Dandan Zhang on AI and Jobs: Neither Pause nor Panic

Leading Chinese labour economist argues that the policy priority should not be to halt innovation, but to monitor labour-market risks, expand retraining, and strengthen safety nets.

Generative AI is already reshaping how people think about work. Its impact is broader and faster than past technologies, affecting not only routine tasks but also cognitive work and parts of the white-collar labour market. Yet the evidence so far suggests that the most dramatic effects on employment are still more potential than fully realised. That gap between what AI could do and what it is already doing gives society a narrow but valuable window to prepare.

This is the argument made by Dandan Zhang, a Professor in Economics and Deputy Dean at the National School of Development (NSD), and Deputy Dean of the Institute of South-South Cooperation and Development, Peking University.

One of China’s leading labour economists, Zhang argues that in today’s global race for technological advantage, no country is likely to slow its investment in AI. The real policy challenge, then, is not to hold back technological progress, but to make the transition as employment-friendly as possible: identifying the skills most at risk, adapting education and retraining, and strengthening safety nets for workers who may otherwise be left behind.

Professor Zhang delivered the speech for TAIXUE, a TED-like initiative by New Economist Think Tank. The video was uploaded on 10 June 2026. Zhang has kindly authorised and reviewed the translation.

— Yuxuan Jia

Hello everyone, I am Dandan Zhang, an economics professor at Peking University. I specialise in labour economics research. It is a great honour to be here at Taixue today to share some thoughts with all of you. In this talk, I will be exploring how artificial intelligence is reshaping the world of work.

On 30 November 2022, a large language model (LLM), namely ChatGPT, was officially released. Its arrival was so sudden that it felt almost otherworldly, as if something from beyond our world had suddenly landed among us. Since then, our generation has found itself standing right at the centre of a massive transformation.

Our jobs may be the first thing it touches. And this is one of the great uncertainties of our time: what will artificial intelligence mean for the employment prospects of our generation?

Today, I would like to examine this question through the lens of labour economics. Drawing also on my own research in recent years, I will share how economists understand the ways in which AI is reshaping employment.

The first fundamental economic question we need to answer is: does technological progress inevitably lead to unemployment? To answer this question, economists often begin by looking back. We ask: what impact did past waves of technological progress have on employment?

Let’s look at what has happened over the past 100 years. The world has experienced several technological revolutions, including electrification, computerisation, and, in the past decade or two, automation. All of these technological advances sparked job anxieties to varying degrees. Yet in every case, they all seemed to reach a state of employment equilibrium.

One study found that around 60 per cent of today’s jobs did not exist in 1940. What does this tell us? It tells us that technology does not simply destroy jobs. It also creates them. As old jobs disappear, new ones continue to emerge. For example, with the information revolution and technological progress in computing, computers replaced typists and some basic text-based work, but they also gave rise to the entire IT industry and millions of new jobs.

So the critical question is whether technological progress can create jobs at a pace that keeps up with the jobs it replaces. In other words, replacement itself is inevitable, and it is not something we need to fear in itself. What we really need to examine is whether, in the process of technological progress, there is a sufficient window of time for new jobs to be created.

Historically, the disappearance of old jobs has typically unfolded over a very long period of time, possibly several decades. That gradual transition gave society an essential window of time for new occupations to emerge and develop.

Next, I would like to summarise two key ways in which this wave of technological progress differs from previous ones.

First, artificial intelligence is not replacing physical labour; it is replacing mental labour. Previous waves of technological progress primarily displaced repetitive, operational work on assembly lines. In other words, replacing physical labour. This time, AI is acting on human cognitive abilities, such as comprehension, reasoning, creativity, and judgment. These are precisely the core competitive advantages of knowledge workers.

This year alone, Claude has dealt a significant blow to the programming profession. Models like Sora and Seedance also had a disruptive impact on video editing.

What deserves even more attention, or perhaps even caution, is a more recent change: AI is no longer only capable of “thinking”. It is now moving towards having both a “brain” and “hands”.

For example, OpenClaw, the personal AI assistant that has recently become very popular on social media, can actually get things done. Large AI models are moving from simply “talking” to being able to roll up their sleeves and help us complete many concrete tasks.

So this wave of technological progress is not only replacing mental work. It is also beginning to replace hands-on execution. This is the first key difference.

The second key difference is that AI technology can be deployed very quickly, and the cost of executing tasks with AI is extremely low. Compared to previous waves of technological progress, generative AI operates on a completely different cost structure. Take intelligent manufacturing, for example: deploying it is a very tedious process. First, you need to remodel the factory, order a bunch of different equipment, and fully reintegrate the production lines, all of which demands massive fixed-asset investments.

In the past, the deployment of new technology took time. It moved forward gradually, industry by industry, or region by region. Equipment had to be purchased unit by unit. Production lines had to be built one by one. As a result, the replacement of jobs also happened gradually and within a relatively limited scope.

This created time. But this time, generative AI is different. Through cloud services, APIs, and open-source models, both enterprises and individuals can access and deploy AI at a very low cost. In particular, with the expansion of many open-source models, the capabilities of large models can now be called upon almost instantly.

The cost of local deployment is also falling as models continue to iterate and as their use becomes more widespread. For example, I recently saw a remark by the renowned mathematician Terence Tao. He said that “AI has basically driven the cost of idea generation down to almost zero.” This means that the cost of a lot of knowledge-based tasks — for example, editing, copywriting, or contract reviewing — is dropping drastically.

So if you were a business owner, obviously you would want to use this new technology to replace relatively expensive human labour if the cost to implement this technology is low. The low cost incentivises businesses to replace human labour with technology.

So, to summarise, this wave of AI-driven technological progress is no longer just about replacing physical labour. It is replacing both mental work and hands-on execution. On top of that, compared to previous waves of technological progress, it carries significantly lower deployment and usage costs. The spread of this technology no longer has to unfold gradually from region to region or industry by industry; instead, a single model can emerge and take effect globally and simultaneously. It may affect an entire technological layer, or an entire category of occupations, all at once. And that is precisely what may shrink the window of time available for new jobs to be created.

Since this impact is inevitable, let’s take a look at how economists believe what kind of effects AI would have on employment. I will introduce three very frontier research methods. Through these different methodologies, we can see how economists approach this question.

The first method is called the AI Exposure Index analysis. In the AI era, the way economists look at employment, occupations, and industries is very different. We would deconstruct occupations, stripping them down into individual tasks or specific skills, so that we can analyse them along different dimensions.

If we want to understand how an occupation will be reshaped, we need to analyse it at a very micro, almost atomic level. Tasks are the basic atoms of a job. We need to define which atoms make up the jobs in today’s labour market.

For this, we need something like the periodic table of elements in chemistry. In labour economics, we do have something similar. For example, O*NET and other international databases have compiled tasks into a kind of dictionary. In total, there are around 20,000 tasks that make up more than 1,000 occupations in today’s labour market. Every occupation is deconstructed. Take a university professor, for example, whose role would be broken down into a specific set of work tasks.

The next step is to evaluate, among these 20,000 tasks, how likely it is that a LLM can complete or accelerate each task.

More specifically, we ask whether AI can save more than 50 per cent of the time required for that task. If it can, we label that task as 1, meaning it is exposed. If it cannot, we label it as 0, meaning it is not exposed. At this point, every task within an occupation has been labelled, with each task carrying either a 0 or a 1 — no exposure or exposed, respectively. These task-level scores are then aggregated up to the occupation level, giving a calculated exposure score for each occupation.

The data used in this kind of analysis mainly comes from recruitment data. We look at the requirements in new job postings: what tasks do these new positions require, and what responsibilities do they list? In this sense, we are using flow information from the labour market.

Our team has also applied this method to research on China. We used more than one million job postings from Zhaopin to calculate AI exposure levels for occupations in China.

Our research found that the occupations with higher exposure — that is, those more likely to be affected by AI — are mainly white-collar jobs. This is basically consistent with international findings. Accountants, editors, salespeople and programmers, in particular, may be among the more significantly affected occupations.

Occupations with lower exposure are mainly blue-collar jobs, such as food service workers, industrial workers and domestic workers. So, in general, high exposure is concentrated among white-collar workers, while low exposure is concentrated among blue-collar workers. Again, this is also broadly consistent with findings from other countries.

Over the past few years, some occupations with high exposure have indeed seen a decline in labour demand. This suggests that high exposure may bring a certain displacement effect. However, although the exposure index method is widely used, it also faces challenges and still needs to be improved. One major issue is that it actually measures potential exposure. But exposure does not equal actual replacement, just as standing in the sun does not necessarily mean you will get sunburned. The risk exists, but it is still a potential risk.

So the exposure index does not measure actual replacement. It measures theoretical substitution. It tells us whether AI can complete a certain task, not whether AI will replace a particular job. That’s why we can’t jump to a completely negative conclusion — even if the exposure is high, we shouldn’t assume the worst.

As I mentioned earlier, our research does show a negative relationship between AI exposure and labour demand. In other words, the higher the AI exposure, the lower the demand tends to be. However, using data from China, Singapore, and the United States, we found that the exposure levels for many occupations are concentrated between 0.7 and 0.8. Although the overall relationship is negative, the picture becomes more nuanced within this range. At the same level of exposure, demand falls for some occupations, remains unchanged for some, and rises for others.

For example, computer-related jobs, which are highly exposed, have seen a decline in demand. Yet, other roles with the same level of exposure see an increase in demand, such as retail sales positions. Meanwhile, some remain unchanged, like business operations. So even at the exact same level of exposure, if we look closely at different occupations, demand is falling for some, staying flat for others, and rising for others.

This raises an important point: at the same level of exposure, not all occupations will be replaced. This shows that the exposure index does not automatically point to the inevitable replacement of an occupation.

So what exactly is going on here? One main reason is that, as I said earlier, a job is made up of tasks. If AI replaces only some of these tasks, or completes part of the work, then human workers can redirect their attention to the remaining tasks that AI cannot do. They can spend more time on the parts of the job where human contribution is still needed.

The other reason has to do with what the O-ring theory, which looks at the relationships among the tasks that make up a job. If the tasks within a job are in an O-Ring relationship — meaning the different tasks are complementary — then focusing more effort on certain tasks can greatly increase output and improve labour productivity. But if the tasks are parallel, meaning that they can substitute for one another, then AI may lead to unemployment.

So, ultimately, whether an LLM will actually replace a job depends on two things: the combination of tasks within that job, and the nature of the relationships among those tasks. Occupations may have similar levels of exposure, but their actual likelihood of being replaced may be very different. This depends on the two factors I have just described.

Therefore, looking at exposure level alone is not enough. We also need to determine, at the occupational level, whether AI’s impact is pushing in a positive or negative direction. The goal is to give the exposure index more depth and nuance — one that can actually tell us whether displacement will eventually occur. That is what our team is currently working on, refining the methodology behind the exposure index.

Now let me introduce the second method, which we might call AI job-integration analysis. The previous method looks at the potential, or theoretical, impact of AI. This second approach looks instead at real-world implementation at the firm level.

This line of research comes from a 2025 NBR working paper by two PhD students from Harvard University. The paper, titled Generative AI as Seniority-Biased Technological Change, has been widely cited. It argues that generative AI does not favour high-skill workers, but rather those with seniority.

The authors analyse the full text of firms’ recruitment advertisements over the past several years. They identify whether a firm is hiring for roles specifically responsible for implementing or integrating AI technologies into the company’s workflow. If a firm posts this kind of job opening on a given day, the authors treat it as evidence that the firm has already started using AI and is using it in a relatively deep way.

So, as you can see, this method follows a completely different logic from the previous one. It no longer uses theory to infer whether AI may affect certain jobs. Instead, it directly observes firms’ actual behaviour: are they hiring people to bring AI into their workflows? And the findings are very interesting.

The authors found that generative AI represents a form of seniority-biased technological change. Among firms that had adopted AI, hiring for junior positions declined significantly, while hiring for senior positions remained stable, and in some cases even increased. In other words, the impact of AI appears to fall most heavily on those at the lower end of the career ladder, especially entry-level workers.

The authors also found that the main effect was a slowdown in new hiring, rather than layoffs or resignations. In other words, firms were not necessarily cutting existing workers; they were hiring fewer new ones.

Furthermore, when looking across different education levels, from elite universities to mid-skilled technical colleges and low-skilled institutions, the authors found that mid-skilled technical graduates bear the heaviest impact, while graduates from top-tier universities and lower-skilled workers remain largely unaffected by the slowdown.

This research offers an important insight: AI may not first affect the existing employed population in the labour market — what we might call the stock of employment. Instead, it may gradually compress labour demand by reducing the number of new positions. There may still be a window between potential exposure and actual substitution. AI may first reduce the entry of new workers into the labour market, while the existing stock of employed workers may, at least for now, be less affected.

This method also has limitations. It has very high data requirements. It needs access to a company’s complete hiring history to determine whether AI is genuinely being used.

The third method I am going to introduce is based on actual AI usage data. Rather than theorising about what AI can do, which is what the exposure index approach does, it directly observes how people interact with LLMs, and through that interaction data, draws conclusions about what people are actually using AI tools for.

This is a very recent and widely circulated study. Many of you may have seen it circulated on WeChat official accounts. It was released by Anthropic, the company behind Claude, in an employment report published earlier this year. In this report, Anthropic introduced an important concept: the observed exposure.

The first method I introduced measures theoretical exposure. This third method measures observed exposure. The data used in this research comes from actual conversations between users and Claude. The researchers looked at how users interact with Claude every day and what tasks they use Claude to complete. They then identified the specific things users were doing when they used a LLM and extracted those tasks for analysis.

The next step was to examine whether the LLM directly completed those tasks, or whether it helped humans complete them. If the model completed the task directly, this was treated as substitution. If the model helped the person complete the task and enhanced their capability, this was treated as augmentation.

Their findings are very interesting. They found that, so far, actual replacement has not really occurred on a large scale. What we mainly see is complementarity. LLMs are mainly improving our work efficiency, rather than producing obvious substitution.

They also compared observed exposure with theoretical exposure, which is what the first method captures. They found that theoretical exposure is much higher than actual observed exposure. For example, in computer and mathematics-related occupations, theoretical exposure suggests that more than 94 per cent of tasks could be completed by AI. But in actual Claude usage, Claude only covers about one third of these tasks. The difference between more than 90 per cent and just over 30 per cent is very large.

Why can AI theoretically do more than 90 per cent of the tasks, while in practice it only does a little over 30 per cent? There are many reasons for this gap.

One reason is that model capabilities are still limited. Another is legal and compliance constraints. In many cases, when models use certain types of data, human verification is still needed. There are also barriers related to software integration. So there is still a gap between actual use and theoretical substitution.

Using the observed exposure index, the researchers found that 57 per cent of tasks were augmentation-type tasks, while 43 per cent were substitution-type tasks. So, overall, augmentation remains the main pattern.

Augmentation-type use improves productivity, but it does not reduce employment. Substitution-type use is what directly threatens employment. So their findings are somewhat reassuring: AI is making us better at what we do, while not yet taking away our jobs on a large scale.

That said, this method has its limitations. It relies on a single LLM — Claude — and the range of tasks it covers is quite narrow. As mentioned earlier, the full task dictionary contains 20,000 tasks, yet this study only captures 800 of them as they appear in Claude interactions. The analysis is therefore partial and incomplete. Nevertheless, the study introduces a valuable new perspective: by looking directly at how people actually interact with AI, it suggests that the reality may not be nearly as alarming as it sounds in theory.

I’ve now finished introducing these three methods, so let’s do a quick comparison. These three methods actually complement each other. They seem to capture three different stages of AI’s impact: from theory, to firms, to people. The exposure index, the first method, mainly looks at the potential exposure, or the theoretical exposure. It relies on flow data, especially new recruitment information.

The second and third methods both look at actual adoption, drawing on existing employment data to understand how much of the current employment structure may be affected. These methods can complement one another, and I believe only by combining them can we get a more comprehensive and more precise analysis.

So this is how economists are now analysing the question of how LLMs affect employment. These methods are all at the frontier of current research. Our own research team is also trying to apply these methods to study China.

The first point I want to emphasise is that we now need to build a dynamic monitoring and early-warning system. This is really important. We need to track, in real time, how the tasks and skills contained in today’s jobs are changing. Which skills are needed in this new era? Which skills may be phased out? This lets us spot the risk signals and take action well before any changes start showing up in the employment data for the existing workforce.

Second, policymakers need to develop education and training programs based on the monitoring results. People need to know quickly which abilities are needed and which skills are becoming more important, so that they can fill their own skill gaps as soon as possible.

Here, I want to emphasise individual initiative and creativity. Through our research and conversations with Xianyu [the Chinese equivalent of eBay], we found that new, user-created forms of work emerge on the platform every year. Some are quite unconventional: people offer to argue with someone on your behalf, retouch photos with AI, bathe pets, and provide all kinds of niche services. These examples show just how extremely creative individuals can be.

So I think the first thing we need to do is send clear signals to the market. We need to allow individuals to face and embrace the challenges brought by AI, to reinvent themselves, and to create more employment opportunities for themselves.

But creativity also needs to be guided by information. The hope is that these early warning signals can help each person make better career choices, shorten the time it takes to retrain for new work, and lower the cost of trial and error during this transition.

The third point is about safety-net policies. First of all, technological progress cannot be stopped. So while AI is getting smarter and smarter, policymakers need to develop much more robust risk management strategies during this transition period, particularly creating strong social safety net policies.

In the current environment of international competition, no country is going to slow down its investment in AI or the pace of AI development. So the key policy question, I think, is not about preventing the use or the development of AI. On the contrary, technological progress should be encouraged. At the same time, policymakers need to prevent a situation where workers, who cannot get through this short-term adjustment period, end up permanently locked out of the new employment landscape.

For instance, at this year’s Two Sessions [annual plenary sessions of the National People's Congress and of the National Committee of the Chinese People's Political Consultative Conference] we saw the proposal to build an employment-friendly society. This points to an important general direction. What that means is that going forward, as more firms integrate AI into their workflows, policymakers need to assess whether the employment landscape is actually employment-friendly. This involves a whole range of policy measures, and I believe it can help us open up a real window for employment transition — one where, even as some jobs are being replaced, more new opportunities are being created at the same time.

Finally, let me take some time to summarise today’s talk. The arrival of generative AI is reshaping our labour market. According to the main studies so far, it mainly affects white-collar jobs, entry-level positions and some mid-skill workers. AI may improve work efficiency, or it may bring substitution. The final trajectory of each occupation and each job is still unclear. We cannot say that any particular occupation will definitely disappear.

This process depends, to a large extent, on the initiative of each worker: whether they are willing to adapt and change. It also depends on the way the tasks and skills within each job are combined.

One last point I want to make is that, at present, what we are seeing is still mostly potential impact. The real shock has not yet fully arrived. This means that we are still in a very precious window of opportunity. The question is whether we can make the most of this window and prepare ourselves well. For example, can policy analysis stay a step ahead of each new wave of technology, so it can send the market clearer, more accurate signals and make the right adjustments? That is a shared challenge facing both researchers and policymakers, and it’s what my team and I are working towards. That’s everything I wanted to share with you today. Thank you, everyone.