In order to limit the effects of the biases that taint the practices of sustainable AI oriented towards ESG, “coalitions” made up of institutions, companies, scientists and think tanks, strive to define ethical frameworks – in the form of codes and charters – based on philosophical, sociological and psychological reflections, but this framework provokes debates between regulators and libertarians.
The Foundations of AI Ethics
Ethical deliberations on AI practices follow three main approaches: universal, normative, and applied. The first – of an axiological nature and inspired by Kant and Rousseau – is based on the principles and values that underpin life in society: respect for man, truth, justice, nature, etc. The second – called legalistic or prescriptive – covers moral judgments and social values, such as true or false, good or bad, fair or unfair, etc. The third – of a praxeological nature – measures the consequences, externalities or impacts of a system, behavior or object on the economy, nature, society or the person. It is most often applied to new technologies, in particular AI, and to management, in particular sustainable management (Pistilli, 2024).
AI ethics frameworks
In order to limit their negative impacts on collective and individual thoughts, decisions or behaviors, guides or ethical charters published by companies and normative codes, references and/or regulations issued by regulators (international organizations, nation-states, associations), aim to frame the design of software and the use of their data and results by companies (Constantinides & al, 2024). Codes generally do not have an ethical or moral dimension, unlike charters, guides or, in France, the “raisons d’être” of companies.
International institutions such as the OECD, the UN, UNESCO and the G7 have been working since 2019 to establish a “normative ethics” and “global governance of AI”. The Vatican, inspired by the work of Bonanti (2018), who inspired several codes, advocates the advent of an “algorithmic ethics” (or ethics of algorithms) based on the principles of “transparency, social inclusion, responsibility, impartiality and reliability”. Overall, according to Menecoeur (2020), the 126 documents on the ethics of AI identified worldwide are divided between public codes (national and international) and private guides (companies, universities and associations). But the rulers of the United States, the People’s Republic of China and the European Union, as well as the leaders of their digital and/or ESG-oriented companies, have applied rules, codes and ethical guides that respond to often different approaches.
The ethical relativism of AI and ESG
European texts – and in particular the AI Act – are the subject of intense lobbying, in particular by GAFAM, in order to avoid open source AI and the downgrading of certain generative AI software. Most European think tanks advise strengthening the regulation of AI, such as the Montaigne Institute, which launched the Objectif IA operation for digital training, the CSR Observatory, which strives to put AI at the service of the application of ESG standards, and the Louis Bachelier Institute, which has launched the Good in Tech program to measurethe impact of AI on society.
But a collective of 30 global AI leaders denounced the European approach, stating that “Europe has become less competitive and less innovative compared to other regions, and it now risks falling even further behind in the AI era due to inconsistent regulatory decisions”. These reactions show that AI codes and guides are subject to a form of “ethical relativism”, because they depend on both technological and economic factors, but also – and increasingly – on geopolitical and cultural considerations. They are the subject of different readings according to the disciplines, professions and ideologies of the actors of AI and ideology.
In the United States, under the influence of the Federal Guidelines for Sentencing Organizations (1991), practical guides (guidelines) are more common among American companies than among European or Chinese firms. The GAFAM have initiated a Partnership on AI which recommends the application of general principles and a collective commitment: “We are committed to conducting open research and dialogue on the ethical, social and economic implications of AI” (Hern, 2016). But the interpretation of these principles differs from one society to another. Google focuses on social criteria (including non-discrimination). Apple displays a charter based on honesty and respect for stakeholders. Meta only declares that it applies the professional standards in force. Amazon takes up the fundamental principles, but paradoxically states that “AI drives humans”; Microsoft and Open AI display “Codes of trust reflecting their cultures and values; they recognize “the potential for bias in algorithms and strive to mitigate their impact”.
In Europe, reflections on the ethics of AI were launched in 2015 and led to a regulation aimed at the protection of personal data (Data Governance Act) published in 2018 (but applied in 2023), then a white paper on AI (2020), a directive on microprocessors (Chips Act, 2023) and a directive on artificial intelligence (AI Act) passed in 2024, (but applicable in 2026).
These texts strive to manage the risks induced by AI and to promote a “trustworthy AI” based on compliance with laws (Lawful AI), ethical values (ethical AI) and technical skills (robust AI). The principles – borrowed from bioethics – relate to human autonomy (respect for the rights of citizens), the prevention of any infringement (protection of people and property), equity (between users) and explainability (of software). These principles were then broken down into “requirements”: systems (data, software, results) must “remain under human control”; they must be “robust, safe and transparent”. Compliance with the requirements is controlled by so-called “technical” methods: audit of “trustworthy architectures” (Trustworthy AI), control of the application of design standards (X-by-design), methods of explanation, tests, validations and implementations. Non-technical methods complete this system: regulation by codes, charters and guides, certification of systems, orientation of AI training and AI research.
Among these methods, the AI Act prioritizes so-called “foundational” AI systems according to three levels of risk to public and private lives: models classified as “unacceptable”, leading to intrusive and discriminatory uses of AI, are prohibited; systems classified as “high risk”, which may affect the health, safety or fundamental rights of people or the environment, are subject to a strict regime of monitoring software bias and data governance (false images, illegal content, images or texts subject to copyright, must in particular be reported and corrected); “moderate risk systems” must be subject to declarations of conformity. The European regulation aims in particular to limit the misuse of codes of ethics for the sole benefit of commercial interests (Ethic or Blue washning). European leaders aim to transform “the European Union into a world leader in AI innovation, while ensuring that AI technologies benefit all European citizens”.