A growing number of scholars have warned about social, political, and economic harms associated with the current trajectory of artificial intelligence (AI) (Acemoglu et al. 2023, Frey 2019, Kasy 2022). In The Anxious Generation, for example, social psychologist Jonathan Haidt points to the link between social media use and teenage depression, as large technology companies seek to maximise engagement to amass personal data for targeted advertising. In a similar spirit, Acemoglu and Restrepo (2020) argue that this path results in “the wrong kind of AI”, emphasising the necessity of developing AI technologies that prioritise ethical considerations and user privacy.
In a new paper, we explore whether the trajectory of AI development can be redirected towards less data-intensive methods (Frey et al. 2024). In doing so, we situate our analysis within the literature on directed technological change, which posits that technological innovation shifts away from scarce or costly factors of production (Acemoglu 1998, 2002, Hanlon 2015). Notable examples of this include high wages and labour shortages prompting investment in automation technologies (Habakkuk 1962, Allen 2009, Hornbeck and Naidu 2014), and oil price shocks or carbon taxes driving the development of green technologies (Acemoglu et al. 2012, Hassler et al. 2021).
Building on this logic, we hypothesise that the General Data Protection Regulation (GDPR) raised the costs of storing and processing personal data (Frey and Presidente 2024), incentivising companies to prioritise investments in data-efficient methods over data-intensive approaches, and thereby altered the composition of AI patenting.
Trends in AI patenting
Drawing upon the existing literature and the advice of computer scientists, we introduce a novel framework for evaluating AI technologies based on their data intensity (Russell and Norvig 2016, Goodfellow et al. 2016). AI technologies built on deep-learning methods require large datasets to effectively tune millions of parameters. In contrast, knowledge-based systems depend on structured rules and expert knowledge rather than extensive data. Bayesian methods enhance efficiency by leveraging prior knowledge. Additionally, various techniques have emerged to reduce data requirements: transfer learning (including zero-shot and few-shot learning) repurposes knowledge across tasks, while synthetic data generation creates artificial training examples. For example, DeepMind’s AlphaFold incorporated some of its own predictions into its training data (Jumper et al. 2021). We categorise these methods as ‘data-saving’, in contrast to deep-learning approaches, which we classify as ‘data-intensive’.
Next, we take our taxonomy to the data: we identify relevant patents by conducting keyword searches of patent titles and abstracts in the European Patent Office’s PATSTAT Global database (2024 Spring Edition), analysing them at the patent family level to avoid double-counting the same invention across jurisdictions. Using this dataset, we document several stylised facts about AI patenting trends worldwide.
First, we observe a notable technological shift from data-saving AI to data-intensive deep learning throughout the 2010s. Between 2000 and 2021, the stock of data-intensive AI patents grew at an impressive annual rate of 52%, whereas data-saving AI patents exhibited a more modest growth rate of 19% per year. Interestingly, data-saving patent activity remained below its 2004 level until 2013 but experienced a notable resurgence following the implementation of the GDPR in 2018 (Figure 1).
Figure 1 Average number of AI patent families per applicant
Between 2018 and 2021, transfer-learning patents surged by 185%, while synthetic data generation and Bayesian methods grew by 86% and 68%, respectively, albeit from modest baselines (Figure 2).
Figure 2 Patenting across AI technology classes per year
The potential impact of the regulatory environment on this trend is further underscored by significant geographic disparities in AI patenting activity and technological specialisations. Since the launch of China’s ‘Made in China 2025’ initiative, Chinese AI patenting has surged (Figure 3), with universities and government institutions accounting for 86% and 54% of global AI patents in these categories, respectively, compared to just 3% and 4% in the US (Figure 4). This dominance is particularly evident in data-intensive AI and extends to the private sector, supported by government procurement efforts that generate valuable training data for commercial use (Beraja et al. 2021, Beraja et al. 2023).
Figure 3 AI patent families filed by country
In contrast, US firms lead in data-saving AI innovation, contributing 45% of global data-saving patents. Although the EU lags significantly in overall AI patenting, it has a relatively higher share of indigenous data-saving patents, with 13% of private-sector AI patents compared to just 5% in China.
Figure 4 AI patent families by institution and country
Thus, while data-intensive AI remains dominant across regions, the balance varies: China strongly favours data-intensive approaches, the US maintains a more balanced portfolio, and the EU shows a relative emphasis on data-saving technologies. This, we note, mirrors patterns of data privacy regulation. In China, state procurement policies have explicitly incentivised AI development for data-intensive surveillance applications (Beraja et al. 2021), while in the US, data privacy relies on a patchwork of sector-specific regulations rather than comprehensive federal protection. The EU, in contrast, has pursued a distinctive regulatory approach through its GDPR, emphasising individual privacy protection over data accumulation.
The impact of privacy regulation on AI innovation
Following the literature on directed technological change, we hypothesise that the GDPR, by increasing the cost of storing and processing personal data (Frey and Presidente 2024), has incentivised companies to invest more in data-saving methods and reduce their reliance on data-intensive ones, thereby altering the trajectory of technological change in AI.
For our empirical analysis, we leverage the GDPR’s reach, which affects patent applicants outside the EU if they target EU consumers, to measure how much companies worldwide rely on EU markets. To gauge a firm’s exposure to the GDPR, we use inter-industry linkages from the OECD Inter-Country Input-Output (ICIO) Tables, capturing their exposure to EU markets as outlined by Frey and Presidente (2024). By breaking down the ICIO Tables into intermediate business use and final household consumption, we exclude business-to-business transactions, which are less impacted by the GDPR, allowing us to concentrate on consumer transactions.
We begin by presenting the baseline results using the full sample of patent applicants, which includes firms, individual applicants, public institutions, and universities. Next, we narrow our focus to corporate applicants, examining how the GDPR’s effects differ based on firm characteristics, such as size and age. Taking advantage of the timing of the GDPR introduction, and the varying exposure of firms to this regulation, we find that:
- Patent applicants (including firms, universities, public institutions, and individuals) affected by GDPR have redirected their inventive efforts towards less data-intensive and more data-saving AI approaches.
- The primary drivers of this shift were older and larger companies based in the EU.
- While altering the technological trajectory of AI, the GDPR also reduced overall AI patenting in the EU while amplifying the market dominance of established firms.
Redirecting AI
Over the past decade, the overwhelming focus of AI research has been on data-intensive deep-learning methods – which incentivise companies to amass personal data – often at the expense of exploring less-data-dependent, rule-based systems (Klinger et al. 2020). These systems are also behind the most recent advances in generative AI. However, despite many productive use cases, some scholars argue that the current direction of AI development may reduce overall welfare (Acemoglu and Johnson 2023).
Consider the historical trajectory of electric vehicles. At the beginning of the 20th century, they were as competitive as their gasoline-powered counterparts. However, a lack of investment in electrical infrastructure, along with significant oil discoveries, shifted market dynamics decisively in favour of internal combustion engines (Taalbi and Nielsen 2021). This shift led to an inefficient technological lock-in. Now, a century later, we are witnessing a renewed shift toward electric vehicles as efforts intensify to correct this path-dependent technological trajectory. Imagine the technological trajectory if we had the foresight to tax carbon in the early 1900s. One could make a similar argument for taxing data.
At the same time, the stronger response from established firms indicates that privacy regulations may inadvertently reinforce incumbent advantages while dampening overall innovation – a pattern our study finds evident in the field of AI. Indeed, a growing body of research highlights the negative effects of the GDPR on smaller businesses and innovation, contributing to increased market concentration (Frey and Presidente 2024, Peukert et al. 2022, Johnson et al. 2023). The upcoming EU AI Act could intensify this trend by imposing greater compliance burdens on smaller firms and potentially shifting technological development toward less-data-intensive methods. Its emphasis on explainability poses particular challenges for deep-learning technologies – responsible for most progress in the field over the past decade.
Investigating how the EU AI Act influences both the volume and direction of AI innovation presents a valuable opportunity for future research. What our study shows that it is possible in principle to shape the trajectory of AI development through policy intervention.
References
Acemoglu, D (1998), “Why do new technologies complement skills? Directed technical change and wage inequality”, Quarterly Journal of Economics 113(4): 1055–89.
Acemoglu, D (2002), “Directed technical change”, The Review of Economic Studies 69(4): 781–809.
Acemoglu, D, D Autor, and S Johnson (2023), “How AI can become pro-worker”, VoxEU.org, 4 October.
Acemoglu, D, and P Restrepo (2020), “The wrong kind of AI? Artificial intelligence and the future of labour demand”, Cambridge Journal of Regions, Economy and Society 13(1): 25–35.
Acemoglu, D, P Aghion, L Bursztyn, and D Hemous (2012), “The environment and directed technical change”, American Economic Review 102(1): 131–66.
Allen, R C (2009), The British Industrial Revolution in global perspective, Cambridge University Press.
Beraja, M, A Kao, D Y Yang, and N Yuchtman (2021), “AI-tocracy”, Quarterly Journal of Economics 138: 1349–402.
Beraja, M, A Kao, D Y Yang, and N Yuchtman (2023), “Exporting the surveillance state via trade in AI”, NBER Working Paper 31676.
Frey, C B (2019), The technology trap, Princeton University Press.
Frey, C B, and G Presidente (2024), “Privacy regulation and firm performance: Estimating the GDPR effect globally”, Economic Inquiry 62: 1074–89.
Frey, C B, G Presidente, and Pia Andres (2024), “Data-biased innovation: Directed technological change and the future of artificial intelligence”, The Oxford Martin Working Paper Series on Technological and Economic Change.
Goodfellow, I, Y Bengio, and A Courville (2016), Deep learning, MIT press.
Habakkuk, H J (1962), American and British technology in the 19th century, Cambridge University Press.
Hanlon, W W (2015), “Necessity is the mother of invention: Input supplies and directed technical change”, Econometrica 83(1): 67–100.
Hassler, J, P Krusell, and C Olovsson (2021), “Directed technical change as a response to natural resource scarcity”, Journal of Political Economy 129(11): 3039–72.
Hornbeck, R, and S Naidu (2014), “When the levee breaks: Black migration and economic development in the American South”, American Economic Review 104(3): 963–90.
Johnson, G, S Shriver, and S Goldberg (2023), “Privacy and market concentration: Intended and unintended consequences of the GDPR”, Management Science 69(10): 5695–721.
Jumper, J, R Evans, A Pritzel, T Green, M Figurnov, O Ronneberger, K Tunyasuvunakool, R Bates, A Žídek, A Potapenko, et al. (2021), “Highly accurate protein structure prediction with alphafold”, Nature 596(7873): 583–89.
Kasy, M (2022), “The political economy of AI regulation: Towards democratic control of the means of prediction”, Institute of Labor Economics Discussion Paper 16948.
Klinger, J, J Mateos-Garcia, and K Stathoulopoulos (2020), “A narrowing of AI research?”, working paper, available at SSRN.
Peukert, C, S Bechtold, M Batikas, and T Kretschmer (2022), “Regulatory spillovers and data governance: Evidence from the GDPR”, Marketing Science 41(4): 746–68.
Russell, S J, and P Norvig (2016), Artificial intelligence: A modern approach, Pearson.
Taalbi, J, and H Nielsen (2021), “The role of energy infrastructure in shaping early adoption of electric and gasoline cars”, Nature Energy 6(10): 970–76.