suggests constraints on newsroom AI use.Source: Figure 3 · p. 26