In the ongoing debate over artificial intelligence and employment, apocalyptic forecasts abound. Reports from think tanks, consultancies, and policymakers often warn that generative AI could automate vast swaths of white-collar work, displacing millions in fields from law and medicine to software development and creative industries.
Yet a new working paper from January 2026 challenges this narrative's foundational assumptions. Titled "O-Ring Automation" (NBER Working Paper No. 34639), economists Joshua S. Gans (University of Toronto) and Avi Goldfarb (Rotman School of Management, University of Toronto) argue that conventional risk assessments dramatically overstate displacement by treating job tasks as independent and separable — when in reality, they are deeply complementary.
The paper draws its name from the famous "O-ring theory" introduced by Michael Kremer in 1993, which posits that production quality depends on the multiplicative interaction of multiple tasks: if one fails (like a faulty O-ring in the Challenger disaster), the entire output suffers. In this framework, tasks aren't additive; their qualities multiply.
Automating one task doesn't simply remove 10% of a job if that task represents one of ten equal components — it reshapes the value of everything else.
Traditional exposure indices, such as those popularized in studies by Frey & Osborne (2013) or later adaptations for AI, calculate automation risk by averaging the automatability of individual tasks within an occupation. If AI can handle 90% of tasks in a role, the model predicts roughly 90% job loss. But Gans and Goldfarb counter that this logic breaks down under complementarity.
When machines take over routine or low-variance tasks, human workers don't become redundant; instead, they reallocate their fixed time endowment to the remaining manual bottlenecks. This "focus" mechanism amplifies the productivity and value of those irreplaceable human elements.
A canonical historical parallel illustrates the point: the introduction of automated teller machines (ATMs) in the 1970s and their widespread adoption by the 1990s. Conventional intuition suggested ATMs would eliminate bank tellers en masse.
Instead, teller employment in the United States remained stable or even grew for decades after ATMs proliferated (peaking around 2000 before gradual declines tied to broader banking consolidation).
Why? ATMs handled routine cash dispensing and deposits efficiently, freeing tellers to shift toward higher-value activities — building client relationships, cross-selling products, advising on loans, and resolving complex issues.
Banks expanded branches into underserved areas, leveraging cost savings to reach more customers, and tellers evolved into "relationship bankers" or "customer relationship team" members. The job didn't vanish; it transformed, often requiring new skills in sales, empathy, and interpersonal communication.
Similar patterns emerge in other domains. Radiologists, once burdened with routine image screening, can now focus on ambiguous cases and patient consultations when AI tools flag anomalies with high accuracy.
Paralegals and junior lawyers, relieved of document review drudgery by legal AI, devote more time to strategy, negotiation, and client counseling — tasks where human judgment, creativity, and ethical nuance remain superior. Data analysts augmented by AI tools produce deeper insights faster, concentrating on interpretation and business strategy rather than raw data cleaning.
The implications are profound. Partial automation can increase labor income by scaling the economic return to bottleneck human tasks. As machines improve smoothly, adoption often occurs discretely and in bundles — firms wait until AI reliably handles a critical mass of complementary tasks before fully integrating, leading to sudden shifts rather than gradual erosion.
Until AI closes every bottleneck in the production chain, workers in affected occupations may become more valuable, commanding higher wages or expanded roles.
Yet the authors highlight a double-edged sword. This same complementarity creates the potential for abrupt collapse. So long as meaningful human-exclusive bottlenecks persist, professions can flourish under AI augmentation. But if (or when) technology advances to substitute the final irreplaceable task — closing the entire O-ring loop—displacement could be rapid and near-total. Instead of slow decline, we might witness boom followed by bust: rising employment and productivity until the tipping point, then sudden obsolescence.
Ironically, Gans and Goldfarb embody their own thesis. While the paper does not explicitly state it in the abstract or main text, contemporary discussions note that many such economic analyses now leverage large language models for drafting, editing, and idea refinement — tools like ChatGPT-5 variants or Claude Opus equivalents.
The researchers themselves reportedly used advanced AI systems to assist in writing the paper, a meta-example of focus: AI handles structuring, literature synthesis, and prose polishing, allowing economists to concentrate on conceptual innovation, model rigor, and policy insight.
The paper's critique extends to policy. Linear exposure metrics, widely used to guide reskilling programs or universal basic income debates, likely exaggerate risks in complementary-task environments. Policymakers should instead examine bottleneck structures, time reallocation dynamics, and the pace of bundled adoption.
As AI capabilities accelerate — OpenAI's o1 series, Anthropic's Claude 4 family, and Google's Gemini advancements all demonstrate leaps in multi-step reasoning — the window for "focus-driven" productivity gains may narrow, but the transition could prove bumpier and less predictable than feared.
In an era of breathless headlines about job apocalypse, "O-Ring Automation" offers a sobering recalibration: AI may not hollow out occupations incrementally. It could supercharge them — right up to the moment it doesn't. Understanding complementarity isn't just academic; it could determine whether the AI era becomes one of widespread augmentation or abrupt rupture.
Also read:
- Cursor's AI Revolution: Building a Browser from Scratch with GPT-5.2 Agents in Just One Week
- LinkedIn Has Become the Internet’s Unofficial Dating Background-Check Tool
- The Religious UX Revolution: HaramBlur and the Fusion of Faith and Tech
Thank you!
