Institute for a Greater EuropeEthical Considerations of European Union Artificial Intelligence Policy and Regulatory Sandboxes
Introduction
In June 2022, the European Parliament released a new briefing on artificial intelligence (AI) regulation.[1] The report was focused primarily on the role of so-called “regulatory sandboxes” as tools for ethical oversight. These regulatory sandboxes would build upon the process of rule harmonisation within the European Union (EU) on AI under the proposed Artificial Intelligence Act. This also expands upon previous actions taken within the EU whereby the European Commission outlined rules for “responsible” AI in April 2021.[2]
In this piece, I will begin by detailing what regulatory sandboxes are and their proposed utility. Following this, I will discuss common ethical concerns at the centre of AI regulation and technological development. I then discuss the novel possibilities which could be provided by the proposed regulatory sandbox approach to AI regulation within the EU. I move, then, to note the remaining ethical challenges regarding the regulatory sandbox approach. Finally, I conclude by discussing future possibilities for addressing ethical concerns regarding the use of AI technology within the EU.
What are regulatory sandboxes?
Regulatory sandboxes are described by the European Parliament as “a tool allowing businesses to explore and experiment with new and innovative products, services or businesses under a regulator's supervision.”[3] In a report by the European Institute of Public Administration (EIPA),[4] the organisation refers to a proposal for the Artificial Intelligence Act from April, 2021 where Article 54 states that regulatory sandboxes for AI, “shall provide a controlled environment that facilitates the development, testing and validation of innovative AI systems for a limited time.”[5] Thus, regulatory sandboxes are best thought of as limited-release testing groups whereby new AI technology would be used by a smaller number of individuals than the full population under more highly scrutinised conditions with the aim of catching widespread issues with the technology before utilising it in a less controlled setting. The regulatory sandbox approach would establish a process in which the regulating authority, in this case, associated with the EU, would create a space for participants, part of civil society, to interact with new AI technologies developed by business entities.[6]
Ethics of AI
Developments in AI raise several significant ethical questions. As autonomous systems are able to increasingly access, collect, and leverage troves of personal information and are tasked with making decisions in place of humans, these changes can bring concerns, particularly regarding their relationship to individuals and their information. The expansion of the power of AI across both private industry and public organisations makes the pertinence of placing greater scrutiny on its appropriate ethical dimensions considerable. Three central ethical considerations regarding the technological development of artificial intelligence include privacy, discrimination, and human judgement.[7] As these three considerations are particularly pertinent, in this section I will focus specifically on elaborating on the main ethical issues which are raised regarding AI and possible concerns in these areas.
Privacy
First, the expansion of AI technology has relied on the collection of massive amounts of data, with autonomous and semi-autonomous AI technology being employed in various sectors such as healthcare, transportation, and production.[8] The protection of privacy in ethics is considered to be the protection of individual choice to decide who or what enters the personal life space on oneself. The growth of AI is directly related to questions of privacy.
For one, the methods through which companies are able to collect this data can raise ethical concerns regarding these processes through AI. There are various possibilities through which privacy violations may occur through AI technologies. Data persistence, when data lasts longer than the humans who created it, data repurposing when data is used for reasons other than its original purpose, and data spillovers, when data is collected on people which were not the intended target. As a significant, and growing, amount of privacy-sensitive data analysis is being driven by machine learning, the relationship between how data analysis through AI technologies occurs and individual choice to enter into personal life space exists should be pertinent for consideration of a regulatory sandbox approach to AI regulation. This significant and growing insertion of AI technologies into everyday life can also raise concerns of bias and discrimination, as I will cover in the next section.
Bias and Discrimination
Besides privacy concerns, the hope for AI technology is that it will reduce the bias that is inherent to human decision-making, the reality is that such products have been found to produce biased or discriminatory decisions. Bias refers to decisions which systematically favour or disfavour one group, or groups, over another group, or groups, when there is no justifiable reason for favour or disfavour to occur. Discrimination is the product of bias against a certain group, whether direct or indirect.
Bias has occurred, for example, when technology used by Amazon to provide autonomous insights in hiring decisions resulted in bias against women, while algorithms in facial recognition programs used by a number of companies, such as IBM and Microsoft, displayed bias on the basis of skin colour.[9]
UNESCO has also released a recommendation on the ethics of AI usage, suggesting that while these technologies can be a great service to humanity, they also raise significant concerns regarding bias and discrimination.[10] Specifically, the organisation notes that this can be a by-product of the expanding number of tasks which are conducted by AI technology. The algorithms underlying developing AI must be adapted to carrying out these new tasks, while also interacting in the human environment in a way in which they do not reproduce or exacerbate programmed biases. This distinction between service to humanity and relationship with humans is also a meaningful consideration in regard to the role of human judgement, which is the focus of the next section.
Human Judgement
As a third issue, one of the central features of advancements in AI technology is that it can make semi-autonomous or fully autonomous decisions within the need for humans to take the time and effort to make these decisions. Yet the design and implementation of AI technology tend to require significant human insights and decisions regarding the development of these products. Taken, for example, autonomous driving vehicles. If the vehicle has been programmed to drive autonomously, it must also be programmed to make decisions in the case of an imminent crash. As a result of this, the decision for how an autonomous vehicle will react, and who it would be more likely to harm, must be made with the input of human judgement. This consideration, the relationship between human judgement in programming AI in decision making, can apply to other situations, particularly as the use of AI expands. After having discussed the ethical concerns, the next part will elaborate on the novel possibilities emerging from regulatory sandboxes.
Novel possibilities
The most promising aspect of regulatory sandboxes is the hope that this would create a better regulatory environment for consumers of AI technology and the public use of autonomous technologies. Previous studies have argued that strict liability regimes, which are designed to provide recoveries for damages and make develops liable for all damages they have found to be caused, have a “chilling effect” on Small and Medium Enterprises (SMEs), and regulatory sandboxes would instead provide greater support to SMEs by providing a different environment for regulation of AI technologies.[11]
Regulatory sandboxes have also been previously used more widely regarding financial technology (fintech). The World Bank released a report in 2020 noting that regulatory sandboxes had become synonymous with fintech and provided insights into how regulatory sandboxes had been applied to fintech to that point.[12] The report found that regulatory sandboxes could help build consensus among stakeholders, allow new firms to enter the market, and build regulator knowledge on fintech trends and innovations, among other insights.
In the European Parliament briefing, they also note how regulatory sandboxes may be especially beneficial for consumers, as this approach can support the introduction of both newer and safer products.[13] The European Commission’s draft of the AI act in April 2021 stated the importance of establishing trust regarding the safe use of AI technologies and respect for fundamental rights.[14] Regulatory sandboxes could be especially useful for high-risk applications of AI technologies where they can be tested within more controlled environments with better safeguards for participants.[15]
Taken together, regulatory sandboxes may allow for new AI technology to better support SMEs, help drive innovation in certain sectors, and create safer products. Such outcomes would be useful in developing trust in the regulatory sandbox approach and be better for both participants and the general public if AI technologies are later allowed and released.
Ethical Challenges
Despite the novel possibilities which come with the proposed approach of regulatory sandboxes from the EU regarding AI technology, ethical difficulties are likely to remain with the use of this approach. The first round of the SHERPA Delphi study conducted by the EU, an EU-funded survey designed by the SHERPA consortium sent out to 231 experts to analyse ethical and human rights issues in smart information systems, found several ethical concerns raised, but three of the most frequently cited concerns were a lack of privacy, bias and discrimination, and loss of human decision-making.[16]
First, regarding privacy, AI development expands the way in which personal information is able to be collected and used in ways which can invade people’s privacy. Regulatory sandboxes do not necessarily address questions of privacy, and the recent European Parliament briefing did not mention information privacy or machine learning. AI-based machine learning in particular can raise particular ethical concerns, requiring access to data and large datasets, meaning that how and when this data is collected must be seriously considered.[17] The European Parliament briefing does not provide detail on privacy concerns as well in the application of regulatory sandboxes, which will be a significant concern regarding AI technology.
Further, properly addressing bias and discrimination will be a central concern of AI regulation. Regulatory sandboxes should provide some advantages to address these ethical concerns. The existing focus on the use of regulatory sandboxes for the development of fintech has suggested that regulatory sandboxes could be modelled through a system whereby a committee of regulators can decide to admit firms and then ban or conditionally approve new AI technologies after determinations have been made based on application within the regulatory sandbox environment.[18]
Additionally, to reduce consumer harm, a policy report from Nicol Turner Lee, Paul Resnick, and Genie Barton at the Brookings Institution has argued for an “algorithmic hygiene” approach to addressing bias and discrimination, which focuses on the understanding context during the early stages of bias detection and comparing outcomes for different groups at the outset to better capture and address discriminatory outcomes in AI technological development.[19] Regulatory sandboxes should thus be concerned with how regulators make decisions to ban or approve specific developments in AI products, as well as ensure that early-stage consideration of context and group comparisons are conducted to better capture possible unjustified biases in expanding AI technology.
Finally, regulatory sandboxes struggle to answer questions of human judgement in the programming of AI. People’s moral decisions are biased by personal perspectives and modes of decision-making.[20] This means that the design of AI products is, to a certain extent, inherently connected to the moral decisions of humans who design and interact with these technologies. Simultaneously, the expansion of AI technology blurs the line between human judgement regarding ethical choices and AI making judgements through its algorithmic function.[21] This raises questions of how judgements themselves should be made regarding the programming of technological products which have implications for interaction with and possible harm to human beings.
Concluding Remarks
Summarizing the main arguments discussed here, regulatory sandboxes are emerging as an expected aspect of the EU approach towards AI technological development in coming years. In the EU, the General Data Protection Regulation (GDPR), applied in all EU member states since May 25, 2018, created significant changes in data protection law within the EU. It has been noted in existing work on the ethical implications of AI that these laws are broader in scope in defining relevant data in areas such as healthcare than in other places, such as the United States.[22] The broader application of EU law to aspects of data collection can mean that regulatory sandboxes must provide particular focus to legal issues as well when considering future ethical dimensions. As mentioned in the European Parliament briefing as well, regulatory arbitrage, the lowering of regulatory safeguards to attract innovators, and fragmentation of the EU single market are also possible consequences.[23]
The recent European Parliament briefing on regulatory sandboxes, along with the European Commission AI act proposal, displays the likelihood that this will be the primary approach to the regulation of AI development in the EU in the coming years. Simultaneously, AI is expected to expand its reach to various areas of human life through public and private applications. In addition to the practical and legal considerations of this approach, ethical issues should be central to assessing the use of regulatory sandboxes for AI technology.
Regulatory sandboxes may limit the initial issues with an AI-based technology, but once released from sandbox oversight new issues may arise which are no longer addressed by regulatory sandbox oversight. This leads to a question of feasibility regarding how broadly regulatory sandboxes can be used to improve ethical issues in AI and how subsequent developments of products previously housed under regulatory sandbox oversight will continue to address questions of privacy, bias and discrimination, and human judgement in AI technological development.
Regulatory sandboxes would ideally take early-stage consideration of establishing standards for privacy and testing the alignment of new AI technologies with these standards before full release. A similar approach to capturing instances of systematic bias in AI products through group comparisons, along with an understanding of context and how to ban or conditionally approach new technologies, could reduce ethical concerns about unjustified differentiation of groups through these products. Addressing ethical concerns of human judgement will be particularly difficult, as regulatory sandboxes do little to inherently reduce these concerns. Such an approach, however, could aim to, as with privacy, set standards which are generally considered to be broadly agreeable and protect people from harm in the expansion of AI.
Endnotes
[1] Madiega, Tambiama and Anne Louise Van De Pol, “Artificial intelligence act and regulatory sandboxes,” European Parliament: European Parliamentary Research Service, (2022), available at https://www.europarl.europa.eu/RegData/etudes/BRIE/2022/733544/EPRS_BRI(2022)733544_EN.pdf.
[2] European Commission, “A European approach to artificial intelligence”, European Commission, (2022), available at https://digital-strategy.ec.europa.eu/en/policies/european-approach-artificial-intelligence.
[3] Madiega and Van De Pol, supra note 1, at 1.
[4] Pop, Florina and Lukas Adomavicius, “Sandboxes for Responsible Artificial Intelligence”, European Institute of Public Administration, (2021), available at https://www.eipa.eu/wp-content/uploads/2022/03/EIPA-Briefing_Sandboxes-for-Responsible-Artificial-Intelligence.pdf.
[5] European Commission, “Proposal for a REGULATION OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL LAYING DOWN HARMONISED RULES ON ARTIFICIAL INTELLIGENCE (ARTIFICIAL INTELLIGENCE ACT) AND AMENDING CERTAIN UNION LEGISLATIVE ACTS”, COM(2021) 206 final, European Commission, (2021).
[6] Yordanova, Katarina, “The Shifting Sands of Regulatory Sandboxes for AI”, KU Leuven Centre for IT and IP Law, (2019), available at https://www.law.kuleuven.be/citip/blog/the-shifting-sands-of-regulatory-sandboxes-for-ai/.
[7] Pazzanese, Christina, “Ethical concerns mount as AI takes bigger decision-making role in more industries”, The Harvard Gazette, (2020), available at https://news.harvard.edu/gazette/story/2020/10/ethical-concerns-mount-as-ai-takes-bigger-decision-making-role/.
[8] Stead, William W. "Clinical implications and challenges of artificial intelligence and deep learning." Journal of the American Medical Association, 320, no. 11 (2018): 1107-1108.
[9] Kerry, Cameron, "Protecting privacy in an AI-driven world.", Brookings Institution, (2020), available at https://www.brookings.edu/research/protecting-privacy-in-an-ai-driven-world/.
[10] UNESCO, “Draft text of the Recommendation on the Ethics of Artificial Intelligence”, SHS/IGM-AIETHICS/2021/JUN/3 Rev.2, UNESCO, (2021), available at https://unesdoc.unesco.org/ark:/48223/pf0000377897.
[11] Truby, Jon, Rafael Dean Brown, Imad Antoine Ibrahim, and Oriol Caudevilla Parellada. "A Sandbox Approach to Regulating High-Risk Artificial Intelligence Applications." European Journal of Risk Regulation 13, no. 2 (2022): 270-294.
[12] Appaya, Sharmista and Mahjabeen Haji, “Four years and counting: What we’ve learned from regulatory sandboxes”, World Bank: Private Sector Development Blog, (2020), available at https://blogs.worldbank.org/psd/four-years-and-counting-what-weve-learned-regulatory-sandboxes.
[13] Madiega and Van De Pol, supra note 2, at 3.
[14] European Commission, supra note 1.
[15] Truby et al., supra note 1, at 274.
[16] SHERPA, "SHERPA Delphi Study—Round 1 Results ." Shaping the ethical dimensions of smart information systems– a European perspective (SHERPA), (2020), available at https://www.project-sherpa.eu/wp-content/uploads/2020/03/sherpa-delphi-study-round-1-summary-17.03.2020.docx.pdf.
[17] Stahl, Bernd Carsten. "Ethical issues of ai." In Artificial Intelligence for a Better Future, pp. 35-53. Springer, Cham, 2021.
[18] Allen, Hilary J. "Regulatory sandboxes." Geo. Wash. L. Rev. 87 (2019): 579.
[19] Lee, Nicol Turner, Paul Resnick, and Genie Barton. "Algorithmic bias detection and mitigation: Best practices and policies to reduce consumer harms." Brookings Institute: Washington, DC, USA (2019).
[20] Frank, Darius-Aurel, Polymeros Chrysochou, Panagiotis Mitkidis, and Dan Ariely. "Human decision-making biases in the moral dilemmas of autonomous vehicles." Scientific reports 9, no. 1 (2019): 1-19.
[21] Spaulding, Norman W. "Is Human Judgment Necessary? Artificial Intelligence, Algorithmic Governance, and the Law." The Oxford Handbook of Ethics and AI, (2020).
[22] Gerke, Sara, Timo Minssen, and Glenn Cohen. "Ethical and legal challenges of artificial intelligence-driven healthcare." In Artificial intelligence in healthcare, pp. 295-336. Academic Press, 2020.
[23] Madiega and Van De Pol, supra note 3.