Table 4.

Main highlights for the reviewed body of literature (divided by key salient ethical considerations).

Concepts	Identified issues	Key reviewed articles on this issue	Key insights and best practices for supporting research ethics stakeholders
Scope and approaches	Intersectoral and interdisciplinary participatory efforts are needed to develop dynamic, adaptative, and relevant normative guidance and practices	Gooding and Kariotis, 2021	Work up new ways of collaboration for REBs.
Diversity and fair representation	Concerns regarding inclusion are often found in RCTs.	Grote, 2021	Relatively little data is found for other types of research. Does not seem to consider research projects with retrospective data.
Biases toward vulnerable population	AI systems are either fed with actual biased data or generate biased results.	Cath et al., 2018; Chassang et al., 2021; Grote, 2021	If the algorithms are not considered unbiased and representative of the general population, results could exclude minorities and, thus, be harmful.
Informed consent	Transparency and accessibility of relevant information could help participants better understand a situation which will allow them to make a conscious decision. For informed consent, there should be a focus on the impacts and risks that arise from interventions using AI and ML. There is a dilemma between giving out all the information to participants and only the relevant ones they need.	Sedenberg et al., 2016; Nebeker et al., 2019; Grote, 2021 Jacobson et al., 2020 Jacobson et al., 2020	The issue of informed consent raises concerns about the following: - Nature of the information may be disclosed in the consent while the model is still at the preliminary stage of development - What risks can be revealed when we do not know the impacts of the technology? Possibility of causing harm to participants if incomplete or unreliable information is disclosed. The problem of the quality of consent and its scope in a complex and rapidly evolving technological field. Questioning the need to develop new tools for consent. Limits on the possibility of revoking consent compared to other types of research. REBs must make sure that researchers are giving intelligible information to participants.
Benefit risks assessment	One trigger point is to establish whether involvement of AI in RCT improves the standard of care. Many factors need to be considered to justify conducting AI RCTs, to risks research participants imposed in studies. Fine subject selection and equal distribution of risks and benefits across different populations must also be considered. Determine risk threshold. Data monitoring and management of the risks intervention requirements (full assessment, based only on study data, etc.) Management of passively collected data (e.g., the content of text messages) vs. predictive algorithms is still under development.	Grote, 2021 Jacobson et al., 2020	REBs should clearly define how AI may reduce the trial burden and improve the benefit-risk ratio in a research project. REBs should ensure that participants are not at a higher risk of being part of a minority in a population. REBs should identify and recommend appropriate measures to mitigate the specific risks that are embedded in AI in RCT.
Safety and security (End user-centered)	The research project and technologies used should not pose any harm to participants. These issues should be evaluated according to users' perspectives; the assessment should consider the reality's context. REB's lack of understanding of AI models makes assessing their impacts on safety difficult. Measures should be established to counteract negative impacts. Anticipate the implications of AI use (human protection, legal act, etc.)	Coeckelbergh et al., 2016 Coeckelbergh et al., 2016 Nebeker et al., 2019 Chassang et al., 2021	Adequate risk mitigation for a new technology implies that the REB has a good knowledge of the technology and its impacts REBs and researchers should identify adverse effects of AI systems and their consequences that may harm participants; identify mechanisms to repair potential harms The possibility that the harm is physical or moral can be an issue
Transparency	This concept will become a challenge with AI's black box, making it difficult to explain each result generated.	Jacobson et al., 2020; Andreotta et al., 2021	Ensure that the research project is explained in a way that is understandable to participants.
Privacy and confidentiality	Confusion between governance and confidentiality protection mechanisms. Greater emphasis on governance to the detriment of specific considerations for confidentiality or other ethical issues. Scientific advancement and data quality could impact individuals' privacy by collecting data extensively and transferring them.	Samuel and Gemma, 2021 Samuel and Gemma, 2021 Jacobson et al., 2020; Gooding and Kariotis, 2021	Pay more attention to algorithm and software development, allowing to broaden analysis and ethical evaluation toward ethical considerations toward privacy and confidentiality. Questioning the limits of current anonymization techniques with the use of AI systems. Questioning the new harms that may result from breaches of privacy and violations of confidentiality. Develop mechanisms to prevent, limit, and, if necessary, repair the damage resulting from these new potential breaches of privacy and confidentiality.
Justice, equity, and fairness	Standard of fair representation	Grote, 2021	Results do not mention the distribution of research benefits from these technologies. To address this issue, REBs should: -Focus more on these issues to rebalance their approach, which is more centered on governance. -Put AI systems and their potential into question to reduce inequalities and strengthen health equity Access to research benefits should be investigated to ensure a return of individual results. Challenges of transmitting general results to the community.
Validity and effectiveness	Consensus to appreciate the normative implications of AI technologies: -technology development -application of technology in real-time conditions The challenging aspect of understanding black box	McCradden et al., 2020c Ienca and Ignatiadis, 2020	REBs do not currently have an effective method to evaluate the validity of results generated by AI. REBs need the right tools to ensure that the expected aims of AI systems are achievable. The development of the system should meet the concrete needs of the populations targeted by the technology In an actual situation, the potential for transformation of the practice and the care offered must be ensured. Adaptations can be complex; in practice, modifications to the protocol are more difficult considering the nature of AI.