A rising quantity of students have warned about social, political, and financial harms related to the present trajectory of artificial intelligence (AI) (Acemoglu et al. 2023, Frey 2019, Kasy 2022). In The Anxious Era, for instance, social psychologist Jonathan Haidt factors to the hyperlink between social media use and teenage despair, as giant know-how firms search to maximise engagement to amass private information for focused promoting. In the same spirit, Acemoglu and Restrepo (2020) argue that this path leads to “the mistaken type of AI”, emphasising the necessity of creating AI applied sciences that prioritise moral concerns and person privateness.
In a brand new paper, we discover whether or not the trajectory of AI improvement could be redirected in the direction of much less data-intensive strategies (Frey et al. 2024). In doing so, we situate our evaluation inside the literature on directed technological change, which posits that technological innovation shifts away from scarce or expensive elements of manufacturing (Acemoglu 1998, 2002, Hanlon 2015). Notable examples of this embrace excessive wages and labour shortages prompting funding in automation applied sciences (Habakkuk 1962, Allen 2009, Hornbeck and Naidu 2014), and oil value shocks or carbon taxes driving the improvement of inexperienced applied sciences (Acemoglu et al. 2012, Hassler et al. 2021).
Constructing on this logic, we hypothesise that the Common Knowledge Safety Regulation (GDPR) raised the prices of storing and processing private information (Frey and Presidente 2024), incentivising firms to prioritise investments in data-efficient strategies over data-intensive approaches, and thereby altered the composition of AI patenting.
Traits in AI patenting
Drawing upon the current literature and the recommendation of laptop scientists, we introduce a novel framework for evaluating AI applied sciences primarily based on their information depth (Russell and Norvig 2016, Goodfellow et al. 2016). AI applied sciences constructed on deep-learning strategies require giant datasets to successfully tune tens of millions of parameters. In distinction, knowledge-based techniques rely on structured guidelines and skilled information fairly than intensive information. Bayesian strategies improve effectivity by leveraging prior information. Moreover, varied strategies have emerged to cut back information necessities: switch studying (together with zero-shot and few-shot studying) repurposes information throughout duties, whereas artificial information era creates artificial coaching examples. For instance, DeepMind’s AlphaFold included some of its personal predictions into its coaching information (Jumper et al. 2021). We categorise these strategies as ‘data-saving’, in distinction to deep-learning approaches, which we classify as ‘data-intensive’.
Subsequent, we take our taxonomy to the information: we establish related patents by conducting key phrase searches of patent titles and abstracts in the European Patent Workplace’s PATSTAT World database (2024 Spring Version), analysing them at the patent household degree to keep away from double-counting the similar invention throughout jurisdictions. Utilizing this dataset, we doc a number of stylised info about AI patenting traits worldwide.
First, we observe a notable technological shift from data-saving AI to data-intensive deep studying all through the 2010s. Between 2000 and 2021, the inventory of data-intensive AI patents grew at a powerful annual price of 52%, whereas data-saving AI patents exhibited a extra modest progress price of 19% per 12 months. Curiously, data-saving patent exercise remained beneath its 2004 degree till 2013 however skilled a notable resurgence following the implementation of the GDPR in 2018 (Determine 1).
Determine 1 Common quantity of AI patent households per applicant
Between 2018 and 2021, transfer-learning patents surged by 185%, whereas artificial information era and Bayesian strategies grew by 86% and 68%, respectively, albeit from modest baselines (Determine 2).
Determine 2 Patenting throughout AI know-how lessons per 12 months
The potential affect of the regulatory atmosphere on this pattern is additional underscored by vital geographic disparities in AI patenting exercise and technological specialisations. Since the launch of China’s ‘Made in China 2025’ initiative, Chinese language AI patenting has surged (Determine 3), with universities and authorities establishments accounting for 86% and 54% of international AI patents in these classes, respectively, in comparison with simply 3% and 4% in the US (Determine 4). This dominance is especially evident in data-intensive AI and extends to the personal sector, supported by authorities procurement efforts that generate worthwhile coaching information for industrial use (Beraja et al. 2021, Beraja et al. 2023).
Determine 3 AI patent households filed by nation
In distinction, US corporations lead in data-saving AI innovation, contributing 45% of international data-saving patents. Though the EU lags considerably in general AI patenting, it has a comparatively increased share of indigenous data-saving patents, with 13% of private-sector AI patents in comparison with simply 5% in China.
Determine 4 AI patent households by establishment and nation
Thus, whereas data-intensive AI stays dominant throughout areas, the steadiness varies: China strongly favours data-intensive approaches, the US maintains a extra balanced portfolio, and the EU exhibits a relative emphasis on data-saving applied sciences. This, we observe, mirrors patterns of information privateness regulation. In China, state procurement insurance policies have explicitly incentivised AI improvement for data-intensive surveillance functions (Beraja et al. 2021), whereas in the US, information privateness depends on a patchwork of sector-specific laws fairly than complete federal safety. The EU, in distinction, has pursued a particular regulatory strategy by its GDPR, emphasising particular person privateness safety over information accumulation.
The affect of privateness regulation on AI innovation
Following the literature on directed technological change, we hypothesise that the GDPR, by rising the price of storing and processing private information (Frey and Presidente 2024), has incentivised firms to take a position extra in data-saving strategies and cut back their reliance on data-intensive ones, thereby altering the trajectory of technological change in AI.
For our empirical evaluation, we leverage the GDPR’s attain, which impacts patent candidates exterior the EU if they aim EU customers, to measure how a lot firms worldwide depend on EU markets. To gauge a agency’s publicity to the GDPR, we use inter-industry linkages from the OECD Inter-Nation Enter-Output (ICIO) Tables, capturing their publicity to EU markets as outlined by Frey and Presidente (2024). By breaking down the ICIO Tables into intermediate enterprise use and ultimate family consumption, we exclude business-to-business transactions, that are much less impacted by the GDPR, permitting us to focus on shopper transactions.
We start by presenting the baseline outcomes utilizing the full pattern of patent candidates, which incorporates corporations, particular person candidates, public establishments, and universities. Subsequent, we slim our focus to company candidates, inspecting how the GDPR’s results differ primarily based on agency traits, comparable to measurement and age. Taking benefit of the timing of the GDPR introduction, and the various publicity of corporations to this regulation, we discover that:
- Patent candidates (together with corporations, universities, public establishments, and people) affected by GDPR have redirected their creative efforts in the direction of much less data-intensive and extra data-saving AI approaches.
- The first drivers of this shift had been older and bigger firms primarily based in the EU.
- Whereas altering the technological trajectory of AI, the GDPR additionally decreased general AI patenting in the EU whereas amplifying the market dominance of established corporations.
Redirecting AI
Over the previous decade, the overwhelming focus of AI analysis has been on data-intensive deep-learning strategies – which incentivise firms to amass private information – usually at the expense of exploring less-data-dependent, rule-based techniques (Klinger et al. 2020). These techniques are additionally behind the most up-to-date advances in generative AI. Nevertheless, regardless of many productive use circumstances, some students argue that the present route of AI improvement might cut back general welfare (Acemoglu and Johnson 2023).
Contemplate the historic trajectory of electrical automobiles. At the starting of the twentieth century, they had been as aggressive as their gasoline-powered counterparts. Nevertheless, a scarcity of funding in electrical infrastructure, together with vital oil discoveries, shifted market dynamics decisively in favour of inner combustion engines (Taalbi and Nielsen 2021). This shift led to an inefficient technological lock-in. Now, a century later, we’re witnessing a renewed shift towards electrical automobiles as efforts intensify to right this path-dependent technological trajectory. Think about the technological trajectory if we had the foresight to tax carbon in the early 1900s. One may make the same argument for taxing information.
At the similar time, the stronger response from established corporations signifies that privateness laws might inadvertently reinforce incumbent benefits whereas dampening general innovation – a sample our examine finds evident in the area of AI. Certainly, a rising physique of analysis highlights the adverse results of the GDPR on smaller companies and innovation, contributing to elevated market focus (Frey and Presidente 2024, Peukert et al. 2022, Johnson et al. 2023). The upcoming EU AI Act may intensify this pattern by imposing larger compliance burdens on smaller corporations and doubtlessly shifting technological improvement towards less-data-intensive strategies. Its emphasis on explainability poses specific challenges for deep-learning applied sciences – accountable for most progress in the area over the previous decade.
Investigating how the EU AI Act influences each the quantity and route of AI innovation presents a worthwhile alternative for future analysis. What our examine exhibits that it’s attainable in precept to form the trajectory of AI improvement by coverage intervention.
References
Acemoglu, D (1998), “Why do new applied sciences complement abilities? Directed technical change and wage inequality”, Quarterly Journal of Economics 113(4): 1055–89.
Acemoglu, D (2002), “Directed technical change”, The Evaluation of Financial Research 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 mistaken type of AI? Artificial intelligence and the future of labour demand”, Cambridge Journal of Areas, Financial system and Society 13(1): 25–35.
Acemoglu, D, P Aghion, L Bursztyn, and D Hemous (2012), “The atmosphere and directed technical change”, American Financial Evaluation 102(1): 131–66.
Allen, R C (2009), The British Industrial Revolution in international perspective, Cambridge College 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 by way of commerce in AI”, NBER Working Paper 31676.
Frey, C B (2019), The know-how entice, Princeton College Press.
Frey, C B, and G Presidente (2024), “Privacy regulation and agency efficiency: Estimating the GDPR impact globally”, Financial Inquiry 62: 1074–89.
Frey, C B, G Presidente, and Pia Andres (2024), “Knowledge-biased innovation: Directed technological change and the future of artificial intelligence”, The Oxford Martin Working Paper Collection on Technological and Financial Change.
Goodfellow, I, Y Bengio, and A Courville (2016), Deep studying, MIT press.
Habakkuk, H J (1962), American and British know-how in the nineteenth century, Cambridge College Press.
Hanlon, W W (2015), “Necessity is the mom of invention: Enter provides and directed technical change”, Econometrica 83(1): 67–100.
Hassler, J, P Krusell, and C Olovsson (2021), “Directed technical change as a response to pure useful resource shortage”, Journal of Political Financial system 129(11): 3039–72.
Hornbeck, R, and S Naidu (2014), “When the levee breaks: Black migration and financial improvement in the American South”, American Financial Evaluation 104(3): 963–90.
Johnson, G, S Shriver, and S Goldberg (2023), “Privacy and market focus: Meant and unintended penalties of the GDPR”, Administration Science 69(10): 5695–721.
Jumper, J, R Evans, A Pritzel, T Inexperienced, M Figurnov, O Ronneberger, Okay Tunyasuvunakool, R Bates, A Žídek, A Potapenko, et al. (2021), “Extremely correct protein construction prediction with alphafold”, Nature 596(7873): 583–89.
Kasy, M (2022), “The political financial system of AI regulation: In direction of democratic management of the means of prediction”, Institute of Labor Economics Dialogue Paper 16948.
Klinger, J, J Mateos-Garcia, and Okay Stathoulopoulos (2020), “A narrowing of AI analysis?”, working paper, out there at SSRN.
Peukert, C, S Bechtold, M Batikas, and T Kretschmer (2022), “Regulatory spillovers and information governance: Proof from the GDPR”, Advertising and marketing Science 41(4): 746–68.
Russell, S J, and P Norvig (2016), Artificial intelligence: A contemporary strategy, Pearson.
Taalbi, J, and H Nielsen (2021), “The position of vitality infrastructure in shaping early adoption of electrical and gasoline vehicles”, Nature Vitality 6(10): 970–76.