Introduction

Generative AI (GenAI) and large language models are spreading swiftly through organisations and workflows that depend on coding, content creation, customer service, and analytical tasks. This has created widespread debate over whether GenAI will augment or displace workers. Based on the Occupational Informational Network (ONET) taxonomy, Eloundou et al. (2023) estimate that about 19 percent of workers in the United States could find at least half their tasks affected by large language models. Although occupations and their tasks will invariably shift, some view this as early warning signs of potential job impacts from GenAI.

These concerns have been corroborated with recent evidence from academic studies, at least for certain occupations and demographics.  Brynjolfsson, Chandar, and Chen (2025) demonstrate that GenAI adoption is associated with notable, seniority-biased changes in the labour market. Their research, drawn from payroll data covering roughly a fifth of U.S. employment at firms with fifty or more employees, reveals that in occupations exposed to AI, there is a striking reduction in the headcount for younger workers .  Compared to the fourth quarter of 2022—just before the launch of ChatGPT—employment of workers under 25 in roles such as customer service and software development is nearly 20 percent lower, while those aged 30 to 45 see only modest increases. Similarly, Lichtinger and Hosseini Maasoum (2025), analyzing LinkedIn data from 62 million résumés across 285,000 firms, confirm these trends: companies recruiting “GenAI integrators” have seen substantial declines in junior staff, while their senior ranks have remained steady or even grown.

These studies are among the earliest to shed light on the effects of AI in the labor market, prompting an urgent question: what is happening globally, and in particular, how is Hong Kong faring? This is especially important for international financial hubs like Hong Kong, where differences in economic structure, regulation, and labour market realities may produce outcomes distinct from those seen in Western economies.

To address this, this chapter brings together several data sources to examine how Hong Kong ranks in terms of GenAI exposure, adoption, and capability. We utilise LinkedIn data via Revelio Labs, patent records from the USPTO, and academic publication trends from OpenAlex, and digitalisation indicators from BuiltWith. While none of these sources is perfect—and the AI landscape itself is evolving rapidly—their combined breadth offers the most complete view yet of Hong Kong’s labour market and AI activity.

Based on measures constructed upon the above mentioned databases, our results indicate that Hong Kong’s occupational structure leaves it highly exposed to GenAI, with a large share of its workforce in roles identified by existing research as vulnerable to AI. Nonetheless, observable AI adoption and innovation—measured by “GenAI integrator” hiring in job postings, patents, and research output—are mostly concentrated among a select group of firms, many of which are multinationals. In general, Hong Kong trails leading AI centers like the US and Singapore in these areas. Drawing on LinkedIn-based worker histories, we have constructed time-series panels of workforce composition for Hong Kong, organised according to standard GenAI exposure metrics. This enables us to track trends in employment in AI-exposed roles as firms begin advertising for AI-related jobs, and to compare these shifts to those in peer economies. While the mix of occupations in Hong Kong suggests a high level of AI exposure—on par with other advanced economies—the actual employment adjustments in exposed fields have been less dramatic, likely complicated by factors such as ongoing COVID-19 disruptions, episodes of social unrest, and migration patterns.

This paper also speaks to the expanding literature on AI’s labour market effects. For instance, Humlum and Vestergaard (2025) find that, in Denmark, the impact has so far been mild, possibly because of institutions, labour regulations, or the presence of complementary skills that dampen AI’s potential for disruption. Similarly, research by Klein Teeselink (2025) in the UK illustrates that although some jobs are displaced, new tasks and complementarities often offset such shifts. 

Finally, we draw out several policy implications relevant to Hong Kong and similar economies. In particular, we highlight the need to adapt education, skills policy, data infrastructure, and labour-market institutions, so that the productivity gains of GenAI can be realised without unduly harming certain groups or amplifying transitional costs.

Labour market and AI in Hong Kong

So far, the labour market impact of AI in Hong Kong seems understudied. The Hong Kong SAR government does not produce high quality longitudinal occupation level data available to academic researchers. Thus, for labour market data, we use Revelio Labs’ harmonised workforce data which is assembled from LinkedIn. They pull regularly updated worker histories and augment the data with job titles, job descriptions, start/end dates, occupation/title taxonomies, estimated seniority based on job title, skills, and geographies, as well as educational history and demographics. This harmonisation allows us to create firm-by-time panels of workforce composition for Hong Kong and to align roles with GenAI exposure categories used in the literature. These ingredients enable the construction of time-varying indicators of occupational mix, junior-versus-senior shares, and exposure-linked groupings, which we can then examine around external adoption signals (e.g., text-identified adoption in job postings or firm communications).

There are a few caveats. First, Hong Kong’s coverage of LinkedIn could be substantially different. For example, the platform may be underused by Chinese firms as its adoption or presence in the mainland varied over time. Second, the composition of the Hong Kong labour force could be quite different. For example, although heavily exposed, Hong Kong has a high proportion of finance or managerial talents which are harder to displace, or features older or more elite trends (the fair comparison perhaps being New York). Meanwhile, much of the impact of GenAI is on the less elite and younger populations. Third, adoption rates and depth by Hong Kong firms could be slower. However, we do not think that the biases in our sample data should be correlated with the emergence of ChatGPT nor specific to GenAI-affected occupations.

For measures of AI exposure and adoption (i.e., integration), there are two data sources: (1) general occupation-level exposure indices, which assume implicitly that all firms adopt AI equally.  This popular metric for occupation-level impact is produced by Eloundou et al (2025).[1] Second, there are firm-level signals of AI adoption, revealed via job postings. Specifically, Lichtinger and Hosseini (2025) implement a classifier of “integrators” or what we call “adopters” of LLMs who seek to integrate AI technologies into their workflows. 

Turning to the first measure, we calculate the fraction of highly exposed jobs in Table 1. Exposed may mean augmented or displaced, but in general means affected. Examples of such jobs include financial analysts, customer service workers and software developers. Hong Kong does not rank particularly high among economies in terms of its share of high exposure based on this definition.

Table 1. Fraction of highly exposed jobs across markets

Next, we plot the 2016 normalised growth of exposed vs non-exposed occupations. We map AI exposure using GPT-4 ratings from Eloundou et al. (2023) at the O*NET SOC occupation level. Occupations are classified into exposure buckets based on percentile rankings: low exposure (0-25th percentile), medium exposure (25-75th percentile), and high exposure (75-100th percentile). This categorical approach captures non-linear effects while maintaining interpretability. The high exposure group corresponds to about 20% of all jobs in Hong Kong in Q4 of 2022. We plot semi-annually and the US and Singapore alongside for comparison. For Singapore and US, there is (1) a sharp growth in GenAI affected youth coincident with the release of ChatGPT. However, (2) while Hong Kong may or may not have had a decline, there seems to be an effect arising from the COVID-19 period. This could be due to restrictions or social protests leading to emigration. Thus, unlike in Singapore and US, there appears to be pre-trends, making the labour market effect of AI somewhat murkier. We attempt to later parse this out with firm-level analysis, whereby we exploit variation across firms as opposed to variation across occupations. However, the conclusion is clear – across advanced economies, there seems to be an effect from the rise of LLMs.

Figure 1. Growth rates of workers, by high v.s. low AI exposure

We now perform a regression analysis at the firm level to exploit firm level variation in AI exposure to better understand whether AI may have had an effect. This is because in aggregate, Hong Kong already seemed to exhibit a potential pre-trend in job decline. For our firm level sample, we include US firms as a comparison point to see if AI effects were larger or smaller in Hong Kong, giving us over 400,000 firms.

To isolate the effect of AI, we do three things. First, we control for the hiring growth in the 2020 -2022 period for each firm. Second, we look at cross-sectional variation in the share of firms pre-ChatGPT employment in vulnerable occupations (low seniority, high LLM impact). We also replicate Lichtinger and Hosseini (2025).[2] Interestingly, by this measure, adoption rates appear higher in Hong Kong. We report summary statistics in the Appendix. 

Under this more scientific specification, we find that in Hong Kong, firms with a higher share of positions vulnerable to the AI shock experience lower headcount growth, especially compared to the US.[3] For example, in model 1, the dependent variable “Headcount growth” is defined as percentage change in headcount after v.s. before the AI shock. So coefficients are in percentage points. For firms in Hong Kong, the marginal effect of Share vulnerable workers on Headcount Growth is the sum of the main effect and the interaction, β(Share vulnerable) + β(InHK× Share vulnerable) = -0.2122 + (-0.3670) = -0.5792. This suggests that, if we compare a firm with 40% workers in positions vulnerable to the AI shock to a firm with only 20% vulnerable workers, the former’s headcount growth is approximately 0.116 percentage point lower (-0.5792 * 0.2). Translating this into a change in the number of jobs, this is approximately –0.6 jobs for a median Hong Kong firm (headcount median pre-ChatGPT is ≈ 527) and –3.9 jobs for a mean-size Hong Kong firm ( ≈ 3,374).

One potential interpretation of the above result is that the labour market effect of AI is worse than in the US. If we consider (1) from summary statistics the average Hong Kong firm does adopt AI more (again, the usage of LinkedIn in Hong Kong may differ from the US), (2) during this period, Hong Kong’s economy faced heavy mobility restrictions, (3) Hong Kong faced sagging Chinese economic conditions, then the point estimates suggest AI may have impacted Hong Kong more than the US. This makes sense as there is evidence to suggest that during bad times, firms restructure their production processes at an accelerated pace (Hershbein and Kahn 2018).

To help visualise the net effect, in the Appendix, we repeat Figure 1 overlaying adopters in exposed occupations/roles and adopters in non-exposed occupations. It appears adopters are slightly more likely to reduce headcount of exposed occupations, but (1) reallocate those headcounts elsewhere, and (2) the effects are likely quite similar across exposed occupations at “non-adopters”.  Overall, the results suggest that on a net basis, adopters shed headcount from exposed occupations slightly faster, but the aggregate effect is close to the effect seen from those who we classify as adopters.

Table 2. Regression analyses: Effects of AI adoption and exposure on the labour market

AI Innovation in Hong Kong

Albeit narrower, another aspect of AI adoption and capability is innovation. Two sources of generally selection-free innovation are available. We also obtain data from various sources, including academic publication on AI from OpenAlex and AI patenting from USPTO. OpenAlex is a comprehensive, open-source database of scholarly works, authors, institutions, and concepts. It serves as a free alternative to proprietary databases like Scopus and Web of Science. For this paper, we use OpenAlex to track the global landscape of AI research, identifying trends in publication output, international collaborations, and the focus of research across different countries and institutions. Its open nature allows for transparent and reproducible analysis of academic AI research.

The US Patent and Trademark Office (USPTO) maintains a dedicated dataset of patents related to AI. This database is curated using machine learning models to identify patents that include AI components. We leverage this data to analyse trends in AI-related innovation, specifically to identify the geographic distribution of inventors and assignees (companies) of AI patents. This provides a view into the commercialisation and corporate R&D side of AI development. We have linked inventors to Revelio through a joint collaboration with Revelio Labs. 

Every year, we adjust the patent or publication based on its cohort-adjusted citations. Higher thresholds or numerical thresholds produce a similar pattern.  In all variations, it seems like relative to its industry application, Hong Kong remains decisively research-first. As shown in Figures 2~4, cohort-adjusted OpenAlex ranks Hong Kong #3 globally in the top-decile AI papers per capita (~173 per million, 2015–2024), but for AI patents per capita, the city sits mid-pack (top 30).

Startups tell the same story, using data from Crunchbase. Crunchbase AI/ML organisations founded in the last 10 years place Hong Kong in the middle of the OECD on per-capita counts, and funded-startup density is similarly modest; capital raised is growing but trails Singapore/US trajectories. In short, Hong Kong over-indexes on publications and under-indexes on commercialisation—closing the translational gap (patents, funded AI/ML venture creation) is still quite important. Israel is a notable contrast, whereby its academic output is far lower than its commercialisation efforts. On both dimensions, the exemplary states are Switzerland, Singapore and US, which produce a great deal of startups as well as academic outputs.

Figure 2. Academic publications on AI

Figure 3. Patents on AI

Figure 4. Comparison of academic research and patents on AI

Conclusions

Due to time constraints, we did not examine the startup ecosystem in Hong Kong. However, our results point to a number of policy implications. First, there appears to be some incremental effect of AI impact in Hong Kong, but the scale seems small. Does this mean that the AI effect here is smaller, and why? Is it that Hong Kong has poor data? We ask policymakers to produce more labour market statistics to track this, especially by occupation level. Whether constructed by agency employees or in collaboration with universities, Hong Kong sits on an inert gold mine of data from its tax filings that contain self-reported occupations. This data can be utilised easily to produce public statistics. With the help of local universities, Hong Kong can easily create a longitudinal dataset that can enable various research efforts and real-time policy monitoring.

Second, while Hong Kong’s strong university system seems to produce world-class research, our research shows there is a big gap in commercialisation. Notable exceptions include SenseTime. There are several potential reasons for this, and may not be specific to AI. Hong Kong has made a proactive effort toward supporting the university system. However, there is a clear decoupling between the publication rate at universities and commercialisation. It is not clear if this implies the universities are not doing well in sharing knowledge with local industry, or there is a lack of innovation talents in local industry, or simply the length of gestation for basic research to translate to application.

It is also not clear that these statistics tell the whole story. As said earlier, the local industry is dominated by finance, and their capture of these technologies may not show up well in the statistics. Another possibility could be that the triumphs of the local industry could spill over to nearby cities such as Shenzhen, with DJI being a prime example.

However, the recent expansion of government initiatives is commendable. Hong Kong has already put in place several co-investment and public venture capital schemes – most notably the Innovation and Technology Venture Fund (ITVF), Hong Kong Science & Technology Parks Venture Fund, and the Cyberport Macro Fund – which operate on a matching basis with private investors. In particular, the ITVF Enhanced Scheme explicitly targets AI and data science. However, the scale of AI-focused capital and the depth of dedicated AI venture expertise remain modest relative to Hong Kong’s research output.

Multiple official and quasi-official reviews characterise Hong Kong’s innovation funding architecture as conservative and administratively onerous. The Advisory Committee on Innovation & Technology report highlights that private R&D investors in Hong Kong are “risk averse in I&T” and that the society is less tolerant of failure than peer economies.  The Innovation, Technology and Industry Bureau and the Financial Services Development Council (2021, 2024) as well as Our Hong Kong Foundation point to under utilised public co-investment vehicles, a lack of risk appetite for earlystage financing, and a need for “more generous, less risk-averse government-led incubation programmes /grant schemes with streamlined administrative procedures.” Policy should therefore focus less on creating co- investment from scratch and more on (i) expanding these vehicles, (ii) sharpening their mandate around AI and other deep tech fields, and (iii) complementing them with true angel-level incentives, which Hong Kong currently lacks. However, policymakers should also pay special attention toward streamlining the administrative burden, coaching bureaucrats to be flexible and open-minded instead of career-concerned and risk-averse (such as pegging utilization as a KPI). Meanwhile, the university can do its part by encouraging entrepreneurship among student founders and faculty and examining whether KPI-driven exercises maximise the social return of its increasingly world-class researchers.

References

1. Revelio Labs assigns seniority scores using an ensemble model that combines three components: (1) current job information (title, company, industry), (2) job history details (previous employment duration and seniority), and (3) individual age. We define junior workers as levels 1-2 and senior workers as levels 4-7, excluding level 3 to maintain clear separation.
2. Our replication appears successful as we find similar results in the US in terms of both the frequency of AI adoption rates as well as its labor market impact. We find similar effects in the US in terms of the impact of AI adoption on young people. However, the effect seems a little bit smaller in Hong Kong, and shows evidence of starting to occur before ChatGPT, echoing the results we find above in aggregate).
3. In unreported results, there seems to be an interaction effect among adopters with high share of vulnerable workers.
4. Abbas Nejad, K., Musillo, G., Wicker, T., & Zaccaria, N. (2025). Labour market signals: The role of large language models. CentER Discussion Paper No. 2025-003. Tilburg University.
5. Bick, A., Blandin, A., & Deming, D. J. (2024). The rapid adoption of generative AI (NBER Working Paper No. 32966). National Bureau of Economic Research.
6. Brynjolfsson, E., Chandar, B., & Chen, J. (2025). Canaries in the coal mine: Early warning effects of generative AI on labour markets. Working paper.
7. Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889–942.
8. Chen, W. X., Srinivasan, S., & Zakerinia, S. (2025). Displacement or complementarity? The labour market impact of generative AI (HBS Working Paper No. 25-039). Harvard Business School.
9. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: An early look at the labour market impact potential of large language models. arXiv preprint arXiv:2303.10130.
10. Gulati, P., Marchetti, A., Puranam, P., & Sevcenko, V. (2025). Generative AI adoption and higher order skills. INSEAD Working Paper.
11. Hershbein, B., & Kahn, L. B. (2018). Do recessions accelerate routine-biased technological change? Evidence from vacancy postings. American Economic Review, 108(7), 1737–1772.
12. Humlum, A., & Vestergaard, E. (2025). Large language models, small labour market effects (NBER Working Paper No. 33777). National Bureau of Economic Research.
13. Hui, X., Reshef, O., & Zhou, L. (2024). The short-term effects of generative artificial intelligence on employment: Evidence from an online labour market. Working Paper.
14. Klein Teeselink, B. (2025). Generative AI and labour market outcomes: Evidence from the United Kingdom. SSRN Working Paper.
15. Labaschin, B., Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2025). Extending “GPTs are GPTs” to firms. AEA Papers and Proceedings, 115, 51–55.
16. Lichtinger, A., & Hosseini Maasoum, M. (2025). Generative AI as seniority-biased technological change: Evidence from US résumé and job posting data. Working paper.
17. Liu, Y., Wang, H., & Yu, S. (2025). Labour demand in the age of generative AI: Early evidence from US job posting data. SSRN Working Paper.
18. Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.
19. Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv preprint arXiv:2302.06590.
20. Wiles, E. (van Inwegen), Munyikwa, Z. T., & Horton, J. J. (2023). Algorithmic writing assistance on jobseekers’ resumes increases hires (NBER Working Paper No. 30886). National Bureau of Economic Research.

Appendix

Table A1. LinkedIn sample characteristics of Hong Kong/ US firms

Note: our notion of a firm is a firm-location (i.e. Goldman Sachs HK is diff from Goldman Sachs US and an employee in the US would contribute to the latter).

Figure A1. Growth rates of workers,

by high v.s. low AI exposure and adopters v.s. non-adopters