Paranoid Android: Revolution or transition?

Read the full article (PDF)open_in_new

What research actually tells us about AI and the labour market

« Please could you stop the noise?
I’m trying to get some rest. »
– Radiohead

Introduction: When fear outpaces facts

Artificial intelligence (AI) has become the new emotional centre of global markets.

Every hint, every research note, every stray headline about its disruptive potential triggers chain reactions that feel more psychological than analytical. The lateFebruary selloff after the now-famous Citrini scenario is the most vivid example to date. A report built around a hypothetical white-collar recession in 2028 was enough to knock down a range of stocks that included gig economy platforms, credit card issuers, and private equity leaders.

Even though global investors most likely understood perfectly that the document described a thought experiment, not a forecast, it didn’t matter. When a theme is this charged, fear can move prices faster than facts. Markets have become like the Radiohead narrator whispering “Please stop the noise” while imagining every possible outcome at once.

In this paper, we aim to bring the noise back to a manageable volume.

AI is changing the economy and will change it in profound ways. But the speed, sequence, and distribution of the changes matter greatly. A firstprinciples reading of the available research offers a calmer and more grounded path than the dramatic scenarios that sometimes capture headlines. Roles and occupations are unlikely to disappear overnight; instead, specific tasks will be replaced gradually. The story is about understanding where complementarity between humans and machines is high and where it is not.

Fear in markets tends to come from category errors. When investors react to headlines as if all AI were equal and all jobs were equally at risk, they drift into the kind of noise that inspired our title.

Tasks before jobs

The most reliable foundation for understanding AI’s economic impact is still the task-based approach. Occupations comprise bundles of tasks with different levels of predictability, structure, and physical coordination. AI will not replace entire occupations; it will replace specific tasks within them. This important distinction is the key to separating hype from economic reality.

Exposure is not the same thing as automation. AI can touch many tasks without replacing the humans who perform them. For example, research from Pizzinelli and colleagues (2023) makes this distinction explicit.

Researchers have developed the concept of occupational exposure to AI (called AI occupational exposure scores, or AIOE), which measures how many of a job’s tasks overlap with what AI can theoretically do, not whether the machine is suited to take the lead. With adjustments for complementarity (C-AIOE), a different picture emerges.

Jobs that depend on empathy, subtle communication, deep expertise, or nuanced judgment tend to invite partnership rather than substitution. In these settings, AI acts like a powerful assistant. It accelerates the routine parts of the work and gives the human more room to think. For example, lawyers reviewing documents, architects refining concepts, or managers interpreting complex situations may all see high exposure in theory, but the practical risk of their being replaced turns out to be much lower.

The research helps explain why this gap appears. These higher-skill roles combine the ingredients that AI has trouble replicating. They involve responsibility for outcomes, unpredictable contexts, and interactions where trust and intuition matter.

That combination pulls their complementarity scores upward and reduces their adjusted exposure. AI can screen information, analyze patterns, and draft preliminary material, but it cannot operate meaningfully without the human who provides context and final judgment.

As a result, the technology becomes a productivity amplifier rather than a substitute. Thus, professionals and managers often see the largest drop between raw exposure and complementarity-adjusted exposure. AI affects their work, but it alters the workload rather than removing the worker.

The opposite is true for routine cognitive work.

Data entry, templated documentation, simple request handling, and other predictable interactions present very little complementarity because the human does not add much beyond the task itself. Firms adopting AI are indeed reducing headcount in these categories. Such disruption has happened several times in the long history of technological innovation. It is not a precursor of economic collapse.

The research also reminds us that AI is not a single technology. Vision models, language models, and autonomous decision-making systems affect different task classes in very different ways.

Vision tools and large language models have their biggest effects in clerical work. Autonomous control systems are more likely to support workers in operational or industrial settings. As recent research has shown¹, some firms adopting these tools are even expanding their manual workforce because AI enhances the value of human experience, leading to more hires and faster wage increases in jobs with high AI exposure.

The conclusion from this first section is simple. The near-term impact of AI is selective, narrow, and uneven. It is neither a tidal wave nor a mirage. Tasks will be reshuffled: some workers will lose their jobs, but many others will become more productive. The outcome will be a different kind of labour market, not mass unemployment.

Labour markets in slow motion

A curious divergence continues to emerge. Blue-collar work has been far less affected by AI than white-collar support functions. Construction, agriculture, and production jobs remain anchored in the physical world, which is still a difficult environment for general-purpose automation. These occupations involve responsibility, unpredictable conditions, and significant physical dexterity. For now, there is little evidence of broad displacement. In fact, when firms introduce advanced equipment or AI-enhanced machinery, they often add workers because the technology raises the value of human oversight. The benefits show up through higher output and safer operations rather than through shrinking payrolls.

But that picture may not hold forever. Progress in robotics is accelerating in ways that could shift this conclusion sooner than expected. Early demonstrations of general-purpose humanoid systems, such as Tesla’s Optimus robot, hint at a world where machines can eventually navigate environments that challenge traditional automation today. Optimus is still an earlystage product, but this frontier could move quickly once hardware, sensors, and AI control systems advance together.

A robot that could walk, grip, circumvent obstacles, and manipulate tools in real time would fundamentally change the calculus for many blue-collar tasks. We are far from that point; yet, what seems to be sheltered today may become exposed tomorrow. Readers should be mindful that the boundary between routine and nonroutine physical work may shift if these platforms reach commercial scale.

That said, lower-skill white-collar roles are currently where the negative part of the story can be found. Clerical support has been shrinking in those firms that adopt AI most aggressively. It shows up as fewer junior hires rather than as mass layoffs. Still, that subtle distinction has large macro implications because it changes the pipeline of experience in the economy and may become a constraint on future productivity growth if it lasts.

The most positive side of the story is currently found at the top end of the skills spectrum.

Demand for high-skill white-collar roles seems to be on the rise, as evidence mounts that firms are hiring more technical experts, more analysts, more engineers, and more managers who can design or supervise AI systems. This demand mirrors the pattern seen in earlier general-purpose technologies where adoption led to upskilling inside firms rather than labour displacement.

On the compensation front, there is also little sign yet of a general, AI-driven wage shock. Recent research from the Federal Reserve Bank of Dallas shows that the “experience premium” – the distribution of wages for workers with different experience levels within an AIexposed sector – is starting to become more evident. Lower-skilled workers are paid less on average, and higher-skilled workers are paid more, but the overall wage bill has yet to move meaningfully.

This is the core message for labour markets. The AI transition is happening but at a measured pace. Gains are concentrated in higher-skilled workers, while the market for recent graduates is becoming more challenging. It does not mean mass unemployment, but it may create subtle bottlenecks in the flow of future experience with implications for long-term growth.

The known unknowns that matter

To summarize, even though we expect AI to disrupt the labour market in significant ways, the positive impact of job augmentation will outweigh the negative impact of job displacement. Even so, major uncertainties remain.

New task creation is the biggest one. Every transformative technology eventually generates new categories of work that absorb displaced labour and raise productivity. For AI, these tasks are only beginning to appear. The present generation of AI-related tasks mostly extends digital work rather than reinventing it. Some of the new tasks have very little social or economic value. Others, especially in health care and education, have far greater potential.

The pace at which truly productive new tasks appear will shape the long-term path of growth and might eventually force us to rethink entirely our approach to the economics of AI.

A second major uncertainty concerns AI’s effect on the innovation engine itself. If AI truly accelerates the generation of ideas, it could reshape long-run growth by expanding the stock of knowledge at a faster pace. Idea production is the ultimate driver of productivity gains in advanced economies, and AI’s ability to search vast datasets, simulate outcomes, and generate new designs could raise the speed at which firms discover new products and processes. In that world, the growth frontier moves outward more quickly, and the economy benefits from a faster flow of breakthroughs.

The picture changes if these capabilities remain tightly held by a small group of dominant firms or countries. Concentration would limit how far and how fast innovations spread across the global economy. Productivity would rise inside a few walls, but the wider economy would see only modest gains. Much depends on how easily competitors or new entrants can gain access to the tools and the compute infrastructure needed to run them, and on whether each country has the infrastructure and political capital to position itself for AI adoption.

If barriers to entry remain high, AI could tilt the playing field toward incumbents and reduce the competitive pressure that normally drives diffusion. The macro impact would then hinge not only on what AI can do but also on who is able to use it.

What this means for markets

The market narrative that dominated at the start of 2026 highlights how sensitive investors have become to the future of AI. The most recent scenarios have become extreme and speculative, with an emphasis on potential disruptions to the world as we know it. In classic hype-cycle behaviour, investors reacted as if such disruption were imminent.

Every investor knows that expectations adjust in bursts, not in smooth lines. The true path of AI adoption is likely to be much slower and more textured than investor psychology often assumes.

Our take is that, in the near term, AI is neither a miracle nor a menace. It is not about to double productivity or to erase millions of jobs. It is a gradual transition unfolding task by task. Markets that overprice immediacy will continue to swing wildly at every unexpected headline. Investors who focus on real-use cases, sustainable productivity benefits, and long-term integration will make better decisions.

The most likely scenario remains one of meaningful but steady gains spread over a decade. Firms that integrate AI in ways that complement human capability should outperform. Firms that ignore genuine AI adoption rather than embracing it will eventually be pushed out of the market. But, in the end, the scenario where AI simply takes over the world is far-fetched.

Final thoughts

Fear often rises when the future is loud and unclear. AI presents an extraordinary opportunity for the world economy, but it requires calm assessment rather than dramatic leaps. The task-based approach grounds us in reality. It shows a path of selective automation, humanmachine collaboration, uneven labour-market shifts, and gradual productivity improvements. Investors who navigate this landscape with realism rather than anxiety will make better decisions and avoid succumbing to the noise.

¹ Example: https://www.dallasfed.org/research/economics/2026/0224