2.1 The Task Framework of Automation and Empirical Studies on Robots

In its capacity of “intelligent machine,” AI can carry out typically human tasks, like pattern recognition, text analysis, and prediction without supervisions. As such, it is akin to other forms of labor-substituting automation technologies studied by the vast literature recently surveyed by Restrepo (2023). As discussed in detail there, the “task model of automation” provides a simple foundation to assess the effect of labor-displacing technological change. The task model postulates that producing goods and services requires the completion of tasks that can be assigned to groups of workers, defined by their skills and abilities, or robots and machines. The crucial takeaway of this framework is that the relative labor demand for groups of workers whose tasks are taken over by robots falls with automation, and the level of labor demand for these groups might even fall depending on the productivity gains that come with robot adoption, and workers’ ability to transition to occupations unaffected by task displacement. Acemoglu and Restrepo (2022) derive explicit closed-form log-linear formulas that summarize the effect of automation on wages of affected worker groups. Aggregate productivity and the demand for labor in non-automated tasks that workers can easily reallocate positively impact wages, while the share of tasks displaced by automation directly lower wages. The overall effect of automation on labor market outcomes thus hinges on whether productivity improvements and their demand spillovers to other sectors outweigh direct substitution effects. When applied to the data, this framework explains 50% of the labor share declines across US industries over the period 1987–2016.

We refer the interested reader to Restrepo (2023) for a full discussion of the automation literature concerning industrial robots. For the purposes of our review, we can summarize the empirical literature in this strand as showing that groups experiencing large task displacement from robot adoption saw decreasing employment and wages, which is consistent with US evidence that firms deployed these technologies primarily with the aim of substituting human labor (Acemoglu, Anderson, et al. 2022). Similarly, firm-level studies in various countries generally found that automation led to a reduction in the labor share and an increase in sales per worker, although sometimes accompanied by an employment expansion. However, as highlighted by Restrepo (2023), these effects are hard to interpret and translate to the aggregate level as adopting firms might just be expanding at the expense of competitors. Accordingly, the impact of this first wave of automation on economy-wide productivity remain uncertain. The above discussion clarifies why the emerging literature on AI has focused on finding a measure of AI task exposure for different occupations, defined as the share of tasks that could be carried out by AI in the relevant occupation or job. Under the lenses of the automation literature, this measure should give a sufficient statistic to predict relative wage changes, as more exposed workers would see relative wage losses relative to less exposed workers. However, the nature of AI and specifically gen-AI puts into question whether this technology will act as a complement or a substitute for workers in related tasks (see, e.g., Autor (2022)). For this reason, the pre-gen-AI literature on ML has looked to empirical evidence on the relation between AI exposure or adoption measures and various outcomes, while more recent studies produced experimental evidence on gen-AI. In what follows, we first present studies on exposure measures and then turn to the empirical evidence that may be used to interpret these measures, and understand whether AI might differ from previous waves of automation.

2.2 AI Task Exposures

The literature on non-AI automation, focusing primarily on industrial robots, found strong exposure for blue-collar and less educated workers (Acemoglu and Restrepo 2022), which has been compounded by increasing substitutability of low-skilled workers with machines (A. Berg et al. 2024). Under the lenses of potential task exposures, however, AI appears different.

Exposure Estimates. Early studies on AI task exposures focused on ML. Namely, Brynjolfsson, Mitchell, and Rock (2018) find that machine learning can displace tasks for occupations throughout the wage distribution in a rather uniform way, which contrast markedly with the low-skill bias of industrial robots. In a similar vein, Webb (2019) finds that AI is mainly directed at high-skilled tasks and will affect highly educated and older workers. The author also estimates that AI could reduce 90:10 wage inequality by 5 to 10%. This finding is under the assumption that the wage elasticity to AI exposure is the same as the wage elasticity to software and robots’ exposure, in line with the task framework of automation of Acemoglu and Restrepo (2018) that sees larger task exposure as leading to lower relative wages. E. Felten, Raj, and Seamans (2021) build a comprehensive measure of exposure to machine learning linking 10 AI applications (e.g., image generation, language modeling, etc.) to 52 human abilities (e.g., oral comprehension and expression, inductive reasoning, etc.) which are in turn mapped to occupations. Highly educated, highly paid, white-collar occupations appear most exposed to machine learning. E. W. Felten, Raj, and Seamans (2023) obtain analogous findings when restricting their measure to ML applications more related to gen-AI such as language models and image generation. Eloundou et al. (2023) estimate that up to 80% of US workforce could have at least 10% of their tasks replaced and 19% of workers stand to lose at least 50% of their tasks to LLMs. This analysis also shows that about 15% of all worker tasks in the US could be completed significantly faster at the same level of quality if LLMs are employed. Once again, this exposure is higher for higher-paid and high-skilled professional occupations. In a similar fashion, Gmyrek, J. Berg, and Bescond (2023) also show that the highest share of tasks exposed to AI technology are clerical jobs, where the majority of tasks fall into medium- to high-level exposure. Additionally, the authors analyze separately automation (labor substitution) and augmentation effects, finding that the latter outweigh the former across both low- and high-income countries. High-income countries see 5.5% of total employment exposed to automation and 13.4% to augmentation from AI, while the same percentages are 0.4% and 10.4% for low-income countries, respectively. Interestingly, the earliest study to incorporate earlier machine learning capabilities into a measure of potential computerization (Frey and Osborne 2017) found a negatively sloped relation between IT (Information Technology)-driven labor substitution exposure and average occupational earnings. Therefore, if we interpret the time of the studies as indicative of the state of machine learning capabilities, it appears that AI has become more biased towards high-skilled occupations in recent years.

Correcting for Complementarity. While some of the studies described above explicitly interpret AI exposure as a measure of potential substitutability, most are agnostic on what exactly task exposure entails. By contrast, Pizzinelli et al. (2023) construct an indicator of AI occupational exposure adjusted for the complementarity of the technology with the exposed occupation (C-AIOE). The resulting measure therefore provides a more direct gauge of the potential displacement from AI. Occupations’ substitutability and complementarity are assessed independently from exposure itself based on US O*NET occupation characteristics.⁵ The authors first compute an unadjusted exposure index, reporting that AEs have a higher share of high-exposure jobs with both high and low levels of complementarity. When considering demographic characteristics, female and high-skilled workers appear more exposed. Adjusting for complementarity, the C-AIOE measure shows a lower exposure for high-skilled occupations relative to its unadjusted counterparts, leading to a more uniform exposure of AI throughout occupations. Employing this framework in model-based scenarios, Cazzaniga et al. (2024) find that AI would raise wage inequality for sufficiently high complementarity, and wealth inequality regardless of complementarity. Similarly, Huang (2024) classifies tasks that are exposed to AI into groups that allow a better assessment of whether the technology will substitute or complement human labor. Using this classification, the author shows that high-skill workers with a high AI exposure would see their employment share in the economy increase relative to less AI-exposed. This increase would be similar to that experienced by progress in IT, which affects predominantly low-skilled workers, but substitutes them instead of complementing them.⁶

Kogan et al. (2023) provide an estimate of wage and employment effects of AI on occupational groups that explicitly incorporates the potential of AI to both augment and displace tasks. The authors calibrate a model that matches empirical evidence on the impact of labor-saving and labor-augmenting technologies patented since the 1980s. In the model, technology substitutes routine tasks, augments non-routine tasks, and causes skill obsolescence for incumbent workers who cannot efficiently use it. Accordingly, each technology has both a labor-saving and a labor-augmenting exposure. Assuming that workers with the highest labor-saving exposure will lose 10% of their earnings from AI, the model computes the direct effects–that is, excluding reallocation across sectors–of current AI technology by occupational group. On average, workers stand to lose 4% of their earnings. This effect stems from skill obsolescence as, for almost all occupations, the potential of task augmentation undo the negative task displacement effects. In terms of skill content, more routine occupations such as office and admin work, production and transportation stand to lose the most (6–8%), while higher-skilled occupations like management, engineering, legal, education and social service occupations log the lowest losses (around 2%).

Estimates of Aggregate Employment Impacts. Think tank and private sector estimates provide some additional information on the potential extent of AI impacts on the labor market. Pew Research (2023) finds that 19% of all US workers are in occupations most exposed to AI.⁷ By the same token, Briggs and Kodnani (2023) find an employment exposure of 25% for the US, 24% for the euro area, and 18% worldwide, with effects concentrated in AEs. Ellingrud et al. (2023) offers somewhat smaller estimates, predicting that 8% of hours currently worked will be automated by 2030 because of gen-AI, ranging from around 16% of all hours in STEM (science, technology, engineering, and math), education, and business and legal professional occupations, to 4% in production work and 3% in agriculture. However, when factoring in other structural factors, these figures translate into slowing—but still increasing—demand for higher skilled occupations like STEM, and decreased demand only for selected customer-facing and production occupations.

Aside from Briggs and Kodnani (2023) and Gmyrek, J. Berg, and Bescond (2023), only Pizzinelli et al. (2023) offer a glimpse into the potential difference of AI impact between AEs and EMs. The authors use worker-level microdata from 2 advanced economies (AEs: US and UK) and 4 emerging markets (EMs: Brazil, Colombia, India, and South Africa). Stronger high-skilled workers’ exposures also translate into advanced economies being more susceptible to employment displacement from AI compared to EM counterparts. However, this difference is smaller when considering task complementarity.

Economic Feasibility of Task Replacement. In closing this section, it should be noted that task exposure measures the theoretical possibility that AI may carry out some tasks. This feature makes exposure a potentially poor measure of the employment effects that we might expect from the adoption of the technology. Svanberg et al. (2024) explicitly assess the technical and economic feasibility of automating vision tasks that have a high theoretical AI task exposure. They find that, given current costs, most US businesses would not automate most tasks with high exposure. Only 23% of exposed jobs (8% of total jobs) in U.S. non-farm businesses have at leas one task that is economically attractive to automate. When considering the contribution of automated tasks to compensation, the authors conclude that only 0.4% of non-farm wages would be potentially lost to automated computer vision.

2.3 Empirical and Experimental Studies on AI

In the previous section, we saw that high-skilled and higher-paid workers seem to be more exposed to AI. However, as discussed, the consequences of exposure are less clear-cut than in the case of previous waves of automation. For this reason, we now move to survey empirical studies on the effect of AI and some of the recent experimental literature focusing more narrowly on gen-AI.

Firm-level Studies on ML Impacts. Empirical papers on AI adoption have to rely on data collected before the introduction of more powerful gen-AI tools. Alekseeva et al. (2021) and Acemoglu, Autor, Hazell, et al. (2022) analyze AI-related vacancies in the US. Alekseeva et al. (2021) find a large increase in the demand of AI skills across industries over the period 2010–2019. Job postings listing AI skills command a 5% wage premium compared to other same-firm, same-job postings, and firms that demand more AI skills tend to be larger and more innovative. Acemoglu, Autor, Hazell, et al. (2022) employ similar data for 2010–2018 and find that establishments that increase postings related to ML skills simultaneously reduce vacancies unrelated to ML and change their skill requirements more quickly, which might be consistent with substitution of non-AI work. The authors cannot detect aggregate effects of AI technology, which they ascribe to limited economy-wide adoption. More recently, Babina, Fedyk, A. X. He, et al. (2024) and Babina, Fedyk, A. He, et al. (2024) combine US firm data with employee resumes. In Babina, Fedyk, A. X. He, et al. (2024), the authors show that firms investing in AI tend to transition to a more educated workforce mostly focused on STEM and IT. At the same time, AI adoption tends to fatten firms’ hierarchical structures, increasing workers in junior positions and decreasing workers in middle management and senior roles. In terms of firm outcomes, Babina, Fedyk, A. He, et al. (2024) find that investment in AI experience led to increased sales, employment, market valuations and product innovation. Further, AI-fuelled growth is more prevalent among larger firms and in industries with higher market concentration.

Copestake et al. (2023) offer a developing country perspective on AI adoption, using demand for ML skills observed in the text of posted job descriptions in India over the period 2010–2019. Consistent with Alekseeva et al. (2021), the authors document a rapid increase in ‘AI demand’ after 2016, particularly in the IT, finance, and professional services industries, as a well as a 13 to 17% salary premium for job postings demanding AI skills. At the same time, similarly to Acemoglu, Autor, Hazell, et al. (2022), they document a fall in non-AI job postings and wage offers. Cornelli, Frost, and Mishra (2023) examine AI-related investments across 86 countries for the same period. Their findings link AI-related investments to a shift from mid-skill jobs to high-skill and managerial positions, accompanied by a decline in the labor share of income, higher total factor productivity (TFP) and increasing inequality at the aggregate level. Alderucci et al. (2021) analyze the outcomes of US firms inventing AI technologies, which they identify through the text of patent grants. AI-related innovations are associated with faster employment and revenue growth, higher output per worker, and an increase in within-firm wage inequality. The World Economic Forum (WEF) Future of Jobs Survey (World Economic Forum 2023) predicts that 23% of global jobs will change in the next five years due to industry transformation, including AI-related. The survey confirms previous findings: the fastest-growing roles are technology-related, with AI and ML specialists at the top, while clerical and secretarial roles are the ones declining faster.⁸ Finally, Milanez (2023) conducted 100 case studies of AI applications over the period 2021–2022 in Austria, Canada, France, Germany, Ireland, Japan, the UK, and the US. Only 23% of firms stated that job quantities declined for affected workers; in these cases, workers were reallocated to different tasks (existing or new) without adverse impact on overall employment.

Experimental Evidence on AI Complementarity. Some more recent experimental studies focus more closely on gen-AI. Brynjolfsson, Li, and Raymond (2023) study gen-AI adoption in customer service call centers. The authors report productivity gains accrued primarily by less experienced and lower-skilled workers. Their conclusion is that AI can disseminate the knowledge accumulated by more knowledgeable workers, helping newer workers move down the experience curve. Similarly, Noy and W. Zhang (2023) show that experimental exposure to ChatGPT increases productivity and output quality in writing tasks, while reducing the performance differential between low- and high-ability workers. In the context of radiology, Agarwal et al. (2023) study the effectiveness of human-AI collaboration through an experiment with radiology experts comparing the performance of humans to AI as well as humans assisted by AI. Their findings suggests that it is optimal to assign cases either to humans or to AI, but rarely to a human with the assistance provided by AI since human radiologists tend to underweight the information provided by AI when it deviates substantially from their own beliefs.

2.4 Policies to Mitigate Displacement and Ease Labor Reallocation

While the empirical evidence surveyed above is still ambiguous on the extent and characteristics of AI-driven labor displacement, several academic and policy studies suggested policies to deal with automating technologies. Therefore, it is important to note that the policy recommendations below chiefly relate to the labor-saving nature of AI.

Skills and Workforce Adaptation. Analogously to previous waves of technological innovation, AI is likely to displace old skills and bring about new ones (Acemoglu and Restrepo 2018; Autor 2022). OECD (2023b) and in particular the section by Lassebie (2023) identify new AI-related skills that complement basic digital and data science competencies, as well as others that can foster diversity in the AI workforce. Among the former, programming, machine learning, and data science stand out, while the latter feature other cognitive and transversal skills—namely, creative problem solving, critical thinking, and mentoring.⁹ This study also points out that AI itself might be used to deliver customized training, thereby fostering inclusion and accessibility. Focusing on AI professionals, Borgonovi et al. (2023) similarly conclude that soft skills will continue to be a key requirement for companies hiring AI professionals. Recent evidence on India’s services sector (Copestake et al. (2023)), instead show that establishments looking for AI hiring tend to reduce their demand of analytical and complex communication skills as seen in their job postings. As a result of this changing demand for skills, Lassebie (2023) suggests promoting greater training provision by employers, as well as extending such training beyond currently vulnerable groups to ensure smooth future AI adoption and development.

Taxation and Fiscal Policy Trade-Offs. Acemoglu, Manera, and Restrepo (2020) adopt a different theoretical framework, featuring reduced-form labor market frictions as well as a finite capital supply elasticity. There, they show that setting capital taxes too low relative to labor may promote excessive substitution of workers compared to what is socially optimal. In other words, when taxes on capital are too low relative to taxes on labor, robots become artificially cheaper than workers in executing tasks, which causes larger employment losses from automation than socially optimal. Based on the outcomes from this paper, Acemoglu, Autor, and Johnson (2023) suggest equalizing tax rates on labor and ownership of equipment and algorithms to level the playing field between humans and labor-replacing machines, with the aim of potentially incentivizing human-complementary technology choices.¹⁰ M. A. Berg et al. (2021) see automation as a technology that substitutes unskilled workers and raising the incomes of skilled workers and owners of capital. At the same time, automation has a large productivity effect, therefore robot taxation is not efficient, but simply a means to redistribute the benefits of automation at the detriment of lost output in the long run. A tax on robots that slows the accumulation of automating capital or a higher capital income tax combined with an unskilled wage tax cut deliver the largest welfare benefits. When imperfect competition is included in the model, a markup tax is the second-best policy after a robot tax. A. Berg et al. (2024) study the returns to corporate tax cuts, increased education spending, and increased infrastructure investment in a model economy, comparing a case featuring a broad definition of “robot” capital with a scenario without it. When “robot” capital is not present, infrastructure investments and corporate tax cuts produce more benefit than education spending. Conversely, the “robot” capital economy sees education spending as the most efficient policy measure, followed by infrastructure investment and tax cuts. To interpret these results, it should be noted that the authors assume that “robot” capital is modelled as a substitute for low-skilled labor, akin to pre-gen-AI automation.¹¹

4.1 Reasons for AI regulation

In this section, we discuss the the main reasons for AI regulation. Agrawal, J. Gans, and Goldfarb (2019) and Acemoglu (2024) provide useful overviews of many economic and ethical rationales. In addition to these sources, we consult specialized literature in other fields, as well as the stated aims of proposed and enacted regulations. We choose to focus on six main areas which are prominent in the public discussion or among the intents of regulators. In order, we cover: the effects of AI on competition; privacy; copyright issues; military uses and national security; ethical and bias consideration; and financial stability. We list other issues, among which the impact of AI on electoral systems, in a separate sub-section. Policy initiatives in China, the EU and the US tackling each aspect are briefly mentioned in the relevant sub-section.

4.2 Theory of regulation

The choice and the setup of a regulation of AI also come with a literature, even if still rather limited. In this section, we review some contributions exploring the welfare consequences of AI regulation, which aim to determine the optimal trade-of between opportunity and–potentially existential–risks from progress in this technology.

Acemoglu (2024) stands out as one of the first papers to provide an organized summary of the risks from unregulated AI, as have been discussed in the previous Sub-Section 4.1. A key claim of the paper is that potential economic and societal risks stem from the current use and development of the technology rather than its nature, and downsides could be avoided with appropriate regulation and policies that redirect the course of AI development towards the most productive uses.

Optimal regulation of transformative technologies is at the center of Acemoglu and Lensman (2023) study. The authors build a multi-sector model where risks are revealed over time as technology is adopted across sectors. As adoption proceeds, agents in the economy learn about the probability of a “disaster” which incurs losses proportional to the output produced using the new technology. The benefits of the new technology consists in higher productivity as well as higher productivity growth. Given this main trade-of, optimal adoption of the technology is gradual and increases its speed over time. It is gradual because this allows learning about the risks of the technology, and its speed increases over time as the disaster probability is assessed to be low enough. A role for taxation and regulation emerges if firms do not fully internalize the potential harms of their technology, which the authors justify with some damages being inherently social in their nature. In this case, adoption is too fast relative to optimal, and the gap between effective and optimal adoption increases with the speed of output growth delivered by the new technology. The more transformative the technology, the larger are potential harms from it, due to the fact that losses are taken to be a fraction of total output produced with new technologies. Acemoglu and Lensman (2023) propose two potential tools to close this gap: sector-specific taxes that lower the rate of adoption across all sector, or, as a second best, a “regulatory sandbox” that fully restricts the use of new technologies in sectors with high potential losses until a specific time period where risks should be better known.

Guerreiro, Rebelo, and Teles (2023) assess different approaches to AI regulations in a model featuring AI developers that choose the novelty of their algorithm as well as the size of the population to which this algorithm is deployed. Higher novelty delivers higher utility, but also comes with higher variance of damages realized ex-post. The authors consider two alternative setting: full deployment or beta testing targeted at gauging the damages. In the first setting, developers cannot learn about damages until full deployment takes place, while in the second they can test the algorithm on a fraction of the population and measure realized damages before deliberating on full implementation. In both cases, the planner prefers a lower level of novelty than developers since damages manifest as an externality that is only partly internalized by developers, much like in the Acemoglu and Lensman (2023) setting. The authors consider the combination of two policy solutions commonly proposed to bring the economy to the social optimum with or without beta-testing: (i) regulatory approval of release and (ii) developers’ liability for realized damages. Without beta testing, it is optimal to set a threshold of novelty equal to the social optimum and forbid algorithms above such thresholds, a policy that the authors compare to the EU Commission’s proposal discussed below. Liability for developers delivers the social optimum in this context only if it is full. When beta-testing is available, the optimal policy consists in mandating beta-testing for all levels of novelty and imposing limited liability for damages in case of full deployment. The authors also show that a commonly proposed solution–conditional approval of algorithms that beta-testing reveals to be sufficiently safe–is not optimal since it involves excessive levels of novelty in the testing phase. The authors highlight that this solution might require international cooperation if spillovers are not limited to single countries. In contrast to this approach, Kretschmer et al. (2023) propose a full liability approach to regulation, as discussed more at length in Section 4.3.1.

Without addressing specific policies, Jones (2023) considers when arresting AI development is optimal in the presence of existential risk and different degrees of societal risk aversion. This simple model weighs additional growth against human lives lost, which are modelled as an increased growth rate and a higher death probability, respectively. The planner chooses how long to use AI, realizing growth at the cost of lost lives. AI usage increases with lower risk-aversion, or in an alternative scenario where AI can also deliver increases in life expectancy. A final consideration involves the discount rate attached to future utility. A lower discount rate increases AI if the technology delivers better life expectancy, and lowers its use otherwise.

4.3 Regulation and Governance of AI across Countries

Following Wheeler (2023), we can identify three main challenges for AI regulation: speed, scope, and approach. These challenges are also at the root of major differences in adopted and proposed country-specific frameworks. First, AI’s rapid transformation makes rules and standard-setting particularly difficult, as the landscape is rapidly evolving.⁴⁰ The speed of AI development motivates some authorities to adopt flexible guidelines that are easier to adapt (e.,g., Japan and UK), while others prefer more rigid regulations to contain risks that might manifest rapidly (e.g., the EU). Second, scope differs widely across regulators, starting from the definition of AI itself and extending to the different areas covered.⁴¹ The Appendix A.1 lists the various definitions contained in EU, US, UK and Chinese source documents.⁴² Regulators have so far adopted different strategies ranging from outright temporary bans to voluntary guidelines. The cases of binding regulations have adopted a range of approaches combining ex-ante risk assessments and ex-post liability. The ex-ante assessment approach involves provisions to evaluate the potential risks of AI based on systems’ characteristics or criticality areas of application before risks manifest. Ex-post liability approaches instead make developers and service providers liable for actual realizations of risk events and related externalities. Further, several initiatives have stressed the importance of multilateral organizations, ranging from the EU AI governance architecture (see sub-section 4.3.1) to proposal to establish an authority analogous to the UN’s Intergovernmental Panel on Climate Change (IPCC).⁴³

Before delving deeper into specific cases, some examples may clarify how scope and approaches vary across the world. In terms of scope, measures proposed and implemented in the EU and US concern broad technologies and their applications. By contrast, China regulates algorithms and their content directly. In terms of approaches, the US decentralized the setting of policies for specific AI applications to relevant departments and agencies, while the EU stands poised to adopt a centralized approach featuring ex-ante risk assessments and a new AI governance structure. Among other examples, Italy stood out for implementing a blanket ban on ChatGPT until adequate data protection issues were clarified at the end of April 2023, while Japan is leaning towards voluntary guidelines.⁴⁴

In the rest of the section, we describe the main regulations developing in selected AEs and EMDEs, including EU, the US, the UK, China, Brazil, and India, with an eye to the current debates and their ramifications. We review the areas that have been covered by each and the regulatory approach followed by relevant authorities or proposals. None of the countries we consider have focused directly on financial stability, while regulations primarily addressed monopoly, ethical, and privacy risks.⁴⁵ US regulation covers explicitly national security and military uses of AI, while the EU has been more ambiguous, as legislation on this matter is decided at the national level. Copyright of AI-generated (or co-generated) material is not explicitly tackled in current frameworks, and in some cases deferred to the discretion of the courts on a case-by-case basis (e.g., in the US and China).

4.3.4 A summary of current regulations in EU, US, and China

We summarize the main characteristics of AI-specific regulations that are in place or under discussion in Table 1. From left to right, the table lists the sources of these regulatory actions, the approaches taken by the relevant authorities, the type of AI regulated according to the definitions reported in Appendix Section A.1 , and the areas subject to regulation defined as in our discussion in Section 4.1. Appendix Section B provides more information on each aspect.

^{Table 1:}

AI-Specific Regulations in selected AEs and China

Note: This table covers AI-specific regulations, while we signal with “GN” (as per “general”) when other regulations/laws include these issues in a general context. We do not consider national laws for the EU nor State-specific regulations for the US. The EU AI Act is not officially an EU law (as of March 13, 2024). The Provisional EU AI Act is as approved by COREPER, by EU Parliament’s Committees, and in plenary session by the EU Parliament (March 13, 2024). For China, the source is an unofficial English translation of the Interim Measures document. For the US, some aspects are also covered in the “Blueprint for an AI Bill of Rights” (October 2022) by the White House Office of Science and Technology Policy (OSTP) aimed to set a roadmap for the responsible use of AI especially on potential human rights. The “’general purpose’ AI model” is defined in the Provisional EU AI Act as”(63) AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are released on the market; “[…] “(66) ‘general-purpose AI system’ means an AI system which is based on a general-purpose AI model, that has the capability to serve a variety of purposes, both for direct use as well as for integration in other AI systems;[…]” The “dual-use model” in the US EO is defined as “AI model that is trained on broad data; generally, uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts; and that exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters […]”

^{Table 1:}

AI-Specific Regulations in selected AEs and China

Country	AI regulation	Approach	Coverage	Competition	Privacy	Copyright	Military	Ethical	Fin. Stability
EU	Provisional EU AI Act adopted by the EU Parliament (March 13, 2024)	Based on ex-ante risk assessment	AI applications, ‘general purpose’ AI models (incl. high risk and systemic) also integrated into AI systems.	GN	✓	✓/GN	X	✓	✓/GN
US	Executive Order (EO) on Safe, Secure, and Trustworthy Artificial Intelligence (October 2023)	Decentralized based on guidelines, principles and priorities	AI and AI models, incl. “dual use” foundation models	✓	✓	✓	✓	✓	GN (but incl. cybersecurity)
China	Interim Measures for the Management of Generative Artificial Intelligence Services (August 2023)	Ethical model design and content, ex-post liability	Only on public usages of gen-AI	✓	✓	GN	X/GN (but incl. National security)	✓	GN

Note: This table covers AI-specific regulations, while we signal with “GN” (as per “general”) when other regulations/laws include these issues in a general context. We do not consider national laws for the EU nor State-specific regulations for the US. The EU AI Act is not officially an EU law (as of March 13, 2024). The Provisional EU AI Act is as approved by COREPER, by EU Parliament’s Committees, and in plenary session by the EU Parliament (March 13, 2024). For China, the source is an unofficial English translation of the Interim Measures document. For the US, some aspects are also covered in the “Blueprint for an AI Bill of Rights” (October 2022) by the White House Office of Science and Technology Policy (OSTP) aimed to set a roadmap for the responsible use of AI especially on potential human rights. The “’general purpose’ AI model” is defined in the Provisional EU AI Act as”(63) AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are released on the market; “[…] “(66) ‘general-purpose AI system’ means an AI system which is based on a general-purpose AI model, that has the capability to serve a variety of purposes, both for direct use as well as for integration in other AI systems;[…]” The “dual-use model” in the US EO is defined as “AI model that is trained on broad data; generally, uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts; and that exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters […]”

View Table

^{Table 1:}

AI-Specific Regulations in selected AEs and China

Country	AI regulation	Approach	Coverage	Competition	Privacy	Copyright	Military	Ethical	Fin. Stability
EU	Provisional EU AI Act adopted by the EU Parliament (March 13, 2024)	Based on ex-ante risk assessment	AI applications, ‘general purpose’ AI models (incl. high risk and systemic) also integrated into AI systems.	GN	✓	✓/GN	X	✓	✓/GN
US	Executive Order (EO) on Safe, Secure, and Trustworthy Artificial Intelligence (October 2023)	Decentralized based on guidelines, principles and priorities	AI and AI models, incl. “dual use” foundation models	✓	✓	✓	✓	✓	GN (but incl. cybersecurity)
China	Interim Measures for the Management of Generative Artificial Intelligence Services (August 2023)	Ethical model design and content, ex-post liability	Only on public usages of gen-AI	✓	✓	GN	X/GN (but incl. National security)	✓	GN

Note: This table covers AI-specific regulations, while we signal with “GN” (as per “general”) when other regulations/laws include these issues in a general context. We do not consider national laws for the EU nor State-specific regulations for the US. The EU AI Act is not officially an EU law (as of March 13, 2024). The Provisional EU AI Act is as approved by COREPER, by EU Parliament’s Committees, and in plenary session by the EU Parliament (March 13, 2024). For China, the source is an unofficial English translation of the Interim Measures document. For the US, some aspects are also covered in the “Blueprint for an AI Bill of Rights” (October 2022) by the White House Office of Science and Technology Policy (OSTP) aimed to set a roadmap for the responsible use of AI especially on potential human rights. The “’general purpose’ AI model” is defined in the Provisional EU AI Act as”(63) AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are released on the market; “[…] “(66) ‘general-purpose AI system’ means an AI system which is based on a general-purpose AI model, that has the capability to serve a variety of purposes, both for direct use as well as for integration in other AI systems;[…]” The “dual-use model” in the US EO is defined as “AI model that is trained on broad data; generally, uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts; and that exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters […]”

4.4 AI regulation, international cooperation, and multilateral actions

As noted by Gopinath (2023), AI’s impacts have a global component that calls for international cooperation. We discussed above how AI poses threats to national security and international relations. In addition to strategic considerations, the need for cooperation emerges from externalities inherent to the technology (Acemoglu 2024) that may spill over national borders, or regulation themselves. Agrawal, J. S. Gans, and Goldfarb (2019) see the potential for a “race to the bottom” in regulatory standards in order to attract AI companies. Guerreiro, Rebelo, and Teles (2023) note the need for international cooperation to implement optimal regulatory frameworks. However, the authors note that cooperation may be complicated by contrasting objectives and definitions of social welfare.

We report below the main initiatives relative to international cooperation and multilateral actions relative to AI, ordered from oldest to newest.⁹⁶ At the end of this section, we briefly discuss private sector initiatives.

OECD “AI Principles.” In May 2019, the OECD adopted a set of “AI Principles” to foster innovative and trustworthy AI that respects human rights and democratic values. The five principles are: 1) inclusive growth, sustainable development and well-being, 2) human-centred values and fairness, 3) transparency and explainability, 4) robustness, security and safety, and 5) accountability.⁹⁷ The principles have since been endorsed by 46 countries worldwide and have been embedded in several national and multinational initiatives. OECD (2023c) reports that, as of May 2023, at least 50 countries have envisioned AI strategies, and AI-related policy initiatives have taken shape in 70 countries, half of which are EMDEs. Initiatives are selected for their implementation of one or more of the AI Principles and differ in scope and range. Examples of initiatives include: promoting the use of AI for environmental sustainability; protecting privacy and human rights; disclosing information about use of AI systems; application or new regulations related to potential risks of AI; and having independent oversight bodies to audit the use of algorithms.⁹⁸ Lastly, within the OECD umbrella, the OECD AI Policy Observatory has been established with the purpose of sharing data, information, and analyses on AI (more info at sub section 5.2).⁹⁹

NATO AI strategy. In its October 2021 meeting, the NATO Allied Defence Ministers formally adopted an AI strategy for defense and national security, committing to the cooperation and collaboration necessary for its implementation. Signatories committed to “Principles of Responsible Use” for the development and deployment of AI. The six principles are: (i) Lawfulness; (ii) Responsibility and Accountability; (iii) Explainability and Traceability; (iv) Reliability; (v) Governability; and (vi) Bias Mitigation.¹⁰⁰

World Economic Forum and UN initiatives. In June 2023, The World Economic Forum (WEF) launched the “AI Governance Alliance” to unite industry leaders, governments, academic institutions and civil society organizations around the goal of meaningful AI governance.¹⁰¹ Ultimately, AI’s opportunities and challenges, including governance and regulation, were at the center of the 2024 Davos meetings.¹⁰²

In October 2023, the UN Secretary-General announced the creation of an AI Advisory Body to explore risks, opportunities and the global governance of artificial intelligence.¹⁰³ In December 2023, The Advisory Body released the Interim Report “Governing AI for Humanity”. The document calls for a closer alignment between international norms and AI’s developed and deployment. The AI Advisory Body recommends five guiding principles: 1. AI should be governed inclusively, by and for the benefit of all; 2. AI must be governed in the public interest; 3. AI governance should be built in step with data governance and the promotion of data commons; 4. AI governance must be universal, networked and rooted in adaptive multi-stakeholder collaboration; 5. AI governance should be anchored in the UN Charter, International Human Rights Law, and other agreed international commitments such as the Sustainable Development Goals. The final recommendations are set to be published in the summer of 2024 after a multi-stakeholder consultation process.¹⁰⁴ In September 2024, the UN Summit of the Future will consider the adoption of a “Global Digital Compact.” The Compact should address several aspects of AI, including its governance.¹⁰⁵

G7: “Hiroshima Process International Code of Conduct for Advanced AI Systems”. On October 30, 2023, G7 leaders agreed to the “Hiroshima Process International Code of Conduct for Advanced AI Systems”.¹⁰⁶ The code provides “voluntary guidance for actions by organizations developing the most advanced AI systems.” The document lays out guidance principles encouraging: risk mitigation in all parts of the AI process; increased transparency during development reporting systems’ capabilities and domains of use; sharing of information and reporting of incidents in development; developing governance, increasing security controls and advance the development of international standards; deploying reliable provenance mechanisms such as watermarking; prioritizing research to increase AI’s safety and on applications that would help sustainable development goals; implementing data input measures and protection for personal data and intellectual property. Several of these measures were in line with the provisional EU AI Act and the US EO. Italy, assuming its 2024 G7 Presidency, announced that it will focus on AI as one of the key themes, and that it will plan a special session before June’s leaders’ summit to assess the impact of AI with the involvement of scholars, managers and experts.¹⁰⁷

The first global “AI Safety Summit”. On November 1–2, 2023 the British government hosted the first global “AI Safety Summit,” gathered representatives from 28 national governments, including the UK, the US, China and several European countries. The summit also saw the participation of multilateral organizations—such as the European Commission, the Council of Europe, the OECD and the UN—as well as academics, entrepreneurs and civil society representatives. The initiative discussed how to approach and regulate AI technologies.¹⁰⁸

On this occasion, the participating national governments and the EU signed the “Bletchley Declaration,” resolving to further national, multilateral and bilateral action to promote on AI safety research and establish risk-based policy across the respective countries.¹⁰⁹ The declaration acknowledges that specific approaches may differ, while stressing the need for international cooperation. As discussed in Section 4.3.6, the Summit saw the launch of the “UK Government’s AI for Development Programme.” The initiative is expected to become permanent, with the next two AI Safety Summits to be hosted by South Korea and France in 2024.

Guidelines for Secure AI System Development. In November 2023, the UK National Cyber Security Centre (NCSC), the US Cybersecurity and Infrastructure Security Agency (CISA), and several other national cyber security agencies across AEs and EMDEs released a set of “Guidelines. It also includes inputs from several private companies, AI providers, and universities.¹¹⁰ It recommends safety guidelines for providers of any system that includes the use of AI (i.e., they apply to all types of AI systems, not just frontier models) along the main phases of a system’s development. These guidelines aim to: a secure design, secure development, secure deployment, and secure operation and maintenance. The report includes suggestions about mitigations to related risks in the view that providers of AI components should take responsibility for the security outcomes of users further down the supply chain.

The Global Partnership on AI and the Ministerial Declaration. The goal of the Global Partnership on AI is to guide the responsible development and use of AI, grounded in human rights, inclusion, diversity, innovation and economic growth. ¹¹¹ On occasion of the December 2023 general-purpose AI summit in Delhi, Ministers of the 29 member countries signed a join declaration reafirming their commitment to promote responsible and trustworthy AI, and their dedication to jointly develop regulations, policies, and standards to uphold general-purpose AI’s principles. The Ministers also embraced the notion of “collaborative AI,” which involves supporting and promoting equitable access to critical resources for AI research and innovation, such as AI computing, high-quality diverse datasets, algorithms, software, test-beds etc.¹¹²

5.1 Open questions and gaps in the literature

The emerging literature on AI firmly established the pervasive and potentially transformative role of new technologies based on machine learning, highlighting that all occupations will be affected. However, the limit to our knowledge on its effects remain substantial. In this section we discuss our knowledge gaps progressively broadening the perspective from the labor market, to economy-wide impacts, to development and cross-country implications.

Lack of Consensus on AI Automation Effects. The consequences of AI for the labor market depend on its automating and complementing potential. First, let us consider the automating character of AI. Theory suggests that the effects of labor-substituting automation technology are ambiguous, but they can be determined through the knowledge of some key parameters and data moments. For example, Acemoglu and Restrepo (2022) derive formulas to obtain the net impact on wages and labor demand for different demographic groups. In that framework, wage effects can be directly computed once two quantities are known for occupation and sector: the share of tasks substituted by AI, and an estimate of the resulting cost savings. In this review, we covered several studies trying to provide an estimate of the former through task exposure measures. Although the correlation between different measures is usually positive (see, e.g., Cazzaniga et al. (2024) for a discussion), there is substantial disagreement on whether to interpret exposures as an indicator of potential task displacement. The comparability of different indicators is also in question, as some are indices and can only be interpreted as capturing relative task exposure, while others directly report a share of tasks that would be directly affected by the introduction of AI. Getting the indicators to speak to each other and producing related sufficient statistics for employment and productivity effects would be a necessary first step to determine the overall impact of AI on wages and employment. For example, instead of reporting “task exposures,” studies could report bounds on the share of employment–at current demand conditions and wages–that could be displaced by AI in each occupation as well as in the overall economy (as in, e.g., Briggs and Kodnani (2023) and Company (2023)). In this context, assumptions on complementarity between AI and tasks carried out by workers should be clearly stated. Ideally, the lower and upper bound of estimated employment effects should be interpretable as arising from the maximum and minimum degree of complementarity, respectively. These numbers could then be used to reach a general equilibrium employment estimates accounting for input-output linkages and worker mobility, as in Acemoglu and Restrepo (2022).

The Need for Task-Level and Sector-Level Productivity Data. The lack of precise productivity estimates further complicates the analysis of employment effects. As discussed in Section 3, available estimates cover specific technologies and contexts and cannot be easily generalized. In the narrow field of gen-AI, research has explored the productivity impact of text-generating AI, while it is silent on image generation, event though the latter has a significant effect on labor market outcomes of affected workers (Hui, Reshef, and Zhou 2023). The debate is still ongoing on how much AI will augment workers. The evidence in this paper points to important areas of complementarity and in this case, exposure measures are not sufficient to assess relative wage and employment effects. Instead, we would need an assessment of the measure of tasks that are complementary or substituting. This could take the form of a two-dimensional exposure measure–that is, measuring at the same time the share of augmented and substituted tasks–similar to Kogan et al. (2023). A similar indicator has however substantially higher data requirement than a simple exposure measure, since it would need to provide information on how much tasks are augmented by AI in addition to which tasks are affected. Just as noted by Restrepo (2023) for the broader automation literature, the key data gap appears to be the lack of task-level data, which would offer information on which tasks are carried out to produce each good and service in establishments with different AI technologies. Efforts to collect this data at scale have a notable historical precedent: the 1899 “Hand and Machine Labor” study, commissioned by the US congress to assess the impact of mechanization on labor (Atack, Margo, and Rhode 2019). The above limitations pose a severe constraint to any study trying to assess the economy-wide consequences of AI progress and adoption. In addition to productivity and task-level data, we would need to further assess the implications of the technology for different economic sectors. The data on occupational exposure only gives us a limited idea of supply-side impacts of the technology. However, we are still uncertain on which sectors’ products or markets are likely to benefit the most, and which industries stand to expand or contract. This aspect would be key to assess capacity constraints and skills gap to be filled, as well as retraining needs for potentially displaced workers.

Knowledge Gaps on Growth and Development Issues. The lack of productivity estimates is even more blatant when it comes to general ML and several crucial applications that have now been in use for years around the world. AI-powered precision agriculture stands out as a field for which tightly identified estimates of economic outcomes–such as value added or employment–would be highly beneficial.¹¹⁷ Such estimates would be key to evaluate the technology’s impact on EMDEs, where dis-employment effects are expected to be limited. In addition to sectoral implications, we remain unaware of the actual impact of AI on health and education. These points deserve a separate mention for two reason. First, both fields could further augment growth rates, through their impact on demographics and human capital, respectively. Second, these areas represent crucial gaps in the attainment of sustainable development goals (SDGs) in EMDEs.¹¹⁸ In this sense, experimental findings on the potential contribution of AI to reducing the needed pupil to teacher or patient to doctor ratio would be very valuable. On the theme of EMDEs, a major knowledge gap pertains to the international consequences of AI adoption. Korinek and Stiglitz (2021) and Alonso et al. (2022) study this question theoretically, but we have little empirical guidance on this topic. While we still lack data to evaluate this question, a first assessment may come from bridging theory with empirical evaluation of previous waves of technology that were simultaneously complementing and substituting. A promising candidate for this exercise could be the spillover of past IT adoption in advanced economies on EMDEs and its consequences on capital flows and factor allocations.

5.2 Policy and Regulation

In this part, we list potential issues facing AI regulators: the length of approval and implementation lags compared to the speed of technological advancements; the effects of regulation on sectors and firms; the seeming disconnect between policy and research; the need for data to inform the debate; and the potential role of unions and labor standards.

Regulation and the Speed of Technological Advancements. Several national and supranational authorities are currently either considering or structuring their AI regulation acts. In Section 4, we have highlighted how these initiatives vary widely both in coverage and scope. In addition, differences arise in the timing and speed of adoption. The EU legislative process started in mid-2021. A final act may follow in early 2024, and its complete application would then occur over the next two to three years. The US instead opted for a decentralized setup with involves Agencies and Departments with deadlines by mid-2024. These processes paint a stark contrast between the time needed to reach an agreement compared and the much faster speed of AI technologies. By the time they are finally implemented, regulations might have already become obsolete, unless they allow for a degree of flexibility in their definitions and coverage. Initiatives of self-regulation by AI companies may fill part of this gap.¹¹⁹

Effects of Regulation on Firms and Sectors. Several knowledge gaps persist on the effects of regulations on firm location choices and sectors. Indeed, regulation may push firms to locate where the regulations are less strict, with major repercussions on a country’s ability to compete on the international markets and effects on its internal labor market as well. It is also theoretically unclear whether regulations should include sector-specific clauses or exceptions. These may relate to the strategic nature of specific sectors in supply chains or input-output networks, shielding of “national champions,” or worker protection concerns. In general, regulation might reduce adoption of AI technologies (Lee et al. 2019) and innovation overall (Aghion, Bergeaud, and Van Reenen 2021). Analyses of firm-level and sectoral effects of new AI regulations appear necessary to guide the research community and policymakers on these aspects.

Policy and Research. Another aspect to consider is the potential disconnect between policy and regulations with research. The main regulations currently under consideration already envision at least some degree of collaboration with research. However, considering the rapid pace of the AI technologies, more exchanges between policy and research may be necessary to understand and measure the potential risks and impacts of AI adoption, and to foster innovation in this field.

Within the new EU AI governance emerging from the Trilogue agreement and the Provisional Agreement, the AI Office (of the EU Commission) will include a scientific panel of independent experts with the goal to advise about general-purpose AI models and to monitor their possible material safety risks. An Advisory Forum—including industry representatives, small and medium enterprises, start-ups, civil society, as well as academia—is also envisioned to provide technical expertise to the AI Board. The same version of the EU AI Act also states that the provisions of the Act would not apply to AI systems used for the sole purpose of research and innovation. In the US, promoting innovation and research is one of the eight principles guiding the EO. Practical provisions include expedited visas for noncitizens who seek to travel to the US to work on, study, or conduct research in AI. Further, the EO envisions a program to identify and attract top talent in AI at universities, research institutions, and the private sector, and to establish and increase connections overseas. The EO is also set to establish at least four new National AI Research Institutes and to produce a publicly available report on the potential role of AI in research aimed at tackling major societal and global challenges. The US Senate AI fora, which included experts in different fields (see Section 4.3.2) have been an important step to bridge policy and research to understand risks and opportunities of AI.. Finally, the Chinese regulation explicitly support the coordination between private sectors, education and research institutions, and public bodies in areas such as innovation in gen-AI technology, data resources, applications, and risk prevention.

Data Availability. The scarcity of updated data on the last generation of AI technologies stands out as a key hindrance to both regulators and researchers. Surveys and data on AI, its adoption, and progress appear highly critical. In addition to more data, it would be important to establish priorities on the types of data needed for policy and research. For example, additional insights would be useful on the market structure of chip producing sectors, trade linkages, and the concentration of exports and raw materials. Further, we have only scant data on the increasing energy requirements of AI data centers, which may impact energy markets through prices and investments in additional power sources like nuclear plants.¹²⁰ The OECD is currently trying to fll some gaps in the data, providing (or planning to cover) a wide set of AI-related information across countries and over time. Some example of datasets included are: AI news around the world and AI search trends; characteristics of AI developers and AI courses offered worldwide; trends in the demand and supply of AI talents, scientific publications, and patents; R&D funding and private equity investments in AI start-ups; AI models, software developments, and datasets from open-source platforms.¹²¹ The OECD also keeps track of regulations in the world based on their principles for the responsible stewardship of trustworthy AI released in 2021, with the latest report issued in May 2023.¹²²,¹²³

Additionally, the project “Our World in Data” (Giattino et al. 2023) now includes aspects related to AI in their interactive charts, such as the language and image recognition capabilities of AI, AI-related patents and new companies, or market share of AI hardware production by countries. The project also collects data from various external sources. Most datasets are updated to 2021.

On surveys, Van Noorden and Perkel (2023) look at researchers on their relationship with AI. They show that researchers are already adopting gen-AI, particularly for writing code, research brainstorming, writing manuscripts and conducting literature reviews.¹²⁴ More surveys could bring a better sense of possible impacts of AI, for example covering firms and their adoption of AI and future plans. Looking at small and medium enterprises in a wider set of countries and regions may prove particularly fruitful.

Labor Markets and Trade Unions. The potential role of unions and labor standards in labor market regulations is another area that stands out as needing further investigation. This is especially true in light of concrete actions taken by affected workers’ bodies. The United Auto Workers and the Writers Guild of America (WGA) included AI in their labor deal, giving them a say in how they use AI, and establishing a precedent for AI to be the subject of bargaining.¹²⁵ In particular, the agreements did not aim to ban AI, but rather demanded a share of the productivity gains from AI.¹²⁶ Among regulations, the US EO includes this aspect in its eight principle on supporting the American workforce: “As AI creates new jobs and industries, all workers need a seat at the table, including through collective bargaining, to ensure that they benefit from these opportunities.” At the same time, the EO does not provide binding measures or guidelines. Instead, it delegates the Department of Labor to issue guidance to employers that deploy AI to monitor or augment employees’ work on how to comply with existing legal requirements.¹²⁷

5.3 Key Takeaways

Takeaway 1: AI Effects Are Uncertain. Task exposure measures and the theoretical growth literature agree that the impact of AI will be extensive. However, the empirical evidence on employment effects and productivity provides wide estimates of the effects, and remains ambiguous on their direction. We believe that a key reason for this uncertainty resides in the limited comparability across measures, as well as the lack of crucial task-level, sector-level, and productivity data. Uncertainty is complicated by the scarcity of data on latest applications, such as gen-AI.

Takeaway 2: Policy and Research Are Partly Disconnected. At the date of writing, the academic literature has focused primarily on employment and productivity effects. By contrast, binding policy actions around the world have been mostly more concerned with ethical and bias concerns, privacy, and safety issues, which, with few notable exceptions, fall outside the scope of economics. The reluctance of regulators to tackle economic issues may be partly motivated by the uncertainty surrounding estimates of AI effects. At the same time, to our knowledge, the academic literature has not provided empirical investigations of key economic issues that are of interest to regulators. In this respect, more research is needed to understand the impact of AI on market concentration—from the competitive landscape of developers and providers to the effects on competition in adopting sectors—, trade networks and linkages in the AI supply chain, and intellectual property. Namely, the effects of AI on incentives to content creation remain unexplored. As discussed above, there are several ongoing initiatives that may contribute to bridging the gap between policy and research.

Takeaway 3: Regulations Differ Widely and Face Difficult Trade-Offs. We have seen how regulators across the world have so far adopted different—or outright contrasting—approaches and covered different areas. Some authorities preferred an ex-ante risk-based approach, other focused on ex-post liability, while some countries preferred a “pro-innovation,” regulation-free stance focused on guidelines and best practices. Similarly, there is substantial disagreement on what should be regulated. Efforts for international coordination have not yet changed this landscape. We believe that persistent differences across countries stem from difficult trade-offs that countries face when choosing to regulate AI. Many countries see the development of vibrant domestic AI sectors as strategic, which comes with two concerns. First, governments are reluctant to burden developers with regulation that may stifle innovation, as many countries acknowledge that first-mover advantages in the technology may be large. Second, regulators may face a trade-of between safety and attracting AI investments, as companies may choose to localize where policies are looser. Closer multilateral cooperation in development and regulation would go a long way toward addressing these issues.