AI4People—An Ethical Framework for a Good AI Society: Opportunities, Risks, Principles, and Recommendations

This section introduces the Recommendations for a Good AI Society. It consists of two parts: a Preamble, and 20 Action Points.

There are four kinds of Action Points: to assess, to develop, to incentivise and to support. Some recommendations may be undertaken directly, by national or European policy makers, in collaboration with stakeholders where appropriate. For others, policy makers may play an enabling role for efforts undertaken or led by third parties.

Preamble

We believe that, in order to create a Good AI Society, the ethical principles identified in the previous section should be embedded in the default practices of AI. In particular, AI should be designed and developed in ways that decrease inequality and further social empowerment, with respect for human autonomy, and increase benefits that are shared by all, equitably. It is especially important that AI be explicable, as explicability is a critical tool to build public trust in, and understanding of, the technology.

We also believe that creating a Good AI Society requires a multistakeholder approach, which is the most effective way to ensure that AI will serve the needs of society, by enabling developers, users and rule-makers to be on board and collaborating from the outset.

Different cultural frameworks inform attitudes to new technology. This document represents a European approach, which is meant to be complementary to other approaches. We are committed to the development of AI technology in a way that secures people’s trust, serves the public interest, and strengthens shared social responsibility.

Finally, this set of recommendations should be seen as a “living document”. The Action Points are designed to be dynamic, requiring not simply single policies or one-off investments, but rather, continuous, ongoing efforts for their effects to be sustained.

Action Points

Assessment

• Assess the capacity of existing institutions, such as national civil courts, to redress the mistakes made or harms inflicted by AI systems. This assessment should evaluate the presence of sustainable, majority-agreed foundations for liability from the design stage onwards, in order to reduce negligence and conflicts (see also Recommendation 5).

• Assess which tasks and decision-making functionalities should not be delegated to AI systems, through the use of participatory mechanisms to ensure alignment with societal values and understanding of public opinion. This assessment should take into account existing legislation and be supported by ongoing dialogue between all stakeholders (including government, industry, and civil society) to debate how AI will impact society opinion (in concert with Recommendation 17).

• Assess whether current regulations are sufficiently grounded in ethics to provide a legislative framework that can keep pace with technological developments. This may include a framework of key principles that would be applicable to urgent and/or unanticipated problems.

Development

• Develop a framework to enhance the explicability of AI systems that make socially significant decisions. Central to this framework is the ability for individuals to obtain a factual, direct, and clear explanation of the decision-making process, especially in the event of unwanted consequences. This is likely to require the development of frameworks specific to different industries, and professional associations should be involved in this process, alongside experts in science, business, law, and ethics.

• Develop appropriate legal procedures and improve the IT infrastructure of the justice system to permit the scrutiny of algorithmic decisions in court. This is likely to include the creation of a framework for AI explainability as indicated in Recommendation 4, specific to the legal system. Examples of appropriate procedures may include the applicable disclosure of sensitive commercial information in IP litigation, and—where disclosure poses unacceptable risks, for instance to national security—the configuration of AI systems to adopt technical solutions by default, such as zero-knowledge proofs in order to evaluate their trustworthiness.

• Develop auditing mechanisms for AI systems to identify unwanted consequences, such as unfair bias, and (for instance, in cooperation with the insurance sector) a solidarity mechanism to deal with severe risks in AI-intensive sectors. Those risks could be mitigated by multistakeholder mechanisms upstream. Pre-digital experience indicates that, in some cases, it may take a couple of decades before society catches up with technology by way of rebalancing rights and protection adequately to restore trust. The earlier that users and governments become involved—as made possible by ICT—the shorter this lag will be.

• Develop a redress process or mechanism to remedy or compensate for a wrong or grievance caused by AI. To foster public trust in AI, society needs a widely accessible and reliable mechanism of redress for harms inflicted, costs incurred, or other grievances caused by the technology. Such a mechanism will necessarily involve a clear and comprehensive allocation of accountability to humans and/or organisations. Lessons could be learnt from the aerospace industry, for example, which has a proven system of handling unwanted consequences thoroughly and seriously. The development of this process must follow from the assessment of existing capacity outlined in Recommendation 1. If a lack of capacity is identified, additional institutional solutions should be developed at national and/or EU levels, to enable people to seek redress. Such solutions may include:

Whichever solutions are developed, these are likely to rely on the framework for intelligibility proposed in Recommendation 4.

• Develop agreed-upon metrics for the trustworthiness of AI products and services, to be undertaken either by a new organisation, or by a suitable existing organisation. These metrics would serve as the basis for a system that enables the user-driven benchmarking of all marketed AI offerings. In this way, an index for trustworthy AI can be developed and signalled, in addition to a product’s price. This “trust comparison index” for AI would improve public understanding and engender competitiveness around the development of safer, more socially beneficial AI (e.g., “IwantgreatAI.org”). In the longer term, such a system could form the basis for a broader system of certification for deserving products and services, administered by the organisation noted here, and/or by the oversight agency proposed in Recommendation 9. The organisation could also support the development of codes of conduct (see Recommendation 18). Furthermore, those who own or operate inputs to AI systems and profit from it could be tasked with funding and/or helping to develop AI literacy programs for consumers, in their own best interest.

• Develop a new EU oversight agency responsible for the protection of public welfare through the scientific evaluation and supervision of AI products, software, systems, or services. This may be similar, for example, to the European Medicines Agency. Relatedly, a “post-release” monitoring system for AIs similar to, for example, the one available for drugs should be developed, with reporting duties for some stakeholders and easy reporting mechanisms for other users.

• Develop a European observatory for AI. The mission of the observatory would be to watch developments, provide a forum to nurture debate and consensus, provide a repository for AI literature and software (including concepts and links to available literature), and issue step-by-step recommendation and guidelines for action.

• Develop legal instruments and contractual templates to lay the foundation for a smooth and rewarding human–machine collaboration in the work environment. Shaping the narrative on the ‘Future of Work’ is instrumental to winning “hearts and minds”. In keeping with ‘A Europe that protects’, the idea of “inclusive innovation” and to smooth the transition to new kinds of jobs, a European AI Adjustment Fund could be set up along the lines of the European Globalisation Adjustment Fund.

Incentivisation

• Incentivise financially, at the EU level, the development and use of AI technologies within the EU that are socially preferable (not merely acceptable) and environmentally friendly (not merely sustainable but favourable to the environment). This will include the elaboration of methodologies that can help assess whether AI projects are socially preferable and environmentally friendly. In this vein, adopting a ‘challenge approach’ (see DARPA challenges) may encourage creativity and promote competition in the development of specific AI solutions that are ethically sound and in the interest of the common good.

• Incentivise financially a sustained, increased and coherent European research effort, tailored to the specific features of AI as a scientific field of investigation. This should involve a clear mission to advance AI for social good, to serve as a unique counterbalance to AI trends with less focus on social opportunities.

• Incentivise financially cross-disciplinary and cross-sectoral cooperation and debate concerning the intersections between technology, social issues, legal studies, and ethics. Debates about technological challenges may lag behind the actual technical progress, but if they are strategically informed by a diverse, multistakeholder group, they may steer and support technological innovation in the right direction. Ethics should help seize opportunities and cope with challenges, not only describe them. It is essential in this respect that diversity infuses the design and development of AI, in terms of gender, class, ethnicity, discipline and other pertinent dimensions, in order to increase inclusivity, toleration, and the richness of ideas and perspectives.

• Incentivise financially the inclusion of ethical, legal and social considerations in AI research projects. In parallel, incentivise regular reviews of legislation to test the extent to which it fosters socially positive innovation. Taken together, these two measures will help ensure that AI technology has ethics at its heart and that policy is oriented towards innovation.

• Incentivise financially the development and use of lawfully de-regulated special zones within the EU for the empirical testing and development of AI systems. These zones may take the form of a “living lab” (or Tokku), building on the experience of existing “test highways” (or Teststrecken). In addition to aligning innovation more closely with society’s preferred level of risk, sandbox experiments such as these contribute to hands-on education and the promotion of accountability and acceptability at an early stage. “Protection by design” is intrinsic to this kind of framework.

• Incentivise financially research about public perception and understanding of AI and its applications, and the implementation of structured public consultation mechanisms to design policies and rules related to AI. This may include the direct elicitation of public opinion via traditional research methods, such as opinion polls and focus groups, as well as more experimental approaches, such as providing simulated examples of the ethical dilemmas introduced by AI systems, or experiments in social science labs. This research agenda should not serve merely to measure public opinion, but should also lead to the co-creation of policies, standards, best practices, and rules as a result.

Support

• Support the development of self-regulatory codes of conduct for data and AI related professions, with specific ethical duties. This would be along the lines of other socially sensitive professions, such as medical doctors or lawyers, i.e., with the attendant certification of ‘ethical AI’ through trust-labels to make sure that people understand the merits of ethical AI and will therefore demand it from providers. Current attention manipulation techniques may be constrained through these self-regulating instruments.

• Support the capacity of corporate boards of directors to take responsibility for the ethical implications of companies’ AI technologies. For example, this may include improved training for existing boards and the potential development of an ethics committee with internal auditing powers. This could be developed within the existing structure of both one-tier and two-tier board systems, and/or in conjunction with the development of a mandatory form of “corporate ethical review board” to be adopted by organisations developing or using AI systems, to evaluate initial projects and their deployment with respect to fundamental principles.

• Support the creation of educational curricula and public awareness activities around the societal, legal, and ethical impact of Artificial Intelligence. This may include: