In the past five years, private companies, research institutions and public sector organizations have issued principles and guidelines for ethical artificial intelligence (AI). However, despite an apparent agreement that AI should be ‘ethical’, there is debate about both what constitutes ‘ethical AI’ and which ethical requirements, technical standards and best practices are needed for its realization. To investigate whether a global agreement on these questions is emerging, we mapped and analysed the current corpus of principles and guidelines on ethical AI. Our results reveal a global convergence emerging around five ethical principles (transparency, justice and fairness, non-maleficence, responsibility and privacy), with substantive divergence in relation to how these principles are interpreted, why they are deemed important, what issue, domain or actors they pertain to, and how they should be implemented. Our findings highlight the importance of integrating guideline-development efforts with substantive ethical analysis and adequate implementation strategies.
Artificial intelligence (AI), or the theory and development of computer systems able to perform tasks normally requiring human intelligence, is widely heralded as an ongoing “revolution” transforming science and society altogether1,2. While approaches to AI such as machine learning, deep learning and artificial neural networks are reshaping data processing and analysis3, autonomous and semi-autonomous systems are being increasingly used in a variety of sectors including healthcare, transportation and the production chain4. In light of its powerful transformative force and profound impact across various societal domains, AI has sparked ample debate about the principles and values that should guide its development and use5,6. Fears that AI might jeopardize jobs for human workers7, be misused by malevolent actors8, elude accountability or inadvertently disseminate bias and thereby undermine fairness9 have been at the forefront of the recent scientific literature and media coverage. Several studies have discussed the topic of ethical AI10,11,12,13, notably in meta-assessments14,15,16 or in relation to systemic risks17,18 and unintended negative consequences such as algorithmic bias or discrimination19,20,21.
National and international organizations have responded to these concerns by developing ad hoc expert committees on AI, often mandated to draft policy documents. These committees include the High-Level Expert Group on Artificial Intelligence appointed by the European Commission, the expert group on AI in Society of the Organisation for Economic Co-operation and Development (OECD), the Advisory Council on the Ethical Use of Artificial Intelligence and Data in Singapore, and the Select Committee on Artificial Intelligence of the UK House of Lords. As part of their institutional appointments, these committees have produced or are reportedly producing reports and guidance documents on AI. Similar efforts are taking place in the private sector, especially among corporations who rely on AI for their business. In 2018 alone, companies such as Google and SAP publicly released AI guidelines and principles. Declarations and recommendations have also been issued by professional associations and non-profit organizations such as the Association of Computing Machinery (ACM), Access Now and Amnesty International. This proliferation of soft-law efforts can be interpreted as a governance response to advanced research into AI, whose research output and market size have drastically increased22 in recent years.
Reports and guidance documents for ethical AI are instances of what is termed non-legislative policy instruments or soft law23. Unlike so-called hard law—that is, legally binding regulations passed by the legislatures to define permitted or prohibited conduct—ethics guidelines are not legally binding but persuasive in nature. Such documents are aimed at assisting with—and have been observed to have significant practical influence on—decision-making in certain fields, comparable to that of legislative norms24. Indeed, the intense efforts of such a diverse set of stakeholders in issuing AI principles and policies is noteworthy, because they demonstrate not only the need for ethical guidance, but also the strong interest of these stakeholders to shape the ethics of AI in ways that meet their respective priorities16,25. Specifically, the private sector’s involvement in the AI ethics arena has been called into question for potentially using such high-level soft policy as a portmanteau to either render a social problem technical16 or to eschew regulation altogether26. Beyond the composition of the groups that have produced ethical guidance on AI, the content of this guidance itself is of interest. Are these various groups converging on what ethical AI should be, and the ethical principles that will determine the development of AI? If they diverge, what are their differences and can these differences be reconciled?
Our Perspective maps the global landscape of existing ethics guidelines for AI and analyses whether a global convergence is emerging regarding both the principles for ethical AI and the suggestions regarding its realization. This analysis will inform scientists, research institutions, funding agencies, governmental and intergovernmental organizations, and other relevant stakeholders involved in the advancement of ethically responsible innovation in AI.
We conducted a scoping review of the existing corpus of documents containing soft-law or non-legal norms issued by organizations. This included a search for grey literature containing principles and guidelines for ethical AI, with academic and legal sources excluded. A scoping review is a method aimed at synthesizing and mapping the existing literature27 that is considered particularly suitable for complex or heterogeneous areas of research27,28. Given the absence of a unified database for AI-specific ethics guidelines, we developed a protocol for discovery and eligibility, adapted from the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework29. The protocol was pilot-tested and calibrated prior to data collection. Following best practices for grey literature retrieval, a multi-stage screening strategy involving both inductive screening via search engine and deductive identification of relevant entities with associated websites and online collections was conducted. To achieve comprehensiveness and systematicity, relevant documents were retrieved by relying on three sequential search strategies (Fig. 1): first, a manual search of four link hub webpages (‘linkhubs’)30,31,32,33 was performed. Sixty-eight sources were retrieved, out of which 30 were eligible (27 after removing duplicates). Second, a keyword-based web search of the Google.com search engine was performed in private-browsing mode, after logging out from personal accounts and erasing all web cookies and history34,35. The search was performed using the following keywords: ‘AI principles’, ‘artificial intelligence principles’, ‘AI guidelines’, ‘artificial intelligence guidelines’, ‘ethical AI’ and ‘ethical artificial intelligence’. Every link in the first 30 search results was followed and screened (1) for AI principles, resulting in ten more sources after removing duplicates, and (2) for articles mentioning AI principles, leading to the identification of three additional non-duplicate sources. The remaining Google results up to the 200th listings for each Google search were followed and screened for AI principles only. Within these additional 1,020 link listings we identified 15 non-duplicate documents. After identifying relevant documents through the two processes described, we used citation chaining to manually screen the full texts and, if applicable, reference lists of all eligible sources in order to identify other relevant documents. Seventeen additional sources were identified. We continued to monitor the literature in parallel with the data analysis and until 23 April 2019 to retrieve eligible documents that were released after our search was completed. Twelve new sources were included within this extended time frame. To ensure theoretical saturation, we exhausted the citation chaining within all identified sources until no additional relevant document could be identified.
Based on our inclusion/exclusion criteria, policy documents (including principles, guidelines and institutional reports) included in the final synthesis were (1) written in English, German, French, Italian or Greek; (2) issued by institutional entities from both the private and the public sectors; (3) referred explicitly in their title/description to AI or ancillary notions; and (4) expressed a normative ethical stance defined as a moral preference for a defined course of action (Supplementary Table 1). Following full-text screening, 84 sources or parts thereof were included in the final synthesis (Supplementary Table 2).
Content analysis of the 84 sources was independently conducted by two researchers in two cycles of manual coding and one cycle of code mapping within the qualitative data analysis software NVivo for Mac version 11.4. During the first cycle of coding, one researcher exhaustively tagged all relevant text through inductive coding37, attributing a total of 3,457 codes, out of which 1,180 were subsequently discovered to pertain to ethical principles. Subsequently, two researchers conducted the code-mapping process in order to reduce subjective bias. The process of code mapping, a method for qualitative metasynthesis38, consisted of two iterations of theming37, whereby categories were first attributed to each code, then categorized in turn (Supplementary Table 3). For the theming of ethical principles, we relied deductively on normative ethical literature. Ethical categories were inspected and assessed for consistency by two researchers with primary expertise in ethics. Thirteen ethical categories emerged from code mapping, two of which were merged with others due to independently assessed semantic and thematic proximity. Finally, we extracted significance and frequency by applying focused coding, a second cycle coding methodology used for interpretive analysis37, to the data categorized in ethical categories. A consistency check was performed both by reference to the relevant ethics literature and a process of deliberative mutual adjustment among the general principles and the particular judgments contained in the policy documents, an analytic strategy known as reflective equilibrium39.
Our search identified 84 documents containing ethical principles or guidelines for AI (Tables 1 and 2). Data reveal a significant increase over time in the number of publications, with 88% having been released after 2016 (Supplementary Table 2). Data breakdown by type and geographic location of issuing organization (Supplementary Table 2) shows that most documents were produced by private companies (n = 19; 22.6%) and governmental agencies respectively (n = 18; 21.4%), followed by academic and research institutions (n = 9; 10.7%), intergovernmental or supranational organizations (n = 8; 9.5%), non-profit organizations and professional associations/scientific societies (n = 7 each; 8.3% each), private sector alliances (n = 4; 4.8%), research alliances (n = 1; 1.2%), science foundations (n = 1; 1.2%), federations of worker unions (n = 1; 1.2%) and political parties (n = 1; 1.2%). Four documents were issued by initiatives belonging to more than one of the above categories and four more could not be classified at all (4.8% each).
In terms of geographic distribution, data show a prominent representation of more economically developed countries, with the USA (n = 21; 25%) and the UK (n = 13; 15.5%) together accounting for more than a third of all ethical AI principles, followed by Japan (n = 4; 4.8%), Germany, France and Finland (each n = 3; 3.6% each). The cumulative number of sources from the European Union, comprising both documents issued by EU institutions (n = 6) and documents issued within each member state (13 in total), accounts for 19 documents overall. African and South-American countries are not represented independently from international or supranational organizations (Fig. 2).
Data breakdown by target audience indicates that most principles and guidelines are addressed to multiple stakeholder groups (n = 27; 32.1%). Another significant portion of the documents is self-directed, as they are addressed to a category of stakeholders within the sphere of activity of the issuer such as the members of the issuing organization or the issuing company’s employees (n = 24; 28.6%). Finally, some documents target the public sector (n = 10; 11.9%), the private sector (n = 5; 6.0%), or other specific stakeholders beyond members of the issuing organization, namely developers or designers (n = 3; 3.6%), “organizations” (n = 1; 1.2%) and researchers (n = 1; 1.2%). Thirteen sources (15.5%) do not specify their target audience (Supplementary Table 1).
Eleven overarching ethical values and principles have emerged from our content analysis. These are, by frequency of the number of sources in which they were featured: transparency, justice and fairness, non-maleficence, responsibility, privacy, beneficence, freedom and autonomy, trust, dignity, sustainability, and solidarity (Table 3).
No single ethical principle appeared to be common to the entire corpus of documents, although there is an emerging convergence around the following principles: transparency, justice and fairness, non-maleficence, responsibility and privacy. These principles are referenced in more than half of all the sources. Nonetheless, further thematic analysis reveals significant semantic and conceptual divergences in both how the 11 ethical principles are interpreted and the specific recommendations or areas of concern derived from each. A detailed thematic evaluation is presented in the following.
Featured in 73 of our 84 sources, transparency is the most prevalent principle in the current literature. Thematic analysis reveals significant variation in relation to the interpretation, justification, domain of application and mode of achievement. References to transparency comprise efforts to increase explainability, interpretability or other acts of communication and disclosure (Table 3). Principal domains of application include data use40,41,42,43, human–AI interaction40,44,45,46,47,48,49,50,51,52, automated decisions43,53,54,55,56,57,58,59,60,61,62,63 and the purpose of data use or application of AI systems41,44,64,65,66,67,68. Primarily, transparency is presented as a way to minimize harm and improve AI53,54,55,61,62,66,69,70,71,72, though some sources underline its benefit for legal reasons54,62,63,66,67,69 or to foster trust40,41,46,50,53,54,65,68,69,73,74,75. A few sources also link transparency to dialogue, participation and the principles of democracy47,58,66,67,69,76.
To achieve greater transparency, many sources suggest increased disclosure of information by those developing or deploying AI systems53,68,77,78, although specifications regarding what should be communicated vary greatly: use of AI62, source code48,69,79, data use52,64,67,75, evidence base for AI use74, limitations42,50,64,68,75,77,80, laws79,81, responsibility for AI57, investments in AI61,82 and possible impact83. The provision of explanations “in non-technical terms”43 or auditable by humans54,77 is encouraged. Whereas audits and auditability45,56,61,62,67,76,78,79,84,85 are mainly proposed by data protection offices and non-profit organizations, it is mostly the private sector that suggests technical solutions44,47,69,76,86,87. Alternative measures focus on oversight62,64,65,72,79, interaction and mediation with stakeholders and the public41,49,53,68,78,88 and the facilitation of whistleblowing53,77.
Justice, fairness and equity
Justice is mainly expressed in terms of fairness40,42,44,45,46,65,67,75,77,83,89,90,91,92,93,94, and of prevention, monitoring or mitigation of unwanted bias40,45,50,57,64,69,71,75,81,86,90,91,95,96,97 and discrimination45,50,53,55,61,62,67,72,73,77,85,98,99,100,101, the latter being significantly less referenced than the first two by the private sector. Whereas some sources focus on justice as respect for diversity48,55,73,76,82,83,87,89,95,97,102,103, inclusion48,62,64,68,89,97 and equality58,62,68,76,77,89,95, others call for a possibility to appeal or challenge decisions45,52,53,54,91,96, or the right to redress50,59,62,63,67,85,102 and remedy62,65. Sources also emphasize the importance of fair access to AI76,87,104, data50,54,61,84,100,105,106,107 and the benefits of AI54,55,97,108. Issuers from the public sector place particular emphasis on AI’s impact on the labour market54,55,72,101,109, and the need to address democratic50,55,76,90 or societal48,65,72,82 issues. Sources focusing on the risk of biases within datasets underline the importance of acquiring and processing accurate, complete and diverse data40,45,69,87,110, especially training data44,50,52,55,69,75.
If specified, the preservation and promotion of justice are proposed to be pursued through: (1) technical solutions such as standards67,85,106 or explicit normative encoding45,54,60,84; (2) transparency71,79, notably by providing information53,55,96 and raising public awareness of existing rights and regulation45,76; (3) testing69,75,84,86, monitoring71,73 and auditing56,63,67,84, the preferred solution of notably data protection offices; (4) developing or strengthening the rule of law and the right to appeal, recourse, redress or remedy54,55,59,62,63,65,85,91,96; and (5) via systemic changes and processes such as governmental action59,62,104,109 and oversight111, a more interdisciplinary64,82,102,110 or otherwise diverse75,76,87,102,104,112 workforce, as well as better inclusion of civil society or other relevant stakeholders in an interactive manner45,50,58,63,72,74,75,82,85,86,96,97,103 and increased attention to the distribution of benefits42,50,55,65,80,93.
References to non-maleficence occur significantly more often than references to beneficence and encompass general calls for safety and security97,107,113,114 or state that AI should never cause foreseeable or unintentional harm40,47,50,73,77,96. More granular considerations entail the avoidance of specific risks or potential harms—for example, intentional misuse via cyberwarfare and malicious hacking68,70,71,95,98,106—and suggest risk-management strategies. Harm is primarily interpreted as discrimination55,61,64,65,67,112,115, violation of privacy40,52,61,81,95,115,116 or bodily harm42,47,48,50,73,109,113,117. Less frequent characterizations include loss of trust47 or skills61; “radical individualism”55; the risk that technological progress might outpace regulatory measures74; and negative impacts on long-term social well-being61, infrastructure61, or psychological52,73, emotional73 or economic aspects61,73.
Harm-prevention guidelines focus primarily on technical measures and governance strategies, ranging from interventions at the level of AI research44,64,81,96,102,118, design40,42,44,49,56,73,75, technology development and/or deployment71 to lateral and continuous approaches50,72,80. Technical solutions include in-built data quality evaluations42 or security40 and privacy by design40,44,56, though notable exceptions also advocate for establishing industry standards47,81,119. Proposed governance strategies include active cooperation across disciplines and stakeholders50,64,70,79, compliance with existing or new legislation44,48,52,98,112,116, and the need to establish oversight processes and practices—notably tests53,55,64,91,96, monitoring53,75, audits and assessments by internal units, customers, users, independent third parties or governmental entities57,65,68,75,98,111,112,115, often geared towards standards for AI implementation and outcome assessment. Many imply that damages may be unavoidable, in which case risks should be assessed57,65,68, reduced57,86,89,90,91 and mitigated51,52,55,70,80,85, and the attribution of liability should be clearly defined48,54,55,61,99. Several sources mention potential “multiple”45,48,49,69,77,96 or “dual-use”8,50,55, take explicit position against military application48,54,117 or simply guard against the dynamics of an “arms race”51,70,119.
Responsibility and accountability
Despite widespread references to “responsible AI”60,68,95,100, responsibility and accountability are rarely defined. Nonetheless, specific recommendations include acting with “integrity”64,69,77 and clarifying the attribution of responsibility and legal liability40,75,95,120, if possible upfront53, in contracts69 or, alternatively, by centring on remedy43. In contrast, other sources suggest focusing on the underlying reasons and processes that may lead to potential harm91,100. Yet others underline the responsibility of whistleblowing in case of potential harm53,72,77, and aim at promoting diversity66,109 or introducing ethics into science, technology, engineering and mathematics education76. Very different actors are named as being responsible and accountable for AI’s actions and decisions: AI developers75,77,90,113, designers53,61, “institutions”57,59 or “industry”86. Further divergence emerged on whether AI should be held accountable in a human-like manner87 or whether humans should always be the only actors who are ultimately responsible for technological artifacts48,49,52,54,69,109.
Ethical AI sees privacy both as a value to uphold61,81,92,116 and as a right to be protected44,45,54,55,70. While often undefined, privacy is frequently presented in relation to data protection40,44,53,70,75,83,88,96,100,115 and data security44,52,81,83,105,115. A few sources link privacy to freedom55,70 or trust91,109. Suggested modes of achievement fall into three categories: technical solutions81,97, such as differential privacy91,106, privacy by design42,44,45,96,115, data minimization53,75 and access control53,75; calls for more research64,81,91,115 and awareness81,91; and regulatory approaches42,69,88, with sources referring to legal compliance more broadly44,49,53,75,77,98, or suggesting certificates121 or the creation or adaptation of laws and regulations to accommodate the specificities of AI81,91,105,122.
While promoting good (‘beneficence’ in ethical terms) is often mentioned, it is rarely defined, though notable exceptions mention the augmentation of human senses103, the promotion of human well-being and flourishing51,107, peace and happiness77, the creation of socio-economic opportunities53, and economic prosperity54,70. Similar uncertainty concerns the actors that should benefit from AI: private sector issuers tend to highlight the benefit of AI for customers40,65, though overall many sources require AI to be shared66,69,93 and to benefit everyone53,76,82,101, “humanity”54,61,77,117,119, both of the above65,83, “society”51,104, “as many people as possible”54,70,116, “all sentient creatures”100, the “planet”54,89 and the environment55,107. Strategies for the promotion of good include aligning AI with human values51,61, advancing “scientific understanding of the world”117, minimizing power concentration119 or, conversely, using power “for the benefit of human rights”99, working more closely with “affected” people82, minimizing conflicts of interests119, proving beneficence through customer demand65 and feedback75, and developing new metrics and measurements for human well-being61,107.
Freedom and autonomy
Whereas some sources specifically refer to the freedom of expression45,90,99,122 or informational self-determination45,107 and “privacy-protecting user controls”75, others generally promote freedom48,86,89, empowerment45,69,116 or autonomy48,50,79,94,98,113. Some documents refer to autonomy as a positive freedom, specifically the freedom to flourish53, to self-determination through democratic means55, the right to establish and develop relationships with other human beings55,109, the freedom to withdraw consent84, or the freedom to use a preferred platform or technology90,97. Other documents focus on negative freedom—for example, freedom from technological experimentation99, manipulation50 or surveillance55. Freedom and autonomy are believed to be promoted through transparency and predictable AI55, by not “reducing options for and knowledge of citizens”55, by actively increasing people’s knowledge about AI53,69,79, giving notice and consent96 or, conversely, by actively refraining from collecting and spreading data in absence of informed consent47,55,61,72,91.
References to trust include calls for trustworthy AI research and technology67,114,116, trustworthy AI developers and organizations68,77,83, trustworthy “design principles”108, or underline the importance of customers’ trust40,69,75,83,91,97. Calls for trust are proposed because a culture of trust among scientists and engineers is believed to support the achievement of other organizational goals116, or because overall trust in the recommendations, judgments and uses of AI is indispensable for AI to “fulfil its world changing potential”41. This last point is contradicted by one guideline explicitly warning against excessive trust in AI98. Suggestions for building or sustaining trust include education50, reliability67,68, accountability73, processes to monitor and evaluate the integrity of AI systems over time68, and tools and techniques ensuring compliance with norms and standards60,80. Whereas some guidelines require AI to be transparent54,60,74,75, understandable53,54 or explainable69 in order to build trust, another one explicitly suggests that, instead of demanding understandability, it should be ensured that AI fulfils public expectations67. Other reported facilitators of trust include “a Certificate of Fairness”121, multi-stakeholder dialogue81, awareness about the value of using personal data91, and avoiding harm47,73.
To the extent that is referenced, sustainability calls for development and deployment of AI to consider protecting the environment50,55,63, improving the planet’s ecosystem and biodiversity54, contributing to fairer and more equal societies82 and promoting peace83. Ideally, AI creates sustainable systems61,93,107 that process data sustainably60 and whose insights remain valid over time65. To achieve this aim, AI should be designed, deployed and managed with care55 to increase its energy efficiency and minimize its ecological footprint48. To make future developments sustainable, corporations are asked to create policies ensuring accountability in the domain of potential job losses54 and to use challenges as an opportunity for innovation55.
While dignity remains undefined in existing guidelines, save one specification that it is a prerogative of humans but not robots109, there is frequent reference to what it entails: dignity is intertwined with human rights118 or otherwise means avoiding harm48, forced acceptance48, automated classification55 and unknown human–AI interaction55. It is argued that AI should not diminish50 or destroy97, but respect99, preserve86 or even increase human dignity53,54. Dignity is believed to be preserved if it is respected by AI developers in the first place113 and promoted through new legislation55, through governance initiatives53, or through government-issued technical and methodological guidelines99.
Solidarity is mostly referenced in relation to the implications of AI for the labour market121. Sources call for a strong social safety net54,101. They underline the need for redistributing the benefits of AI in order not to threaten social cohesion66 and respecting potentially vulnerable persons and groups50. Lastly, there is a warning of data collection and practices focused on individuals that may undermine solidarity in favour of “radical individualism”55.
The rapid increase in the number and variety of guidance documents attests to growing interest in AI ethics by the international community and across different types of organization. The nearly equivalent proportion of documents issued by the public sector (that is, governmental and intergovernmental organizations) and the private sector (companies and private sector alliances) suggests that AI ethics concerns both public entities and private enterprises. However, the solutions proposed to meet the ethical challenges diverge significantly. Furthermore, the underrepresentation of geographic areas such as Africa, South and Central America and Central Asia indicates that global regions are not participating equally in the AI ethics debate, which reveals a power imbalance in the international discourse. More economically developed countries are shaping this debate more than others, which raises concerns about neglecting local knowledge, cultural pluralism and the demands of global fairness.
Our thematic synthesis uncovers the existence of 11 clusters of ethical principles and reveals an emerging cross-stakeholder convergence on promoting the ethical principles of transparency, justice, non-maleficence, responsibility and privacy, which are referenced in more than half of all guidelines. The potentially pro-ethical nature of transparency, “enabling or impairing other ethical practices or principles”123, might partly explain its prevalence, whereas the frequent occurrence of calls for privacy, non-maleficence, justice and fairness can be seen as cautioning the global community against potential risks brought by AI. These themes appear to be intertwined with the theme of responsibility, although only a few guidelines emphasize the duty of all stakeholders involved in the development and deployment of AI to act with integrity.
Because references to non-maleficence outnumber those related to beneficence, it appears that issuers of guidelines are preoccupied with the moral obligation to prevent harm. This focus is corroborated by a greater emphasis on preserving privacy, dignity, autonomy and individual freedom in spite of advances in AI rather than on actively promoting these principles124. It could be due to the so-called negativity bias—that is, a general cognitive bias to give greater weight to negative entities125,126—or as a precautionary measure to ensure that AI developers and deployers are held in check25.
The references to trust address a critical ethical issue in AI governance; that is, whether it is morally desirable to foster public trust in AI. Whereas several sources, predominantly from the private sector, highlight the importance of fostering trust in AI through educational and awareness-raising activities, others contend that trust in AI risks diminishing scrutiny and may undermine certain societal obligations of AI producers127. This later perspective challenges the dominant view that building public trust in AI should be a fundamental requirement for ethical governance128.
Sustainability, dignity and solidarity are significantly underrepresented compared to other ethical dimensions, which suggests that these issues might be currently flying under the radar of the mainstream AI ethics debate. This underrepresentation is particularly problematic in light of recent evidence that AI requires massive computational resources, which, in turn, require high energy consumption129. The environmental impact of AI, however, involves not only the negative effects of high-footprint digital infrastructures, but also the possibility of harnessing AI for the benefit of ecosystems and the entire biosphere. As the humanitarian cost of anthropogenic climate change is rapidly increasing130, the principles of sustainability and solidarity appear strictly intertwined. In addition, the ethical principle of solidarity is referenced even more rarely, typically in association with the development of inclusive strategies for the prevention of job losses and unfair sharing of burdens. Little attention is devoted to promoting solidarity through the emerging possibility of using AI expertise for solving humanitarian challenges, a mission that is currently being pursued, among others, by companies such as Microsoft131 and intergovernmental organizations such as the United Nations Office for Project Services (UNOPS)132 or the World Health Organization (WHO). A better appraisal of currently underrepresented ethical principles is likely to result in a more inclusive AI ethics landscape.
Although numerical data indicate an emerging convergence around the importance of certain ethical principles, an in-depth thematic analysis paints a more complicated picture. Our focused coding reveals substantive divergences among all 11 ethical principles in relation to four major factors: (1) how ethical principles are interpreted; (2) why they are deemed important; (3) what issue, domain or actors they pertain to; and (4) how they should be implemented. These conceptual and procedural divergences reveal uncertainty as to which ethical principles should be prioritized and how conflicts between ethical principles should be resolved, and it may undermine attempts to develop a global agenda for ethical AI. For example, the need for ever-larger, more diverse datasets to ‘unbias’ AI might conflict with the requirement to give individuals increased control over their data and its use in order to respect their privacy and autonomy. Similar contrasts emerge between avoiding harm at all costs and the perspective of accepting some degree of harm as long as risks and benefits are weighed against each other. Moreover, risk–benefit evaluations are likely to lead to different results depending on whose well-being will be optimized for and by which actors. Such divergences and tensions illustrate a gap at the cross-section of principle formulation and their implementation into practice133,134.
These findings have implications for public policy, technology governance and research ethics. At the policy level, greater inter-stakeholder cooperation is needed to mutually align different AI ethics agendas and to seek procedural convergence not only on ethical principles but also their implementation. While global consensus might be desirable it should not come at the cost of obliterating cultural and moral pluralism and may require the development of deliberative mechanisms to adjudicate disagreement among stakeholders from different global regions. Such efforts can be mediated and facilitated by intergovernmental organizations, complemented by bottom-up approaches involving all relevant stakeholders135. Furthermore, it should be clarified how AI ethics guidelines relate to existing national and international regulation. Translating principles into practice and seeking harmonization between AI ethics codes (soft law) and legislation (hard law) are important next steps for the global community.
At the level of technology governance, harmonization is typically implemented in terms of standardizations such as the Ethically Aligned Designed initiative136 led by the Institute of Electrical and Electronics Engineers (IEEE). Although standardization is a necessary step in the right direction, it remains to be seen whether the impact of these non-legal norms will mostly happen on the policy level or if they will also influence individual practitioners and decision-makers. Finally, research ethics mechanisms like independent review boards will be increasingly required to assess the ethical validity of AI applications in scientific research, especially those in the academic domain. However, AI applications by governments or private corporations are unlikely to fall under their oversight, unless significant expansions to the purview of independent review boards are made.
This study has several limitations. First, guidelines and soft-law documents are an instance of grey literature, and are therefore not indexed in conventional scholarly databases. As a result, their retrieval is inevitably less replicable and unbiased compared to systematic database searches of peer-reviewed literature. Following best practices for grey literature review, this limitation has been mitigated by developing a discovery and eligibility protocol, which was pilot-tested prior to data collection. Although search results from search engines are personalized, the risk of personalization influencing discovery has been mitigated through the broadness of both the keyword search and the inclusion of results. A language bias may have skewed our corpus towards English results. Our content analysis presents the typical limitations of qualitative analytic methods. Following best practices for content analysis, this limitation has been mitigated by developing an inductive coding strategy, which was conducted independently by two reviewers to minimize subjective bias. Finally, given the rapid pace of publication of AI guidance documents, there is a possibility that new policy documents were published after our search was completed. To minimize this risk, continuous monitoring of the literature was conducted in parallel with the data analysis and until 23 April 2019. In spite of these limitations, our Perspective provides a comprehensive mapping of the current AI ethics landscape and offers a basis for future research on this topic.
Harari, Y. N. Reboot for the AI revolution. Nature 550, 324–327 (2017).
Appenzeller, T. The AI revolution in science. Science https://doi.org/10.1126/science.aan7064 (2017).
Jordan, M. I. & Mitchell, T. M. Machine learning: trends, perspectives, and prospects. Science 349, 255–260 (2015).
Stead, W. W. Clinical implications and challenges of artificial intelligence and deep learning. JAMA 320, 1107–1108 (2018).
Vayena, E., Blasimme, A. & Cohen, I. G. Machine learning in medicine: addressing ethical challenges. PLOS Med. 15, e1002689 (2018).
Awad, E. et al. The Moral Machine experiment. Nature 563, 59–64 (2018).
Science must examine the future of work. Nature 550, 301–302 (2017).
Brundage, M. et al. The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation (Future of Humanity Institute, University of Oxford, Centre for the Study of Existential Risk, University of Cambridge, Center for a New American Security, Electronic Frontier Foundation, OpenAI, 2018).
Zou, J. & Schiebinger, L. AI can be sexist and racist — it’s time to make it fair. Nature 559, 324–326 (2018).
Boddington, P. Towards a Code of Ethics for Artificial Intelligence (Springer, 2017).
Bostrom, N. & Yudkowsky, E. in The Cambridge Handbook of Artificial Intelligence (eds Frankish, K. & Ramsey, W. M.) 316–334 (Cambridge Univ. Press, 2014). https://doi.org/10.1017/CBO9781139046855.020
Etzioni, A. & Etzioni, O. AI assisted ethics. Ethics Inf. Technol. 18, 149–156 (2016).
Yuste, R. et al. Four ethical priorities for neurotechnologies and AI. Nature 551, 159–163 (2017).
Cath, C., Wachter, S., Mittelstadt, B., Taddeo, M. & Floridi, L. Artificial intelligence and the ‘good society’: the US, EU, and UK approach. Sci. Eng. Ethics 24, 505–528 (2018).
Zeng, Y., Lu, E. & Huangfu, C. Linking artificial intelligence principles. Preprint at https://arxiv.org/abs/1812.04814 (2018).
Greene, D., Hoffmann, A. L. & Stark, L. Better, nicer, clearer, fairer: a critical assessment of the movement for ethical artificial intelligence and machine learning. In Proc. 52nd Hawaii International Conference on System Sciences 2122–2131 (2019).
Crawford, K. & Calo, R. There is a blind spot in AI research. Nature 538, 311–313 (2016).
Altman, M., Wood, A. & Vayena, E. A harm-reduction framework for algorithmic fairness. IEEE Security Privacy 16, 34–45 (2018).
Bolukbasi, T., Chang, K.-W., Zou, J., Saligrama, V. & Kalai, A. Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. Preprint at https://arxiv.org/abs/1607.06520 (2016).
O’Neil, C. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy (Crown, 2016).
Veale, M. & Binns, R. Fairer machine learning in the real world: mitigating discrimination without collecting sensitive data. Big Data Soc. https://doi.org/10.1177/2053951717743530 (2017).
Shoham, Y. et al. The AI Index 2018 Annual Report (AI Index Steering Committee, Human-Centered AI Initiative, Stanford University, 2018).
Sossin, L. & Smith, C. W. Hard choices and soft law: ethical codes, policy guidelines and the role of the courts in regulating government. Alberta Law Rev. 40, 867–893 (2003).
Campbell, A. & Glass, K. C. The legal status of clinical and ethics policies, codes, and guidelines in medical practice and research. McGill Law J. 46, 473–489 (2001).
Benkler, Y. Don’t let industry write the rules for AI. Nature 569, 161 (2019).
Wagner, B. in Being Profiled: Cogitas Ergo Sum: 10 Years of Profiling the European Citizen (eds Bayamlioglu, E., Baraliuc, I., Janssens, L. A. W. & Hildebrandt, M.) 84–89 (Amsterdam Univ. Press, 2018).
Arksey, H. & O’Malley, L. Scoping studies: towards a methodological framework. Int. J. Soc. Res. Methodol. 8, 19–32 (2005).
Pham, M. T. et al. A scoping review of scoping reviews: advancing the approach and enhancing the consistency. Res. Synth. Meth. 5, 371–385 (2014).
Liberati, A. et al. The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: explanation and elaboration. PLOS Medicine 6, e1000100 (2009).
Boddington, P. Alphabetical list of resources. Ethics for Artificial Intelligence https://www.cs.ox.ac.uk/efai/resources/alphabetical-list-of-resources/ (2018).
Winfield, A. A round up of robotics and AI ethics. Alan Winfield’s Web Log http://alanwinfield.blogspot.com/2019/04/an-updated-round-up-of-ethical.html (2017).
National and international AI strategies. Future of Life Institute https://futureoflife.org/national-international-ai-strategies/ (2018)
Summaries of AI policy resources. Future of Life Institute https://futureoflife.org/ai-policy-resources/ (2018).
Hagstrom, C., Kendall, S. & Cunningham, H. Googling for grey: using Google and Duckduckgo to find grey literature. In Abstracts of the 23rd Cochrane Colloquium Vol. 10, LRO 3.6, 40 (Cochrane Database of Systematic Reviews, 2015).
Piasecki, J., Waligora, M. & Dranseika, V. Google search as an additional source in systematic reviews. Sci. Eng. Ethics 24, 809–810 (2017).
Moher, D., Liberati, A., Tetzlaff, J. & Altman, D. G., The PRISMA Group. Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. PLOS Med. 6, e1000097 (2009).
Saldaña, J. The Coding Manual for Qualitative Researchers (SAGE, 2013).
Noblit, G. W. & Hare, R. D. Meta-Ethnography: Synthesizing Qualitative Studies (SAGE, 1988).
Daniels, N. Justice and Justification: Reflective Equilibrium in Theory and Practice (Cambridge Univ. Press, 1996).
Guidelines for artificial intelligence. Deutsche Telekom https://www.telekom.com/en/company/digital-responsibility/details/artificial-intelligence-ai-guideline-524366 (2018).
Transparency and trust in the cognitive era. IBM https://www.ibm.com/blogs/think/2017/01/ibm-cognitive-principles/ (2017).
Initial code of conduct for data-driven health and care technology. GOV.UK https://www.gov.uk/government/publications/code-of-conduct-for-data-driven-health-and-care-technology/initial-code-of-conduct-for-data-driven-health-and-care-technology (2019).
Diakopoulos, N. et al. Principles for accountable algorithms and a social impact statement for algorithms. FATML http://www.fatml.org/resources/principles-for-accountable-algorithms (2016).
AI principles of Telefónica. Telefónica https://www.telefonica.com/en/web/responsible-business/our-commitments/ai-principles (2018).
Declaration on Ethics and Data Protection in Artificial Intelligence (Commission Nationale de l’Informatique et des Libertés, European Data Protection Supervisor & Garante per la protezione dei dati personali, 2018).
Everyday Ethics for Artificial Intelligence (IBM, 2018).
Ethics Commission: Automated and Connected Driving (Federal Ministry of Transport and Digital Infrastructure, 2017).
Position on Robotics and Artificial Intelligence (Green Digital Working Group, 2016).
Principles of robotics. EPSRC https://epsrc.ukri.org/research/ourportfolio/themes/engineering/activities/principlesofrobotics/ (2011).
Ethics Guidelines for Trustworthy AI (High-Level Expert Group on AI, 2019).
Artificial intelligence principles and ethics. Smart Dubai http://www.smartdubai.ae/initiatives/ai-principles-ethics (2019).
Dawson, D. et al. Artificial Intelligence: Australia’s Ethics Framework (Australian Government, 2019).
Artificial intelligence and machine learning: policy paper. Internet Society https://www.internetsociety.org/resources/doc/2017/artificial-intelligence-and-machine-learning-policy-paper/ (2017).
Top 10 Principles for Ethical AI (UNI Global, 2017).
Statement on Artificial Intelligence, Robotics and ‘Autonomous’ Systems (European Group on Ethics in Science and New Technologies, 2018).
Big Data, Artificial Intelligence, Machine Learning and Data Protection. (ICO, 2017).
Universal guidelines for artificial intelligence. The Public Voice https://thepublicvoice.org/ai-universal-guidelines/ (2018).
Science, law and society (SLS) initiative. The Future Society https://web.archive.org/web/20180621203843/http://thefuturesociety.org/science-law-society-sls-initiative/ (2018).
Statement on Algorithmic Transparency and Accountability (ACM, 2017).
Dutch Artificial Intelligence Manifesto (Special Interest Group on Artificial Intelligence, 2018).
Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems, Version 2 (IEEE, 2017).
The Toronto declaration: protecting the right to equality and non-discrimination in machine learning systems. Human Rights Watch https://www.hrw.org/news/2018/07/03/toronto-declaration-protecting-rights-equality-and-non-discrimination-machine (2018).
Floridi, L. et al. AI4People—an ethical framework for a good ai society: opportunities, risks, principles, and recommendations. Minds Mach. 28, 589–707 (2018).
SAP’s guiding principles for artificial intelligence (AI). SAP https://www.sap.com/products/leonardo/machine-learning/ai-ethics.html#guiding-principles (2018).
Ethical Principles for Artificial Intelligence and Data Analytics (SIIA, 2017).
Koski, O. & Husso, K. Work in the Age of Artificial Intelligence (Ministry of Economic Affairs and Employment, 2018).
Digital decisions. Center for Democracy & Technology https://cdt.org/issue/privacy-data/digital-decisions/ (2019).
Ethics framework. MI Garage https://www.migarage.ai/ethics-framework/ (2019).
Business Ethics and Artificial Intelligence (Institute of Business Ethics, 2018).
Asilomar AI Principles. Future of Life Institute https://futureoflife.org/ai-principles/ (2017).
The responsible AI framework. PwC https://www.pwc.co.uk/services/audit-assurance/risk-assurance/services/technology-risk/technology-risk-insights/accelerating-innovation-through-responsible-ai/responsible-ai-framework.html (2019).
Whittaker, M. et al. AI Now Report 2018 (AI Now Institute, 2018).
Discussion Paper on AI and Personal Data — Fostering Responsible Development and Adoption of AI (Personal Data Protection Commission Singapore, 2018).
Artificial intelligence (AI) in health. RCP London https://www.rcplondon.ac.uk/projects/outputs/artificial-intelligence-ai-health (2018).
Responsible bots: 10 guidelines for developers of conversational AI. Microsoft https://www.microsoft.com/en-us/research/publication/responsible-bots/ (2018).
Villani, C. For a Meaningful Artificial Intelligence: Towards a French and European Strategy (AI for Humanity, 2018).
The Japanese Society for Artificial Intelligence Ethical Guidelines (Japanese Society for Artificial Intelligence, 2017).
Demiaux, V. How Can Humans Keep the Upper Hand? The Ethical Matters Raised by Algorithms and Artificial Intelligence (CNIL, 2017).
European ethical charter on the use of artificial intelligence in judicial systems and their environment. Council of Europe https://www.coe.int/en/web/cepej/cepej-european-ethical-charter-on-the-use-of-artificial-intelligence-ai-in-judicial-systems-and-their-environment (2019).
Ethics of AI in Radiology: European and North American Multisociety Statement (American College of Radiology, 2019).
Charlevoix Common Vision for the Future of Artificial Intelligence (Leaders of the G7, 2018).
DeepMind ethics and society principles. DeepMind https://deepmind.com/applied/deepmind-ethics-society/principles/ (2017).
Sony Group AI Ethics Guidelines (Sony, 2018).
Artificial Intelligence and Privacy (Datatilsynet, 2018).
White Paper: How to Prevent Discriminatory Outcomes in Machine Learning (WEF, 2018).
ITI AI Policy Principles (ITI, 2017).
The Ethics of Code: Developing AI for Business with Five Core Principles (Sage, 2017).
Commitments and principles. OP https://www.op.fi/op-financial-group/corporate-social-responsibility/commitments-and-principles (2019).
Tieto’s AI Ethics Guidelines (Tieto, 2018).
Introducing Unity’s Guiding Principles for Ethical AI. Unity Blog https://blogs.unity3d.com/2018/11/28/introducing-unitys-guiding-principles-for-ethical-ai/ (2018).
Discussion Paper: National Strategy for Artificial Intelligence (NITI Aayog, 2018).
AI in the UK: Ready, Willing and Able 183 (House of Lords, 2018).
Unified Ethical Frame for Big Data Analysis: IAF Big Data Ethics Initiative, Part A (The Information Accountability Foundation, 2015).
Fenech, M., Strukelj, N. & Buston, O. Ethical, Social, and Political Challenges of Artificial Intelligence in Health (Future Advocacy, 2019).
Responsible AI and robotics: an ethical framework. Accenture https://www.accenture.com/gb-en/company-responsible-ai-robotics (2019).
Artificial intelligence at Google: our principles. Google AI https://ai.google/principles/ (2019).
Microsoft AI principles. Microsoft https://www.microsoft.com/en-us/ai/our-approach-to-ai (2017).
Éthique de la Recherche en Robotique (Allistene, 2014).
van Est, R. & Gerritsen, J. Human Rights in the Robot Age: Challenges Arising from the Use of Robotics, Artificial Intelligence, and Virtual and Augmented Reality (Rathenau Institute, 2017).
The declaration. Montreal Declaration https://www.montrealdeclaration-responsibleai.com/the-declaration (2017).
Mid- to Long-Term Master Plan in Preparation for the Intelligent Information Society: Managing the Fourth Industrial Revolution (Government of the Republic of Korea, 2017).
Crawford, K. et al. The AI Now Report: The Social and Economic Implications of Artificial Intelligence Technologies in the Near-Term (AI Now Institute, 2016).
Report on Artificial Intelligence and Human Society: Unofficial Translation (Ministry of State for Science and Technology Policy, 2017).
Preparing for the future of Artificial Intelligence (NSTC, 2016).
Artificial Intelligence: The Public Policy Opportunity (Intel, 2017).
Machine Learning: The Power and Promise of Computers that Learn by Example (Royal Society, 2017).
Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems, Version 1 (IEEE, 2019).
Report with Recommendations to the Commission on Civil Law Rules on Robotics (European Parliament, 2017).
Report of COMEST on Robotics Ethics (COMEST/UNESCO, 2017).
Campolo, A., Sanfilippo, M., Whittaker, M. & Crawford, K. AI Now 2017 Report (AI ow Institute, 2017).
Policy Recommendations on Augmented Intelligence in Health Care H-480.940 (AMA, 2018).
Avila, R., Brandusescu, A., Freuler, J. O. & Thakur, D. Artificial Intelligence: Open Questions about Gender Inclusion (World Wide Web Foundation, 2018).
Draft AI R&D Guidelines for International Discussions (The Conference toward AI Network Society, 2017).
The National Artificial Intelligence Research and Development Strategic Plan (NSTC, 2016).
Tenets. The Partnership on AI https://www.partnershiponai.org/tenets/ (2016).
Ethics Policy. IIIM http://www.iiim.is/2015/08/ethics-policy/ (2015).
Latonero, M. Governing artificial intelligence: upholding human rights & dignity. Data & Society https://datasociety.net/output/governing-artificial-intelligence/ (2018).
OpenAI Charter. OpenAI https://blog.openai.com/openai-charter/ (2018).
L’Intelligenzia Artificiale al Servizio del Cittadino (AGID, 2018).
Gilburt, B. Women leading in AI: 10 principles of responsible AI. Towards Data Science https://towardsdatascience.com/women-leading-in-ai-10-principles-for-responsible-ai-8a167fc09b7d (2019).
Privacy and Freedom of Expression in the Age of Artificial Intelligence (Privacy International/Article 19, 2018).
Turilli, M. & Floridi, L. The ethics of information transparency. Ethics Inf. Technol. 11, 105–112 (2009).
Taddeo, M. & Floridi, L. How AI can be a force for good. Science 361, 751–752 (2018).
Rozin, P. & Royzman, E. B. Negativity bias, negativity dominance, and contagion. Person. Soc. Psychol. Rev. https://doi.org/10.1207/S15327957PSPR0504_2 (2016).
Bentley, P. J., Brundage, M., Häggström, O. & Metzinger, T. Should We Fear Artificial Intelligence? In-Depth Analysis (European Parliament, 2018).
Bryson, J. AI & global governance: no one should trust AI. United Nations University https://cpr.unu.edu/ai-global-governance-no-one-should-trust-ai.html (2018).
Winfield, A. F. T. & Marina, J. Ethical governance is essential to building trust in robotics and artificial intelligence systems. Philos. Trans. R. Soc. A 376, 20180085 (2018).
Strubell, E., Ganesh, A. & McCallum, A. Energy and policy considerations for deep learning in NLP. Preprint at https://arxiv.org/abs/1906.02243 (2019).
Scheffran, J., Brzoska, M., Kominek, J., Link, P. M. & Schilling, J. Climate change and violent conflict. Science 336, 869–871 (2012).
AI for humanitarian action. Microsoft https://www.microsoft.com/en-us/ai/ai-for-humanitarian-action (2019).
Lancaster, C. Can artificial intelligence improve humanitarian responses? UNOPS https://www.unops.org/news-and-stories/insights/can-artificial-intelligence-improve-humanitarian-responses (2018).
Hagendorff, D. T. The ethics of AI ethics — an evaluation of guidelines. Preprint at https://arxiv.org/abs/1903.03425 (2019).
Whittlestone, J., Nyrup, R., Alexandrova, A. & Cave, S. The role and limits of principles in AI ethics: towards a focus on tensions. In Proc. 2019 AAAI/ACM Conference on AI, Ethics, and Society 195–200 (2019).
Mittelstadt, B. AI ethics — too principled to fail? Preprint at https://arxiv.org/abs/1906.06668 (2019).
The IEEE global initiative on ethics of autonomous and intelligent systems. IEEE Standards Association https://standards.ieee.org/industry-connections/ec/autonomous-systems.html (2019).
The authors would like to thank J. Sleigh for her help with creating the colour-coded map.
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Tables 1–3