The Wrong Operating System for AI
The AI economy is colliding with America’s broken social contract
TLDR
The United States did not design its institutions for an AI economy that repeatedly reshapes work. It built a digital system that rewards data extraction, leaves adoption uneven and shallow, pushes the burden of upskilling onto workers, and relies on a safety net built for older forms of disruption.
The result is not necessarily one sudden employment collapse, but something more plausible and more corrosive… privacy erosion, widening capability gaps, weaker bargaining power, and a labor market where too many people are expected to adapt without the time, money, or institutional support to do it.
The United States has not yet proven that AI will produce mass unemployment. It has proven something arguably more dangerous. It has built an economic and regulatory stack that makes privacy extraction profitable, makes deep technology adoption uneven, and leaves workers largely responsible for financing their own adaptation inside institutions that were not designed for continuous mid-career retraining. The problem is not just the models. The problem is the operating system around them (U.S. GAO 2019; Federal Trade Commission 2024c; U.S. Census Bureau 2025; U.S. Department of Labor 2007).
Problem
Modern American society runs on a simple asymmetry. Firms can collect, combine, and monetize data faster than workers can secure time, money, and institutional support to reinvent themselves.
The public story says innovation is democratizing.
The institutional reality is that value concentrates while transition risk gets pushed downward.
That is why privacy erosion and labor fragility keep showing up together. They are not separate failures. They come from the same incentive structure (Federal Trade Commission 2024c; Greenstein 2025; U.S. Department of Labor 2021).
Privacy as business model
There is still no comprehensive federal internet privacy law governing private companies’ collection, use, and sale of consumer data. Instead, the United States relies on a patchwork of sector rules, FTC enforcement, and a growing mix of state privacy statutes (U.S. GAO 2019; IAPP 2026).
At the same time, the FTC has documented that the business models of major platforms incentivize mass collection of user data for monetization, especially through targeted advertising, and its recent enforcement actions show how browsing and location data can be harvested, profiled, and sold with little meaningful user control (Federal Trade Commission 2024a, 2024b, 2024c).
Privacy loss is not some accidental byproduct of the system. For much of the digital economy, it is the revenue engine.
Underadoption is not the same as readiness
This is where the labor story gets strange. Public discourse makes it sound like AI is already everywhere. The evidence says adoption is both broad and shallow.
McKinsey’s 2025 global survey found that 88% of respondents say their organizations use AI in at least one business function, but most are still in experimentation or pilot phases and only about one-third say they have begun scaling AI programs across the enterprise (McKinsey & Company 2025a).
Census data tell a harder truth about the real economy. Only 3.9% of U.S. businesses reported using AI to produce goods or services in late 2023, and overall business use was still only around 10% by May 2025 (U.S. Census Bureau 2023). Earlier Census research also found AI adoption highly uneven, concentrated among large firms and a relatively small set of sectors and places (Foster et al. 2023).
So yes, AI is spreading. But no, the country has not deeply reorganized work around it. Much of America is not over-adopting AI. It is barely operationalizing it.
That matters because shallow adoption is deceptive. It gives leaders the feeling of progress without forcing the hard work of workflow redesign, role redesign, and real capability building. McKinsey’s own survey is blunt on this point, the biggest difference between firms that get real value from AI and those that do not is not mere tool access. It is whether they redesign workflows and embed AI into operating reality (McKinsey & Company 2025a).
That means the current American pattern is almost the worst possible one. Enough adoption to create anxiety, concentration, and hype, but not enough disciplined transformation to spread the gains.
Why upskilling keeps failing
Workers are not oblivious to the shift. They are already telling us they see it coming. Gallup found that only 45% of U.S. employees participated in training or education to build new skills for their current job in 2024 (Gallup 2025). McKinsey, citing Indeed’s 2024 workplace survey, reported that 75% of U.S. workers expect AI to shift their roles within five years, but only 45% have received recent upskilling (Bérubé, Metakis, and Ocampo 2025). The OECD adds the most important part, the biggest barrier to adult learning is still time, with cost and other structural obstacles following behind (OECD 2025).
Americans are not failing to adapt because they are too comfortable. They are trying to adapt inside a system that makes adaptation hard.
The burden is also not distributed evenly. A recent Nature Communications paper analyzing 167 million U.S. job posts found that lower-skilled occupations face greater upskilling pressure than higher-skilled ones (Tong, Wu, and Evans 2026). At the same time, research highlighted by Harvard Business School found a major digital divide across more than 28,000 ZIP codes, with rural areas lagging urban ones in actual computer use and digital fluency (Lorch 2025).
So the people most likely to need rapid adaptation are often the least institutionally equipped to get it. That is how capability gaps become class gaps and regional gaps.
The Department of Labor made the underlying problem clear long before the current AI cycle. The traditional higher-education model was not built for adult learners, most of whom are “employees who study” rather than “students who work” (U.S. Department of Labor 2007).
That remains the core institutional mismatch. The United States keeps asking workers to solve a structural problem with individual heroics.
The safety net is real, but it is a misfit
It is important to be precise here. The U.S. does have a safety net, and it matters. The Hamilton Project finds that the safety net now cuts poverty nearly in half and has substantially reduced the share of Americans without health insurance (Greenstein 2025). But that same record shows the mismatch: unemployment insurance coverage has declined, and real spending on TANF and its predecessor AFDC fell 78% from 1993 to 2016 (Greenstein 2025). This is not a nonexistent safety net. It is a fragmented one, with major holes in exactly the places where labor-market volatility hits hardest.
The labor model makes that worse. The Department of Labor warned that misclassification can reduce workers’ access to fringe benefits such as health insurance, retirement plans, unemployment insurance, and workers’ compensation (U.S. Department of Labor 2021). In an AI economy, that matters because the same firms most eager to automate coordination, service, logistics, and back-office functions are often the ones most comfortable externalizing labor risk. Benefits remain too tied to formal employment relationships, while learning remains too detached from paid working time. That is a bad combination for an era of repeated skill shocks.
What the evidence actually says about unemployment
The honest version is not that mass unemployment has already arrived. It has not. Census reporting on the 2023 Annual Business Survey found that adoption of technologies including AI generally had little impact on overall worker numbers between 2020 and 2022 (U.S. Census Bureau 2025). Overclaiming here is a mistake because it weakens the case.
But the opposite mistake is worse. The absence of an immediate employment cliff does not mean the system is healthy.
It means the adjustment may arrive as a slower, more politically confusing form of damage: uneven adoption, wage pressure, occupational downgrading, regional divergence, and a widening gap between workers who can continuously retool and workers who cannot. The country may not get one clean employment apocalypse. It may get a long unraveling of privacy, bargaining power, and labor mobility instead. That is harder to measure, easier to deny, and more consistent with the evidence we actually have on shallow adoption, unequal reskilling pressure, and digital divide dynamics (McKinsey & Company 2025a; Tong, Wu, and Evans 2026; Lorch 2025).
Default recommendation
If that diagnosis is right, then the response is structural. The United States needs a real federal privacy framework, adult-centered learning systems, training leave or similar income support for mid-career transitions, portable benefits that are less dependent on a single employer relationship, and a much harder distinction between shallow AI experimentation and genuine operational adoption.
Until then, the country will keep doing the same foolish thing… privatizing the upside of data and automation while socializing the instability that follows (U.S. GAO 2019; Federal Trade Commission 2024c; OECD 2025; Greenstein 2025; U.S. Department of Labor 2007).
That is the core answer. Modern American systems are not built to guarantee a clean future of shared AI prosperity.
They are built to extract data efficiently, diffuse responsibility for worker adaptation, and respond to labor disruption after the damage is already visible.
Privacy erosion and labor fragility are not separate problems. They are two expressions of the same institutional logic. The machine learns fast. The society around it does not.
References
Bérubé, Vincent, Marc Metakis, and Maria Ocampo. 2025. “Redefine AI Upskilling as a Change Imperative.” December 1, 2025. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-organization-blog/redefine-ai-upskilling-as-a-change-imperative
Federal Trade Commission. 2024a. “FTC Cracks Down on Mass Data Collectors: A Closer Look at Avast, X-Mode, and InMarket.” March 2024. https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/03/ftc-cracks-down-mass-data-collectors-closer-look-avast-x-mode-inmarket
Federal Trade Commission. 2024b. “FTC Order Will Ban InMarket from Selling Precise Consumer Location Data.” January 2024. https://www.ftc.gov/news-events/news/press-releases/2024/01/ftc-order-will-ban-inmarket-selling-precise-consumer-location-data
Federal Trade Commission. 2024c. “FTC Staff Report Finds Large Social Media and Video Streaming Companies Have Engaged in Vast Surveillance of Users with Lax Privacy Controls and Inadequate Safeguards for Kids and Teens.” 2024. https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-staff-report-finds-large-social-media-video-streaming-companies-have-engaged-vast-surveillance
Foster, Lucia, Emin Dinlersoz, Kristina McElheran, Nikolas Zolas, Erik Brynjolfsson, Zachary Kroff, and J. Frank Li. 2023. “AI Adoption in America: Who, What, and Where.” Center for Economic Studies Working Paper CES-23-48R. September 2023. https://lehd.ces.census.gov/applications/creat/paper-profile/1211
Gallup. 2025. “Addressing the Barriers Blocking Employee Development.” July 22, 2025. https://www.gallup.com/workplace/692642/addressing-barriers-blocking-employee-development.aspx
Greenstein, Robert. 2025. “Changes in the Safety Net over Recent Decades and Their Impact.” The Hamilton Project, May 1, 2025. https://www.hamiltonproject.org/publication/paper/changes-in-the-safety-net-over-recent-decades-and-their-impact/
IAPP. 2026. “US State Privacy Legislation Tracker.” Last updated February 10, 2026. https://iapp.org/resources/article/us-state-privacy-legislation-tracker/
Lorch, Danna. 2025. “America’s Digital Divide: Where Workers Are Falling Behind.” February 10, 2025. https://www.library.hbs.edu/working-knowledge/americas-digital-divide-where-workers-are-falling-behind
McKinsey & Company. 2025a. “The State of AI in 2025: Agents, Innovation, and Transformation.” November 5, 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
OECD. 2025. “Why Is Participation Not More Common?: Trends in Adult Learning.” 2025. https://www.oecd.org/en/publications/trends-in-adult-learning_ec0624a6-en/full-report/why-is-participation-not-more-common_73893b53.html
Tong, Di, Lingfei Wu, and James A. Evans. 2026. “Lower-Skilled Occupations Face Greater Upskilling Pressure in U.S. Job Ads.” Nature Communications 17: 1237. https://www.nature.com/articles/s41467-025-67992-y
U.S. Census Bureau. 2023. “How Many U.S. Businesses Use Artificial Intelligence?” November 28, 2023. https://www.census.gov/library/stories/2023/11/businesses-use-ai.html
U.S. Census Bureau. 2025. “How AI and Other Technology Impacted Businesses and Workers.” September 17, 2025. https://www.census.gov/library/stories/2025/09/technology-impact.html
U.S. Department of Labor. 2007. “Adult Learners in Higher Education: Barriers to Success and Strategies to Improve Results.” January 1, 2007. https://www.dol.gov/index.php/agencies/eta/research/publications/adult-learners-higher-education-barriers-success-and-strategies
U.S. Department of Labor. 2021. “US Department of Labor to Withdraw Independent Contractor Rule.” May 5, 2021. https://www.dol.gov/newsroom/releases/whd/whd20210505