- Nearly 90% of CEOs surveyed in an NBER study report zero AI impact on employment or productivity at their companies so far.
- MIT field experiments show AI coding assistants increase task completion by 26%, but real project-level gains fall to 3-8% after accounting for non-coding work.
- Customer service job postings have rebounded to pre-COVID levels by mid-2025, contradicting predictions of mass automation in that sector.
- Economist Daron Acemoglu estimates AI will contribute only about 0.05% to GDP growth per year, far below market expectations.
What Happened
A wave of dire predictions about AI replacing white-collar jobs swept through early 2026. Microsoft AI chief Mustafa Suleyman estimated most professional work will be automated within 18 months. Former presidential candidate Andrew Yang declared “the AI jobpocalypse is here.” Anthropic published research modeling a potential “Great Recession for white-collar workers,” referencing how unemployment doubled from 5% to 10% during the 2007-2009 financial crisis.
But a growing body of evidence and analysis is pushing back against these predictions. James Wang, a general partner at Creative Ventures and former analyst at Bridgewater and Google X, published a detailed rebuttal arguing that the white-collar apocalypse “isn’t around the corner.” Independent analyst Martynas Miliauskas called the narrative unsupported by actual employment data. Writer Freddie deBoer has publicly wagered that the economy will remain “basically normal” through 2029 despite AI advances.
Why It Matters
The gap between AI predictions and observed outcomes has become difficult to ignore. A study from the National Bureau of Economic Research found that nearly 90% of CEOs surveyed report zero AI impact on employment or productivity at their companies. Customer service job postings, widely expected to be among the first casualties of AI automation, rebounded to pre-COVID levels by mid-2025.
Wang argues that the debate misunderstands which skills AI actually threatens. “Mechanical skills become less valuable over time. They always have,” he wrote, comparing the current moment to how photography eliminated the premium for realistic painting but elevated the value of artistic judgment.
Technical Details
The productivity data tells a more nuanced story than headlines suggest. MIT field experiments show AI coding assistants increase task completion by 26% on isolated tasks. However, developers spend only 11-32% of their time actually writing code. When accounting for non-coding work like requirements gathering, debugging, and code review, real project-level gains fall to an estimated 3-8%.
After one year of tool availability, adoption rates among developers reach approximately 60%. But quality concerns persist: studies indicate that 40-62% of AI-generated code contains serious issues requiring human review and correction.
Wang calculates that even a generous 10% productivity gain across the entire U.S. software development workforce of 1.7-2.1 million developers, with total compensation of approximately $357 billion annually, would yield about $36 billion in additional output, representing roughly 0.12% GDP growth.
Who’s Affected
Workers with primarily mechanical, routine skills face the highest displacement risk. Wang’s framework suggests that people whose value comes from judgment, such as deciding what to build and why, will become more valuable as AI handles execution tasks. Knowledge workers in roles requiring adversarial problem-solving or edge-case handling appear less vulnerable than initial predictions suggested.
Markets may be pricing in unrealistic expectations. The S&P 500 forward price-to-earnings ratio sits at 22.2x, 18% above its historical average, while the cyclically adjusted price-to-earnings ratio has reached 40, approaching dot-com era levels. These valuations exceed what software-only productivity gains can justify.
What’s Next
The skeptics acknowledge AI will create real changes but argue the timeline matters. Economist Daron Acemoglu, a leading AI skeptic, estimates AI will contribute only about 0.05% to GDP growth per year. Even optimistic projections from Wang suggest gains of 0.12% GDP growth from software productivity alone.
A recurring pattern undermines automation forecasts: teams that build systems capable of automating 90% of cases often find the remaining 10% requires most of their staff’s time, making full replacement impractical. One ex-big-tech engineer described building an LLM system that automated 90% of customer support cases, only to have the project shelved because the edge cases still required a full team. As Miliauskas summarized it: “They built an FAQ you can talk to.”
The technology continues to improve, but the path from impressive demos to reliable workplace deployment remains longer than many executives publicly predict. Historical parallels reinforce this pattern: the Luddites faced genuine disruption from power looms, yet textile industry employment ultimately expanded as the technology matured.
Source: Weighty Thoughts | Martynas M.
Related Reading
- Young Workers Pivot to Trades and Healthcare as AI Reshapes Entry-Level White-Collar Jobs
- Palantir CEO Alex Karp Says Only Trade Workers and Neurodivergent People Will Thrive in AI Era
- India AI Summit Draws 300,000 Attendees, Yields $250 Billion in Investment Pledges
- Research Shows AI Chatbots Misrepresent News Content 45 Percent of the Time
