Here we call an ethical challenge a question that the law alone cannot settle: the law sets the framework or part of it, and the organisation is left with a choice to justify. This convention echoes the Ethics Guidelines for Trustworthy AI of the high-level expert group set up by the European Commission, according to which laws “are not always up to speed with technological developments, can at times be out of step with ethical norms or may simply not be well suited to addressing certain issues” (paragraph 26). The ethical challenges of generative AI are read with this grid.
We retain six, which the public texts consulted on 12 November 2026 address: bias, misleading content and disinformation, copyright and training data, transparency towards people, effects on work and environmental footprint. For each, we read the OECD Recommendation, these guidelines and, for what the law has settled, Regulation (EU) 2024/1689, known as the AI Act, in its consolidated version of 27 July 2026. This reading is not an ethicist’s opinion: it situates questions that each organisation settles.
Texts of principle and texts of law
The OECD Council Recommendation on Artificial Intelligence (OECD/LEGAL/0449) was adopted on 22 May 2019, then revised on 8 November 2023 and 3 May 2024. According to its background information, the 2024 revision aimed in particular to reflect the growing importance of addressing misinformation and disinformation in the context of generative AI, and to add an explicit reference to environmental sustainability.
The European guidelines, published on 8 April 2019 according to the Commission’s page, describe trustworthy AI as lawful, ethical and robust. None of their elements can create legal rights or impose legal obligations towards third parties (paragraph 25). The AI regulation cites the “Union ethical guidelines for trustworthy AI” among the elements of voluntary codes of conduct (Article 95(2)(a)); seeing this document in it is our reading.
The UNESCO Recommendation on the Ethics of Artificial Intelligence is not cited, as it could not be read.
Bias and misleading content
According to paragraph 80 of the guidelines, data sets used by AI systems, for training as for operation, “may suffer from the inclusion of inadvertent historic bias, incompleteness and bad governance models”. OECD principle 1.5 lists among the risks to be managed those related to harmful bias. An organisation that uses a large language model supplied by a third party as it stands has no control over its training or its design, only over the task entrusted and the use of the outputs. A use that touches on recruitment may fall under high-risk systems, presented in our article on Article 6 of the regulation.
Against misleading content, Article 50 requires providers of systems generating synthetic content to mark the outputs “in a machine-readable format” (paragraph 2). It requires deployers to disclose deep fakes and texts “published with the purpose of informing the public on matters of public interest”, except, for these texts, where there is human review and editorial responsibility is assumed (paragraph 4). OECD principle 1.2 addresses misinformation and disinformation amplified by AI. In our reading, to be submitted to the lawyers, an internal note or a reply to a customer produced with the tool falls outside paragraph 4; its accuracy depends on the review planned.
Copyright and training data
Directive (EU) 2019/790 of 17 April 2019 defines text and data mining as “any automated analytical technique aimed at analysing text and data in digital form in order to generate information” (Article 2, point 2). Its Article 3 reserves an exception for research organisations and cultural heritage institutions, for the purposes of scientific research. Its Article 4 opens one for works that are “lawfully accessible”, unless their use has been “expressly reserved by their rightholders in an appropriate manner, such as machine-readable means” (paragraph 3).
Article 53(1) of the regulation requires providers of general-purpose AI models to have a copyright compliance policy, which identifies and complies with this reservation (point (c)), and a “sufficiently detailed” public summary of the training content (point (d)).
These texts do not say, in our reading, whether training a model on protected works falls under Article 4 in all cases. The question, which is open, belongs to the lawyers. The organisation can read this summary, ask for this policy and check its rights over the documents entrusted to an assistant.
Transparency towards people and effects on work
Article 50(1) requires providers to ensure that people are informed that they are interacting with an AI system, “unless this is obvious” from the context, and “at the latest” at the time of the first interaction (paragraph 5). The guidelines add that the option to prefer human interaction should be offered where appropriate (paragraph 78). Article 50 does not provide for it. OECD principle 1.3 extends information about these interactions to the workplace.
On work, OECD recommendation 2.4 is addressed to governments. The guidelines, addressed to all stakeholders, write that systems should support human beings in the working environment (paragraph 50) and cite the “information, consultation and participation of workers” (paragraph 82). The regulation settles one case: under Article 26(7), the employer informs workers’ representatives and the affected workers before putting a high-risk system into service at the workplace. Outside this category, the regulation is silent; employment law remains to be examined.
The environmental footprint, information to request
OECD principle 1.1 refers to “environmental sustainability”. The guidelines cite as an example a “critical examination of the resource usage and energy consumption during training” (paragraph 84). The regulation includes “the known or estimated energy consumption of the model” in the documentation that the provider of a general-purpose AI model supplies, on request, to the AI Office and national authorities (Annex XI, Section 1, point 2(e)). The deployer is not its recipient.
The organisation can include the question in its choice: the information the provider agrees to communicate, the size of the model in relation to the task, the processing that can do without generation. Our page generative AI in the enterprise deals with this choice.
One sheet per use to justify the choices
The guidelines describe trade-offs in paragraph 90: the interests at stake are identified, and “any decision about which trade-off to make should be reasoned and properly documented”.
We propose deriving from this one sheet per use: for each challenge, what the law requires according to the lawyers, the decision taken beyond that, its reason and its date. It supplements the system register on our AI governance page; users’ rules fall under the corporate AI charter, and the application dates of Article 50 under our article on AI Act obligations.
The first sheet can cover the use whose outputs leave the organisation, texts sent to customers or published, because transparency and accuracy are at stake there in front of third parties.