Closer to home for most of us, the release of large language models like ChatGPT have renewed worries about automation, reminiscent of earlier fears about mechanization. AI has advanced far beyond rote data-storage tasks and can even pass the bar exam, or write news, or research papers, leading to fears of massive white-collar unemployment.

But, as new research looking at data of job churn over the past two decades finds, the impact of automation on workers and industries is, in fact, pretty hard to predict given the complexity of the labor market, requiring carefully crafted policies that take these nuances into account.

First, changes in exposure to automation are not intuitive: they do not easily mesh with “blue-collar” and “white-collar” jobs, as typically defined. Instead, automation is more closely linked to the tasks and characteristics of each job, such as repetitiveness and face-to-face interactions. That translates to the three most automation-exposed jobs: office and administrative support, production, and business and financial operations occupations.

Meanwhile, the three least automation-exposed jobs are in personal care; installation, maintenance and repair occupations; and teaching. In other words, even with the Internet of Things controlling your HVAC system, it cannot fix itself when it needs new refrigerant, but its smart-panel interface can help the technician diagnose the problem remotely quickly and know what equipment to bring for a repair. But back-end accountants in that company may not fare as well in the AI jobs sweepstakes.

While automation can displace workers, history suggests that new technology also tends to boost productivity and create new jobs. Consider the automobile: while horses and buggies are outdated, we still need humans to drive (at least until autonomous vehicles come to full fruition), and the assembly line helped automate manufacturing with entire new classes of jobs created for every part of a car and all its electronic systems, with almost 1 million U.S. workers in auto manufacturing today.

But automation has continued in the auto industry over the decades, with robots helping to make hard and heavy physical labor tasks easier, without fully displacing workers.  So there is a push-pull with automation, and the relative sizes of these countervailing effects remains an area of active scholarly debate.

Second, it is rare for an entire job class to disappear overnight; changes mainly take place over generations. The research shows that newer generations of workers, perhaps deterred by the job insecurity observed in earlier generations and lured by high wages in the technology sector, are less inclined to enter automation-prone jobs than those before them. However, after embarking down those career paths, workers tend to stay in their fields, even if the prospects of automation loom large, likely because reskilling is time-consuming and expensive. It is relatively easy for recent high school graduates to opt for tech-centric college degrees like computer science, but learning new skills like coding is more difficult for mid-career professionals in automation-susceptible fields like manufacturing.

Adjustments to automation can be slow on the business side as well. Incorporating automated technology takes time because modern production tasks tend to be so intertwined that automating one part of a business can affect all other operations. For example, when AT&T, once the country’s largest firm, began replacing telephone operators with mechanical switchboards, they found that operators had become central to the complex production system that grew around them, which is why there are fewer operators today, but some still exist.

Third, the research found that the share of workers in highly automation-exposed occupations tends to be clustered, ranging from about 25% to 36% across commuting zones. The least-exposed areas in the U.S. are across the Mountain West, thanks to the area’s high shares of workers in management, retail sales and construction (which hasn’t had much automation or productivity improvement in decades but additive manufacturing may be a game-changer), as well as those on the East and West coasts, with their more innovative finance and tech industries.

On the other hand, those most exposed to automation tend to be located in the Great Plains and Rust Belt, namely due to agriculture. In spite of the fact that U.S. agriculture has been exposed to automation for over a century (more efficient machines and advances in biotechnology), it has become even more technology-driven recently, making ag workers more likely to be impacted by automation.

Read: How artificial intelligence can make hiring bias worse

So will the robots take over your job soon?  More likely, they will make our jobs easier and more efficient. Trying to slow the adoption of technology is both futile and counterproductive: taxing or overregulating tech adoption may backfire, especially given global competitiveness and other countries who may not pause. While the advent of a new era of automation is likely to be both gradually incorporated and result in complements to human labor rather than full replacement, thoughtful policies can help disrupted workers transition to new and better opportunities, ensuring we can harness the transformative power of automation and foster a future of work that benefits all.

Eric Carlson is associate economist at the Economic Innovation Group; DJ Nordquist is EIG’s executive vice president.

More: AI is ready to take on menial tasks in the workplace, but don’t sweat robot replacement (just yet)

Also read: ‘Make friends with this technology’: Yes, AI is coming for your job. Here’s how to prepare.