The question everyone asks about AI and jobs is the wrong question. People wonder whether AI will take their jobs, but this assumes we know enough about the underlying economics to give a meaningful answer. We don’t. The entire framework most people use to think about automation risk, the one built on “task exposure” scores and percentage overlaps between what AI can do and what your job description says, is measuring something real but nearly useless for predicting what will actually happen to employment. Economists have identified a single missing variable that could change this: price elasticity of demand. And the uncomfortable truth is that almost nobody is collecting it.
The conventional wisdom runs something like this: researchers map every task in a given occupation, then check which of those tasks AI systems can already perform or will soon be able to perform. The higher the overlap, the greater the risk. Research studies have done exactly this, using massive US government catalogues of thousands of job tasks first launched in 1998 to analyze which occupations are most exposed to AI capabilities. For example, analyses suggest that significant portions of certain white-collar jobs could be exposed to AI based on task analysis.
This feels rigorous. It has numbers. Percentages. Government data. But economists argue it tells us almost nothing about whether anyone will actually lose their job.
Why task exposure is the wrong measure
The logic gap is surprisingly simple once you see it. Knowing that AI can perform a significant portion of certain tasks does not tell you whether firms will employ fewer workers. It depends on something else entirely: what happens to demand when AI-driven productivity makes the service cheaper or faster.
Think about it concretely. If AI handles the tedious parts of real estate work (drafting listings, answering routine queries, scheduling), an agent can close more deals per week. Productivity goes up. Costs per transaction might fall. But does that mean the firm needs fewer agents, or does it mean the lower cost of each transaction brings more buyers into the market, creating more work for agents?
The answer depends on price elasticity: how much does demand for a good or service increase when its price drops? If demand is elastic (a small price drop creates a big surge in demand), then AI-driven productivity could lead to more hiring, not less. If demand is inelastic (people buy roughly the same amount regardless of price), then the productivity gain simply means you need fewer workers to serve the same market.
This is the piece of data economists are talking about. And right now, for most of the economy’s job categories, we simply don’t have it.
The grocery store knows more than the labor economist
What makes this gap almost absurd is that we do have granular price elasticity data for one sector of the economy: groceries. Research institutions have partnered with supermarkets to get data directly from their price scanners. We know, with fine-grained precision, how a 5% drop in the price of canned tomatoes affects volume sold.
We do not have the equivalent data for legal services, graphic design, software engineering, accounting, marketing, logistics, or any of the white-collar and service-sector jobs that dominate the AI displacement conversation. The data infrastructure simply wasn’t built for this question.
So every headline about AI replacing X million jobs by 2030 is built on a foundation that ignores the most important economic variable. The exposure data tells you where AI could intervene. It tells you nothing about the economic consequences of that intervention.
The prediction industry doesn’t want uncertainty
Definitive predictions generate clicks; uncertainty does not. “AI will replace 40% of jobs” makes headlines, while “we don’t yet have the data to know what AI will do to jobs” does not. This creates a dangerous feedback loop. Silicon Valley executives have strong incentives to project confidence about AI’s capabilities. AI company leaders have described AI as potentially transforming vast swaths of human labor in the coming years. Industry researchers have warned of potential economic disruptions and challenges to traditional career paths. These are serious claims from serious people.
But notice the structure of every argument in this space. It starts with what AI can do (task capability), leaps to what AI will replace (labor substitution), and skips the middle step entirely: what actually happens to prices, demand, and employment when those capabilities are deployed in real markets with real consumers making real purchasing decisions.
Research analyzing AI’s labor market impact has attempted to evaluate the current state of affairs, and the picture it paints is far more ambiguous than the apocalyptic or utopian narratives suggest. The actual displacement data, where it exists, is messy, industry-specific, and stubbornly resistant to sweeping conclusions. Even as the economics profession takes the question more seriously, it still lacks the empirical tools to answer it properly.
Why this missing data matters more than any AI benchmark
The gap between task exposure and actual job displacement isn’t academic. It plays out in real industries in ways that no capability benchmark would have predicted.
Consider the content industry. When AI could produce “good enough” articles, images, and marketing copy at near-zero marginal cost, the price of content collapsed. But demand for content turned out to be elastic with respect to quality thresholds and inelastic with respect to volume at premium price points. People wanted more content, but they weren’t willing to pay more for it. AI filled that gap. Certain business models were destroyed while others were created.
The specific demand shifts mattered enormously. Would consumers pay a premium for human-written content? Some, but not enough to sustain existing business models. Would advertisers value AI-adjacent content less? Yes, eventually. Would the sheer volume of AI content depress attention prices across the board? Yes, dramatically. Each of those outcomes is a price elasticity question. Each has a different answer depending on the specific market segment. And none of them could have been predicted by task exposure analysis.
Now multiply that across every industry in the economy. Without sector-by-sector data on how demand responds to AI-driven cost reductions, we are flying blind.
What “a Manhattan Project for data” would look like
Economists don’t just diagnose the problem. Some have called for a comprehensive, economy-wide effort to collect granular data on AI’s economic impact—an ambitious undertaking comparable in scale to the Manhattan Project.
What would that actually involve?
First, you’d need to track, in real time, how AI adoption in a given sector affects the price of that sector’s output. When a law firm deploys AI for document review, does the cost of legal services drop? By how much? How quickly?
Second, you’d need to measure how demand responds to those price changes. If legal services get 20% cheaper, do companies hire more lawyers? Do more individuals pursue legal claims they previously couldn’t afford? Or does the same volume of work get done by fewer people?
Third, you’d need this data across enough industries and enough time periods to spot patterns. Some industries might see massive demand surges when costs fall (think medical diagnostics, where cheaper testing could mean far more people getting screened). Others might see flat demand (people probably won’t buy twice as many accounting audits just because they cost less).
This kind of data collection doesn’t happen on its own. It requires coordination between government agencies, industry, and academic researchers. It requires new survey instruments, new partnerships with firms willing to share pricing data, and sustained funding over years.
As Silicon Canals has reported, the dataset that could predict AI job displacement barely exists, and nobody is collecting it in any systematic way. The political incentives are part of the problem: governments prefer to project confidence about managing technological transitions, and acknowledging that we lack basic data undermines that posture.
The industries where this matters most
Consider three contrasting cases that illustrate why elasticity is the key variable.
Healthcare. If AI-powered diagnostic tools cut the cost of a preliminary screening by 80%, what happens? In most countries, there is enormous unmet demand for healthcare. Billions of people who currently can’t afford diagnostic services would, in theory, use them if the price dropped far enough. High elasticity. The likely result: AI doesn’t replace doctors and technicians, it creates vastly more demand for follow-up care, interpretation, and treatment. Employment in healthcare could increase.
Legal services. If AI handles routine contract review and due diligence at a fraction of current cost, does demand for legal work surge? Probably not proportionally. Corporate legal needs are driven by regulation and deal flow, not by the price of the service itself. Moderate to low elasticity. The likely result: fewer junior associates needed per deal, but not a collapse in legal employment. A restructuring, concentrated at the entry level.
Content creation. If AI can produce articles, images, and video at near-zero cost, demand for content explodes (it already has). But demand for human-created content at premium prices is limited. Mixed elasticity depending on segment. The result: massive growth in total content output, substantial decline in paid human content work, and growth in curation and editorial roles. The early data bears this out: platforms are flooded with AI-generated material while freelance rates for routine writing have declined sharply, even as demand for high-end editorial strategy and brand voice work has remained stable or grown.
Each of these plays out differently. Each requires different policy responses. And none of them can be predicted by task exposure scores alone.
The political economy of not knowing
There’s a reason this data doesn’t exist, and the reason isn’t purely technical. Collecting economy-wide price elasticity data for AI-affected sectors would produce answers that some powerful actors don’t want.
If the data showed that AI displacement risk was concentrated in a few specific sectors (say, customer service, basic content production, and routine data entry) rather than spread across the entire white-collar workforce, it would undermine the narrative that AI companies need unlimited investment because they’re transforming everything. The narrative that AI will replace all jobs, paradoxically, serves the interests of AI companies by inflating the perceived value of their products.
On the other hand, if the data showed that displacement was more severe and widespread than current models suggest, it would create political pressure for regulation, retraining programs, and redistribution that both tech companies and governments are eager to avoid.
In my recent piece on Japan deploying robots because there’s no one left to hire, I looked at a country where the labor market dynamics are almost the inverse of what most Western countries fear. Japan faces a worker shortage so severe that automation is filling gaps rather than creating them. The same technology, radically different economic context, completely different outcome. That’s what happens when you pay attention to demand conditions rather than just capability conditions.
The ambiguity serves everyone with something to sell. AI companies sell inevitability. Consultants sell transition plans. Media companies sell fear and hope in alternating cycles. The one thing nobody is selling is the boring, granular, sector-specific demand data that would actually tell us what’s going to happen.
What you can actually do with this understanding
If you’re a worker trying to assess your own risk, the task exposure framework gives you a starting point but not a conclusion. Knowing that a significant portion of your tasks overlap with what AI can do is useful only if you then ask: what happens to demand for my industry’s output when AI makes it cheaper?
If you work in a field where lower prices create a lot more demand (education, healthcare, personalized services), AI might make you busier, not redundant. If you work in a field where demand is essentially fixed regardless of price (certain back-office functions, routine compliance, standardized reporting), AI is more likely to reduce headcount.
If you’re a policymaker, the lesson is starker. Every AI workforce strategy built on task exposure data alone is built on sand. The political response to AI and jobs should start with the most unsexy, most necessary step: fund the data collection. Build the equivalent of the grocery-store scanner partnerships, but for law firms, hospitals, engineering consultancies, marketing agencies, and school districts.
If you’re a business leader deciding whether to invest in AI-driven automation, the same variable matters: will making your output cheaper attract enough new customers to offset the reduced need for labor per unit of output? That’s a question about your specific market, not about AI in general.
The honest answer right now is that we don’t know. And admitting that is the first step toward actually finding out.
The cost of pretending we know
I write about systems of power, and one of the patterns I keep seeing is that the people with the most confident predictions about AI and jobs are the ones with the least accountability for being wrong. CEOs who predict mass automation don’t lose their jobs if it doesn’t happen. Economists who predict a smooth transition don’t bear the cost if millions of workers fall through the cracks.
I own stock in companies building these technologies. I use AI tools daily. I am not outside this system. Writing about AI’s effects on the economy while benefiting from AI’s capabilities is the kind of contradiction I’ve learned to sit with rather than pretend away. The question isn’t whether AI changes the labor market. It will. The question is how, and for whom, and at what speed, and in which directions. Price elasticity data, collected at scale across the economy, is the closest thing we have to a real answer.
Right now, that answer doesn’t exist. Not because the question is too hard, but because no one has funded the work. In a world where states are debating whether to pause data center construction and AI companies are raising tens of billions of dollars per quarter, the research budget to actually measure AI’s economic effects is a rounding error.
Price elasticity won’t tell us everything. But it would tell us the one thing we actually need to know: when AI makes something cheaper to produce, does the world want more of it or the same amount? The answer to that question, industry by industry, is the answer to whether your job grows, shrinks, or transforms.
We should find out before we have to guess.
Feature image by Tara Winstead on Pexels
