Will AI replace workers? Not if we build it right

AI is amazing, and will likely become even more so.

You can see the excitement in viral blogs, op-eds and social media posts that proclaim not just the tremendous benefits we will get from AI but how this is already happening today. You can also see it in the record-breaking amount of investment pouring into graphics processing units (GPUs), data centers and AI startups.

At the same time, we are told “very soon most of you will lose your jobs”.

Americans believe this message, and 52% of them are worried about how AI will impact their jobs. They are also concerned about how AI will transform many different aspects of their lives. Others fear that AI will pollute our informational ecosystem and end democracy as we know it. Many go one step further and view AI as a “civilization-ending” threat that can wipe out or subjugate humanity.

But the hype, the excitement and the extreme fears need to be tempered, not just because they are misplaced, but because they are misguiding us on how we should develop and use this promising technology.

In fact, AI’s effects will come only slowly, because achieving the kind of reliability and versatility that is required for automating whole jobs isn’t here yet. Integrating any new technology, let alone one as complex and untried as AI, can only happen slowly and will necessitate myriad changes in many organizations.

Worse, on the current trajectory, AI will be nowhere as beneficial as the industry insiders expect. We may be on the verge of squandering the tremendous potential of artificial intelligence, because we are not conceptualizing it the right way, and we are not developing it the right way.

We are making the mistake of attempting to force distinctly different artificial intelligence to mimic human intelligence. These two intelligences are fundamentally different in nature, and when two things are different, the natural way to proceed is to complement one with the other—not to pretend those differences don’t exist and strive for having one take over everything.

Everywhere but in the productivity statistics

Many people have been forecasting that AI will drive remarkable improvements in productivity and economic output.

But reality seems to be closer to the great economist Robert Solow’s 1987 quip about computers—“everywhere but in the productivity statistics”. AI advances aren’t in the productivity statistics either, at least as of now.

Achievements of current AI models in writing computer code and designing software or in radiology diagnostics are truly remarkable. But that is in the lab, where context is clear and tasks are neatly delineated. The evidence from firms that adopt these tools for the same tasks “in the wild” isn’t as impressive. Often, engineers have to spend a lot of time debugging AI-written code and radiologists and doctors cannot combine their expertise with AI tools, leading to suboptimal outcomes. Several studies and reports have failed to find much in the way of productivity gains from adoption of AI by most companies. Of course, things may be different, and much better, with new AI agents, but so far the disappointing productivity performance isn’t confined to the aggregate. It is also there for firms that are experimenting with AI.

None of this should be surprising. Every new technology takes time to be productively adopted, and typically, the more transformative the technology, the slower this process is. Some draw parallels between AI and fire. A less hyperbolic comparison might be to electricity.

Famously, electricity, despite all of its revolutionary potential, took several decades to be adopted and used widely, even in manufacturing. Already in 1881 New York and London had central generating stations and many were foreseeing that electricity would completely transform manufacturing, homes and transport. But as late as the 1920s, only about half of factories and homes were using electricity. We may in fact expect productive adoption of AI to be even slower because of the necessary reorganization of firms. Job tasks will have to be reallocated between AI models and workers, and many specialists will have to learn how to work with AI. Managers will have to develop their own understanding of what it is that they want from AI and how they can work towards getting it. Worse, many firms may attempt to go for full automation of certain occupations, which is likely to remain impossible in the near future.

All of this will not just delay the productivity gains from AI, but imply that job displacement effects will be slow as well. Several rounds of forecasts predicted the end of radiology or drivers or accountants: none of that has happened yet. Hospitals still need radiologists (in fact more than before), there are still journalists, paralegals, accountants and office workers.

Two years ago, I estimated that within 10 years, we shouldn’t expect much more than about 5% of what humans do to be replaced by AI.

Where does this 5% number come from? Though it is admittedly no more than a guesstimate, it was based on an assessment of what AI models could achieve then and how these capabilities would develop in the second half of the 2020s.

To put it simply, my forecast was that by the end of the 2020s AI would be in a position to successfully automate simple office and cognitive tasks, making up about 20% of work in the US economy. I predicted that the impact on more complex cognitive and office jobs would be slower in coming, because it would take AI much longer to acquire the judgment, multi-dimensional reasoning and social skills that most jobs require. Nor would society be comfortable with AI fully taking over air traffic control, mental health therapy or the management of companies.

AI-based automation of jobs that involve heavy interaction with the physical world will also have to wait, because it will take time for flexible robotics, AI models with a full sense of spatial-causal relations. It may be even longer for computer vision technologies to reach the point where AI could be seamlessly combined with robots to perform tasks in manufacturing and construction.

I also estimated that, given the diffusion of other technologies, such as electricity, computers and previous computer vision technologies, only about one quarter of the 20% that can be fully automated would be automated within 10 years. This gave the 5% number.

My forecast (with the full warning that it was and remains impossible to know how quickly these models would advance) was that GDP would increase by about 1.5% over 10 years. Since then, advances in models have been more rapid than many, including me, predicted. On the other hand, we haven’t seen many applications that can easily be adopted and transform the production process, and, as I explain below, the magic of AI needs to happen in the applications for it to have a sizable impact on the economy.

In balance, those predictions don’t look too bad—yet.

Lessons from voice recognition

Why is it so hard for such a revolutionary technology to have a big impact on the economy? The saga of voice recognition gives us some of the answers.

Turning speech into text has been an aspiration since the 1930s. It became a huge step closer to reality thanks to the work by the American husband and wife team, James and Janet Baker, starting in the 1980s, when they founded the iconic company Dragon Systems.

Dragon truly revolutionized voice recognition, and it would not be an exaggeration to say that everything else that has followed in voice recognition since then is thanks to the Bakers. Their breakthrough was to eschew previous attempts to build speech recognition based on syntax and meaning, and instead deploy machine learning techniques in order to build a pure prediction system which would map speech into the “best match” text.

Already in 1997, the year Dragon launched Dragon NaturallySpeaking, the first program that was capable of voice recognition from natural continuous speech, the momentousness of this was obvious. (Previous software, including the company’s Dragon Dictate, required the speaker to pause between each word.)

I can vouch for all of this, because after developing repetitive strain injury on both arms in 1997, I was an early user. Even with the very slow processors of the time, Dragon NaturallySpeaking was like magic. It would score close to 95% accuracy. I didn’t know at the time that I would be a user of their products for the next 30 years. But what I knew immediately is that as the processing power of computers improved, I could have access to much more powerful voice recognition software, enabling me to carry out my work despite a continued inability to type.

By 2000, everything was going as I and others had predicted. The software had gotten better and with better computer chips, accuracy and functionality were improving at a fast pace.

What nobody could predict was that during the next two decades, despite advances in the underlying technology, consumers would see little of the gains—thanks to a series of corporate disasters. Dragon was acquired in 2000 by a Belgian company, Lernout & Hauspie, that went bankrupt within months of the purchase, with its co-founders later being found guilty in a corporate fraud case. Dragon was then purchased cheap by ScanSoft—later rebranded as Nuance—which progressively deprioritized the consumer product in favor of medical and enterprise markets. Microsoft acquired Nuance in 2022 and continued the same model. The result: voice recognition software on PCs today is only marginally better than what was available in 2000—far less progress than the underlying technology would have led me to expect.

This disappointing history is one of the reasons why even greater caution is warranted when making predictions about the fast diffusion of AI models today.

Breakthroughs in the “infrastructure” of a new technology need to go hand-in-hand with the right kind of applications built on this infrastructure. And then of course you need the right marketing structure to reach the potential consumer base as well as inputs from consumers to act as the impetus for modifications and improvements of the applications.

This is what we are not seeing for AI. This is what was completely bungled for voice recognition.

But there were also other problems that might foretell AI difficulties. Dragon 16 claimed 99 percent accuracy when it was released but that is under ideal conditions and does not take into account the incompatibilities and difficulties of using Dragon with many programs that users have to contend with daily. It also understates how debilitating an even 1% inaccuracy may be—because that 1% can completely change the meaning, sometimes in embarrassing ways, make your command misunderstood, or may lead to a large amount of time spent on proofreading every email or text to catch difficult-to-detect “speakos”. This is all because the pure machine learning approach will have difficulty with the last mile where there is a choice between words that sound similar or an unusual turn of phrase appears.

Full automation, like full speech recognition, is hard. It requires all the last mile challenges to be resolved, and it necessitates well-designed applications in line with the wishes and needs of end-users. It would also require AI to take over every task in every occupation—including high-stakes judgment and social interaction. We are nowhere near the point in which humans would find equal satisfaction in banter with a machine as with another person.

AI’s Real Potential

But that’s besides the point. Such mimicry should not even be the ultimate aim of AI models.

The word “intelligence” has many meanings. A calculator is intelligent in that it can perform some tasks much better than humans (and many in the early 1900s would have thought that impossible). But it is obvious to everybody that the intelligence of calculators is entirely different from human intelligence.

The same is true of AI.

Humans have a distinct style of learning, based on taking a few examples, forming hypotheses, mentally simulating possibilities, and then engaging in a continuous process of experimentation and trial and error in the real world. Think, for example, of how children learn.

A critical part of that learning is social. Adults and children form their hypotheses by looking at how others (often those presumed to have better knowledge or authority) act. They then hone in those skills by further exchanges, both verbal and otherwise, with people in their community.

Human cognition is also clearly multimodal. We reason with our entire body—many different areas of the brain, the nervous system, our senses, the digestive system and more. We do not just engage in prediction, but inference, deduction, induction, various different forms of verification and recursive thinking. This is why when humans “hallucinate,” it is not a snag but their secret sauce. This is how we simulate different worlds and then decide which ones form valuable options to be explored further and which ones involve huge risks we should watch out for, while irrelevant ones are quickly eliminated (okay that’s a generalization, and some humans do go down the rabbit holes of conspiracy theories and so on).

Humans can quickly learn new things and skills, such as different languages, provided that they have the necessary “learning modules” or mental bauplans. They are, on the other hand, quite bad at sifting through large amounts of information or engaging in abstract reasoning that does not relate to their experiences.

AI models are very different. They learn from massive training data sets and they are excellent at sifting through and recognizing patterns in massive corpora. They learn on the basis of prediction, and once a particular mode of reasoning is mastered, they can apply it, regardless of whether it’s in an abstract or concrete setting. This is where their amazing power lies.

On the other hand, AI models do not currently have capabilities similar to real-time social learning or trial and error. When they hallucinate, it’s a disaster, because they do not have the human filters that eliminate unrealistic simulations and turn the ones with potential into useful cases to be watched for.

When two things are different, it would be a fool’s errand to use one to mimic the other. Modern computing was revolutionized by the design and usability of Macintosh computers, which resulted from the collaboration between Steve Jobs and Steve Wozniak. It would have been a colossal mistake for Jobs to try to replace Wozniak’s engineering skills or Wozniak to ignore Jobs’s genius in design and marketing. They succeeded because they collaborated, combining their different skills in order to complement each other. This is what humans and AI can and should do as well.

The case for pro-worker AI

Under the banner of AGI, all major AI developers are putting their effort into getting AI to mimic human skills. A more productive path would be to recognize that AI should work alongside humans, not replace them.

This perspective throws cold water on dreams of full automation. If AI cannot do every task in an occupation, it has to complement the humans who can. The more it enables them—making them more productive, expanding what they can do—the more transformative it becomes.

This is the basis of what I have called “pro-worker AI”: rather than pursuing full-scale automation, build AI that makes humans better at their jobs and creates new skills and expertise for them.

This is the same vision that was articulated 65 years ago by JCR Licklider, who deserves the accolade of “grandfather of the Internet”. He wrote in 1960 : “The hope is that, in not too many years, human brains and computing machines will be coupled together very tightly, and that the resulting partnership will think as no human brain has ever thought and process data in a way not approached by the information-handling machines we know today.” What he was foreseeing then is what we need from AI models now.

Pro-worker AI is a double whammy for society. It will likely have better productivity outcomes than attempts at full automation. Moreover, while automation displaces workers, leads to job losses and creates inequalities (and via these processes, endangers not just societal cohesion but a democratic future), pro-worker AI leverages and amplifies human capabilities.

You might think pro-worker AI is like pie in the sky—something better, but even harder and less likely than the industry’s dreams of automation. It isn’t.

Pro-worker AI is within the frontier of existing models and can be developed easily, as David Autor, Simon Johnson and I have argued recently.

Take teaching as an example. Although the majority of ed-AI applications have focused on automation (taking over some of the tasks teachers perform), it is beyond unrealistic to think that we could have full automation of teaching anytime soon. Pro-worker AI is entirely feasible and could spearhead a transformative change in education. Rather than seeking to eliminate teachers, AI tools could identify which part of curricular material different students are having problems with, and then make recommendations for reorganizing the classroom into smaller groups where the original teacher, perhaps with the support of additional educators, can target the weaknesses and strengths of each subgroup. Individualized education programs are very helpful to students who have fallen behind or are facing learning challenges. But until now they have been prohibitively expensive. AI could change that.

Why is this pro-worker or pro-teacher? Simple: rather than seeking to take over the teacher’s tasks, the tool would provide new information, and teachers could do things they never could.

How do we know that this is feasible? For one, because my research team and I have built prototypes of this technology and are evaluating it. But you don’t need to take our word for it. What is needed for such a tool is a medium-size language model trained on the relevant curriculum and high-quality data of skilled educators dealing with the vast range of difficulties that students have with the curricular material. All of this is within the frontier.

If, instead of running after AGI dreams, we prioritize helping workers acquire better information for better problem-solving across different tasks and create new tasks and capabilities for humans, AI could have much more positive outcomes.

But this then leads to the next question.

Why this preoccupation with automation?

Four distinct forces have made automation the main focus of both the tech industry and corporations. The first three are economic, while the last one is ideological.

First, the existing business models of large tech developers are focused on selling software to corporations and raising revenues from digital advertisements. These business models have not engaged with creating pro-worker tools. Many of the existing companies are pursuing these same business models, while the business of complementing humans remains untested.

Second, the competition landscape does not encourage new players to come and try these new business models either. The tech industry has become very concentrated, and it is increasingly difficult for new players to outcompete the established behemoths. Worse, these incumbents have monopolized key resources—capital investments, data and essential human resources including engineers and innovators. If a scrappy upstart company were to break through, it will most likely be acquired by tech giants. This means the same business models ruling supreme.

Third, there is a “nuts and bolts” problem. Nuts are not useful without bolts, and bolts are useless without nuts. If you don’t think somebody else will produce nuts, you shouldn’t be investing in bolts, and vice versa. We have a similar situation in that businesses reckon that they will only be offered new automation tools by tech companies, and they plan accordingly. Tech companies think that businesses will continue to predominantly demand automation tools and there wouldn’t be a large market for pro-worker AI, and they invest and develop their models accordingly.

Fourth, and as powerfully, the entire industry is taken over by the idea of super-human machines and AGI. This thinking goes back to foundational figures such as Alan Turing who launched the modern field of computer science and laid the foundations of AI by reasoning about how machines could think like humans. His iconic 1950 paper starts with the now famous “imitation game.” This vision, together with a heavy dose of science fiction literature and movies, conditioned the mindset of many leading figures in today’s industry, making AGI the dominant paradigm.

Even before AGI, “reaching human parity” had become the main metric of success across all sorts of applications throughout Silicon Valley. All of this became exacerbated by the fact that many tech leaders came to view humans as fallible and error-prone. They repeatedly underappreciated human skills and expertise. It was therefore natural for the industry to go down the direction of full-scale automation, without paying much attention to complementing humans.

Can we steer AI on a better path?

We can make the right choices: choosing to complement human skills and contribute to the flourishing of humans rather than to their material, social and spiritual impoverishment.

But this requires a specific focus in designing AI models, and a new philosophical vision.

The good news is that such a path is possible.

The bad news is that this is not where we are heading. In fact, the way the industry is building and conceptualizing AI is likely to lead us not just astray, but somewhere worse.

The most important tool for steering AI onto a better path is not a silver bullet policy: there is none. It is a change of perspective.

If tomorrow we wake up and a significant fraction of the very creative programmers, developers and engineers in the industry made pro-worker AI their priority, we would get plenty of human-complementary, pro-worker AI.

The problem is changing the current preoccupation with automation.

The first and most important step is to change the conversation away from AGI and toward the pro-worker possibilities; away from doomsday scenarios of human extinction and toward the real and present danger that on our current path we will likely amplify inequalities, contribute to massive job losses and not even get the promised productivity benefits.

Such a change in conversation can also help activate the democratic process so that the AI promise is fulfilled.

There are of course specific policy tools that can help as well.

To redirect AI away from automation, we need to accomplish four things at once: create corporate demand for pro-worker tools; generate demonstration effects that prove the approach works; build a market environment where new ideas and business models can flourish; and invest in the infrastructure for pro-worker AI.

The corporate sector: It will demand more pro-worker tools when it focuses on increasing productivity and innovation, rather than only labor cost-saving. This doesn’t just require a change of perspective, but also the removal of artificial barriers that encourage short-termist cost-saving. The biggest distortion is the tax system. In the United States, we tax labor income much more than we tax capital income—on average with an effective marginal tax rate on labor of more than 25% and one on capital that’s close to zero. This translates into a subsidy for using capital tools rather than labor and discourages firms from hiring and training their employees. The situation is similar in several other advanced economies. Treating capital and labor income symmetrically would be an important first step towards the market valuing human-complementary technologies.

Demonstration effects: It is hard to kickstart investment in an area that has long been neglected. This was the case for renewables, after almost no investment in these technologies for decades. The energy transition was helped greatly not just by carbon taxes but also subsidies, competitions and encouragement for clean technology innovations, spearheading demonstration effects. The same needs to be done in the area of pro-worker AI. Even modest government programs that yield results in terms of workable technologies that complement human skills would go a long way. These subsidies must not become a boondoggle for tech firms or another bureaucratic nightmare. What’s needed are lean and simple programs that encourage experimentation with pro-worker AI.

Room for new technologies and business models: The tech sector has never been as concentrated as it is today, and humanity has never seen such powerful and large corporations as the leading tech companies. The largest seven tech giants (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla) now account for 60% of the NASDAQ. Such concentration isn’t just bad for the prices consumers face, but it is damaging for innovation. There are plenty of tech startups in the United States and elsewhere, but most of them build their technologies with the hope of being bought up by one of the major tech companies. Worse, many believe that anyone who steps out of the ecosystem will struggle to compete. The tech sector needs genuine competition. An important part of this is antitrust, which has all but been neglected in the United States. Antitrust statutes have to be updated for the age of AI, but more importantly, they need to be enforced.

Creating room for new business models may also require controlling the worst excesses of the tech industry—especially those that have been laid bare in the path taken by social media over the last 15 years. An additional tool that could play a useful role is a digital advertisement tax, so that new technologies and entrants that experiment with alternative platforms and methods of monetizing AI are not crushed by giants that make their money from digital advertisements.

Infrastructure for pro-worker AI: AI needs data, but what kind of data? For large foundation models any data is useful, because it enables them to better approximate all of the variegated dimensions of human speech and communication. For pro-worker AI, high quality data, where the best experts in the world deal with complex problems, edge cases and new challenges are crucial. If you want to build a tool that will enable electricians to perform newer and more complex tasks, then that tool needs the highest quality data of seasoned electricians dealing with multitudes of challenging problems. This will not be possible unless we have a well-functioning data market, where people can create, control and sell data. That means an end to the ability of tech companies to use data without data-creators’ permission and without compensation, and more importantly, it requires the legal infrastructure to encourage and protect experts to use their time in order to produce useful, high-quality data. A properly functioning data economy is key for the future of AI and especially for pro-worker AI.

Data markets would also facilitate challenges from new entrants against incumbents, since they would limit the anti-competitive effects of having built a large war chest of data.

The future of AI is not written in stone. But rewriting it in a more pro-worker, socially-beneficial way requires us to temper our enthusiasm for superintelligent machines and AGI. We need to recognize that a better future lies in cultivating human skills, not sidelining them.

This is where we need human agency most. That will not come from AI models or those that are most convinced that the future of AI is inevitably sidelining humans.

Only human ingenuity, open-mindedness and our ability to simulate different worlds and act on them can put AI—and our entire democratic system—onto a better path.