For millions of workers, the conversation around artificial intelligence has shifted from a distant curiosity to a source of acute professional anxiety. This tension has manifested in tangible political pressure, including recent efforts to pause the construction of data centers in certain regions as communities grapple with the infrastructure of the AI boom.
The uncertainty is compounded by a lack of policy clarity. Lawmakers have yet to articulate a coherent plan for the labor market transformation that follows the integration of large language models into the white-collar workforce. Even economists, who previously suggested that AI would not lead to an immediate “job cliff,” are now acknowledging that the technology could have a unique and unprecedented impact on how humans operate.
The central problem is that we are trying to predict the future of employment using a flawed set of metrics. Most current analyses rely on “AI exposure”—a measure of how many tasks within a job can be performed by an AI. While this data provides a surface-level view of vulnerability, it fails to account for the actual economic drivers of displacement or growth.
Alex Imas, an economist at the University of Chicago, argues that our current tools for predicting these shifts are abysmal. According to Imas, the industry’s reliance on exposure data creates an illusory understanding of risk. To truly understand the impact of AI on the labor market, economists necessitate to move beyond what an AI can do and start collecting data on what employers actually do when productivity increases.
The Flaw in the ‘Exposure’ Metric
To understand why current predictions are failing, one must seem at how “exposure” is calculated. Most researchers use a massive catalogue of job tasks maintained by the U.S. Government via O*NET, a system first launched in 1998 to chronicle the specific duties of thousands of different occupations.
For example, a real estate agent’s job is broken down into individual tasks, such as interviewing clients about their property preferences. By comparing these tasks to the capabilities of AI, researchers can assign an exposure percentage. In one such study, OpenAI researchers found that real estate agents were 28% exposed to AI. Similarly, Anthropic analyzed millions of Claude conversations to see where actual user behavior overlapped with these predefined task lists.
However, Imas contends that “exposure alone is a completely meaningless tool for predicting displacement.” Knowing that a machine can perform 30% or even 70% of a person’s tasks does not tell us if that person will be fired, if their role will evolve, or if the company will simply produce more output with the same headcount.
When Exposure Leads to Displacement
There is one specific scenario where exposure data is an accurate predictor: the “gloomy case.” This occurs when every single task in a job can be performed by AI without human direction and the cost of running the AI is lower than the human worker’s salary.
Here’s the modern equivalent of the elevator operator—a role that vanished once automatic controls became reliable and cheaper. In today’s economy, this might look like a customer service agent whose sole responsibility is phone call triage. If an AI agent can handle 100% of the triage and do so at a fraction of the cost, the job is likely to disappear.
But this is the exception, not the rule. For the vast majority of professional roles, the relationship between AI and employment is far more complex. The cost of AI is not always negligible; reasoning models and agentic AI can rack up significant operational costs, meaning the “cheaper” alternative isn’t always a guarantee.
The Productivity Paradox: More Output or Fewer Workers?
The real mystery lies in the gap between productivity and headcount. Consider a software engineer building a premium dating app. With the help of AI coding tools, a developer might complete in one day what previously took three. This represents a massive spike in productivity.
At this point, the employer faces a strategic choice that “exposure” data cannot predict:
- Expansion: The employer keeps the developer and uses the reclaimed time to build three times as many features, expanding the product and potentially hiring more people to manage the growth.
- Contraction: The employer decides they only need the original amount of output and reduces their engineering staff by two-thirds, as one developer can now do the work of three.
Without data on employer intent and market demand, we cannot grasp which path a company will take. The impact of AI on the labor market is determined not by the tool’s capability, but by the employer’s response to that capability.
The Path Toward Better Data
To move past these “abysmal” predictions, Imas is calling for a new approach to data collection. Instead of focusing on the theoretical capabilities of the software, economists need to track the actual shifts in hiring patterns and task reallocation in real-time.
| Metric | Focus | Predictive Power | Limitation |
|---|---|---|---|
| AI Exposure | Task capability | Low (for displacement) | Ignores market demand & cost |
| Productivity Gain | Efficiency increase | Moderate | Doesn’t specify headcount change |
| Employer Intent | Hiring/Firing trends | High | Difficult to collect at scale |
The goal is to move toward a “call to arms” for economists to identify the specific triggers that lead to displacement versus those that lead to job enrichment. This requires a granular look at how roles are being rewritten in the wake of AI adoption.
As the industry moves toward more autonomous “agentic” AI, the stakes for this data increase. The ability of AI to not just suggest text, but to execute complex workflows, may push more jobs into the “gloomy case” category, making it imperative that policymakers have a grounded understanding of the labor shift before the cliff arrives.
The next critical checkpoint for understanding these shifts will be the release of updated quarterly labor statistics and the ongoing analysis of AI’s effect on specific sector employment, which will provide the first empirical evidence of whether productivity gains are leading to expansion or contraction.
How has AI changed your daily tasks? We invite you to share your experiences and perspectives in the comments below.
Keep reading
