These Are the Jobs People Actually WANT AI to Automate
Study Document: These Are the Jobs People Actually Want AI to Automate
Overview
This episode of the AI Daily Brief (dated 2025-10-13) examines evolving public and worker sentiment toward AI-driven job automation, drawing on two academic studies and real-world adoption data. The host, Nathaniel Whittemore, synthesizes findings from a Harvard Business School study on moral acceptability of AI automation, a Stanford study on worker automation preferences, a U.S. Senate report projecting 100 million displaced jobs, and a survey of blue-collar trades professionals. The central argument is that the discourse around AI and employment has matured beyond binary fears and now allows for nuanced, task-level analysis of where AI is genuinely wanted, where it is resisted, and why those two things sometimes diverge.
Source video: (URL not provided)
Prerequisites
- Basic familiarity with generative AI tools (e.g., ChatGPT, Gemini)
- General understanding of labor economics and job displacement concepts
- Awareness of ongoing U.S.–China technology trade tensions
- Familiarity with the concept of AI capability benchmarking and token-based usage metrics
- Some background on AI policy debates and the concept of a “social contract”
Main Points
Google’s Token Processing Reaches 1.3 Quadrillion Per Month
- Google is now processing 1.3 quadrillion tokens monthly across its AI products, up from 480 trillion in May and 980 trillion in July of the same year — roughly 104% growth in two months.
- DeepMind CEO Demis Hassabis reframed the scale: 500 million tokens per second, or 1.8 trillion per hour.
- Growth is attributed in part to expanded AI coding use cases, which are token-intensive.
- Gemini led all AI platforms in web traffic growth in September (46% jump), more than tripling the growth of second-place Perplexity (14%); Grok was the only platform to see a decline (−7.4%).
Meta Recruits High-Profile Researcher from Thinking Machines Lab
- Andrew Tulloch, a founding member of Thinking Machines Lab (TML) and former OpenAI researcher, left to join Meta’s superintelligence lab.
- Meta had previously been reported to have offered Tulloch a six-year, $1.5 billion package in August; current rumors place the accepted offer at $3.5 billion, though this is unconfirmed.
- An alternative explanation circulating is that the compute and infrastructure gap between TML and Meta may have influenced the decision.
- The move intensifies scrutiny on Meta’s AI talent acquisition strategy and raises expectations for Llama 5.
XAI Enters the World Model Space
- XAI hired two researchers from NVIDIA’s Omniverse platform team (a leader in simulation-based world models) to develop world modeling capabilities.
- Elon Musk announced plans for an AI-generated video game from XAI Game Studio before year-end, leveraging world model technology.
- One analyst framed this as a “classic Elon strategy”: world models are necessary for long-term robotics (Optimus), but games provide near-term revenue for the same underlying technology.
- Speculation exists that this could be a medium-term narrative to support a potential Tesla acquisition of XAI.
Chip War Escalation: China Cracks Down on NVIDIA Imports; Netherlands Seizes Chinese Chipmaker
- Chinese customs officials began physically searching ports for NVIDIA H20 and RTX Pro 6000D chips and reviewing import documentation for false declarations.
- Beijing’s crackdown on advanced chip imports now appears stricter than U.S. export enforcement in some respects.
- The Dutch government invoked a 1952 law (Goods Availability Act) — for the first time ever — to seize control of Nexperia, a Dutch subsidiary of China’s Wingtech Technology, citing national security and governance concerns.
- Analysts characterize the Nexperia seizure as a “frontline moment” in the broader Western effort to prevent Chinese access to advanced semiconductor know-how embedded in Western-based facilities.
Harvard Business School Study: Public Moral Acceptability of AI Automation
- Researchers asked the American public (not just workers) how they feel about AI replacing humans across various occupations — a departure from capability-focused studies.
- Key finding: Americans support automating 30% of occupations given current AI capabilities; that figure rises to 58% when AI is described as outperforming humans at lower cost.
- A narrow subset (~12% of occupations) — including caregiving, therapy, and spiritual leadership — is considered “morally repugnant” to automate regardless of capability.
- The study produced a four-quadrant framework:
| Quadrant | Capability | Moral Repugnance | Label | Example Occupations |
|---|---|---|---|---|
| Green | High | Low | No Friction | Financial analysts, economists, SFX artists |
| Yellow | High | High | Moral Friction | OBGYNs, history teachers, school psychologists |
| Red | Low | High | Dual Friction | Oral surgeons, nannies, nuclear technicians |
| Blue | Low | Low | Technical Friction | Cashiers, semiconductor techs, mail sorters |
- Core conclusion: For most occupations, resistance to AI is rooted in performance concerns, not principled moral objections — meaning resistance will erode as capabilities improve.
Stanford Study Comparison: What Workers Themselves Want Automated
- An earlier Stanford Digital Economy Lab study (led by Erik Brynjolfsson) asked workers — not the public — about their automation preferences, producing a similar four-quadrant map.
- Key distinction: workers consistently show a higher threshold for what they want automated in their own jobs compared to what the broader public thinks is acceptable to automate.
- When the two studies are mapped together (worker automation desire vs. public moral acceptability), four combined zones emerge:
| Zone | Worker Desire | Public Acceptability | Label | Examples |
|---|---|---|---|---|
| Green | High | High | Full green light | Scheduling, payroll fixes, database maintenance |
| Yellow | Low | High | Augment carefully / co-pilot | Film editing, graphic layout, story assignment |
| Orange | High | Low | Assistive only | Care and therapy intake summaries |
| Red | Low | Low | Defer/govern | Hiring/firing decisions, parole risk calls, ethics reviews |
- The “augment carefully” quadrant reveals a craft knowledge gap: workers in fields like film editing understand nuance that the general public does not perceive.
- The “assistive only” quadrant (e.g., caregiving) highlights a tension: the public sees automation as morally repugnant, but frontline workers recognize significant administrative automation value that could reduce burnout and neglect.
Senate Report: 100 Million U.S. Jobs at Risk
- A report by Democrat staffers on the Senate HELP Committee projected that AI and automation could displace nearly 100 million U.S. jobs over the next decade — over half the current 170-million-person workforce.
- Methodology relied heavily on querying ChatGPT directly, which the staffers themselves acknowledged as limited.
- Projected impacts: 89% of fast food workers, 64% of accountants, 47% of truck drivers displaced.
- Policy recommendations included: a 32-hour workweek, $17 minimum wage, stronger worker protections, and eliminating tax breaks for companies that automate.
- The host notes the report’s purpose was not methodological precision but to provoke a policy conversation about a new social contract, which he views as a legitimate and necessary goal even if the numbers are uncertain.
Blue-Collar Trades: Real-World AI Adoption
- A Housecall Pro survey of 400 home service professionals found 40% actively use AI and 60% use it at least somewhat, primarily for content creation and administrative tasks.
- Reported time savings averaged 3.2 hours per week (~160 hours/year) — equivalent to four full weeks of administrative work annually, highly impactful for owner-operated small businesses.
- 73% of tradespeople surveyed said AI had not affected their hiring rates — augmentation, not replacement.
- Real-world examples:
- Oak Creek Plumbing (Milwaukee): 20 plumbers using ChatGPT for field troubleshooting and diagnostics.
- Gulf Shore Air Conditioning (Niceville, FL): Fully AI-powered booking system; technicians use AI to diagnose issues and retrieve technical manuals in seconds; AI-optimized marketing drove significant revenue growth.
- Trades are well-suited to AI augmentation because they combine vast technical knowledge libraries (where AI excels) with hands-on physical skill (where AI cannot substitute).
Key Concepts
- Token processing volume: A measure of AI infrastructure usage; Google’s 1.3 quadrillion monthly tokens indicates scale of real-world AI deployment.
- No Friction quadrant: Harvard study category where AI capability is high and public moral repugnance is low — occupations most likely to see smooth automation.
- Moral friction: Occupations where AI is technically capable but the public perceives automation as morally problematic (e.g., roles involving human judgment, education, or healthcare).
- Technical friction: Occupations where the public has no moral objection to automation but AI capability is not yet sufficient — a strategic opportunity zone for AI developers.
- Dual friction: Occupations with both low AI capability and high moral repugnance — lowest-priority automation targets.
- Task-level analysis: Evaluating AI’s impact at the level of individual job tasks rather than whole roles, enabling more precise assessment of displacement and augmentation.
- Social contract (AI context): The renegotiation of societal expectations about work, contribution, and compensation in an economy where AI can perform a significant share of labor.
- World models: AI systems that learn a physical or simulated representation of environments, used to train embodied AI (robots) and potentially to generate interactive games.
- Goods Availability Act (Netherlands, 1952): Dutch law invoked in 2025 for the first time to seize control of Nexperia from Chinese parent Wingtech, citing national security.
- Augment carefully / co-pilot by default: Combined-study quadrant where public finds automation acceptable but workers prefer human oversight — suggesting AI-assisted rather than AI-autonomous workflows.
Summary
The episode argues that the conversation about AI and employment has matured from broad existential fear to a more granular, empirically grounded analysis of where automation is wanted, where it is accepted, and where it faces genuine resistance. Two complementary academic frameworks — Harvard’s moral acceptability quadrant and Stanford’s worker-preference quadrant — together begin to sketch a societal map of AI deployment priorities. The key insight from combining these studies is that public acceptance of automation and worker acceptance of automation frequently diverge, and that resistance to AI is far more often rooted in capability doubts than in principled moral objections. Real-world adoption among blue-collar tradespeople further demonstrates that AI is already delivering meaningful productivity gains in unexpected sectors without displacing jobs. The Senate’s 100-million-job projection, while methodologically weak, is viewed as a useful provocation for a necessary policy conversation about the social contract in an AI economy. The host concludes that while long-term AI-driven abundance is likely, the near-term disruption will be significant and demands serious, good-faith societal deliberation — a process that these emerging studies and on-the-ground examples are finally making possible.