The current normative guidelines do not make up for the few AI-related guidelines (Aymerich-Franch and Fosch-Villaronga, 2020). However, in addition to the ethical standards used as a basis for AI use guidelines, the UN published a first set of guidelines to regulate AI (Chassang et al., 2021). Many projects, like the Human Brain Project (HBP), took the initiative to encourage discussions from different parties to anticipate issues that could result from their research (Stahl and Coeckelbergh, 2016; Aicardi et al., 2018, 2020). Researchers and developers can access tools that help orient their reflections on their responsible use of technology (Aymerich-Franch and Fosch-Villaronga, 2020). Furthermore, the implementation of ethical approval committees (i.e., Human Research Ethics Committees in Australia) that uses a soft-governance model, which leans toward ethical regulation and is less restrictive than legal regulations, would help prevent studies or companies abuse their participants or users (Andreotta et al., 2021). Many are contemplating using digital health ELSI to encourage the implementation of ethical standards in AI when laws and regulations are lacking in it (Nebeker et al., 2019).

Articles have mentioned many leading countries in AI research. Supplementary Table 1 showcases the progress and effort the European Union (EU) and other countries have made regarding AI regulation. The countries, alongside the EU, that were often mentioned throughout our final sample were the following: Australia, Canada, China, the European Union, the United Kingdom, and the United States. Since this information is strictly from our selected articles, some information was unavailable. While noticeable progress is being made regarding AI development and regulation, most countries have shown little indication, if any, of AI research ethics.

While guidelines and norms are shifting to fit AI standards, many questions on moral status and rights are raised to adapt to this new reality. Authors argue that we cannot assign moral agency to AI and robots. There are multiple reasons for it: robots do not seem capable of solving problems ethically (Stahl and Coeckelbergh, 2016), AI's lack of explanation regarding its generated results, and the absence of willingness to choose (Farisco et al., 2020) which might impact decision making in research ethics.

Rights are attributed to different living entities. For instance, in the EU, the law protects animals as sentient living organisms and unique tangible goods. Their legal status also obliges researchers not to harm animals during research projects, making us question the status and rights we should assign AIS or robots (Chassang et al., 2021). Indeed, Miller pointed out that having a machine at one's disposal raises questions on human-machine relationships and the hierarchical power it might induce (Miller, 2020).

The literature invoked the following classic theories: the Kantian-inspired model, utilitarianism, principlism (autonomy, beneficence, justice, and non-maleficence), and the precautionary principle. Table 3 illustrates these essential ethical approaches found in our final sample, along with their description in terms of AI research ethics.

Public education and ethical training implementation could help governments spread awareness and sensitize people regarding research ethics in AI (Cath et al., 2018). Accountability of AI regulation and decision-making should not strictly fall into stakeholders' hands but also be based on solid legal grounds (Chassang et al., 2021). Digital mental health apps and other institutions will now be attributed responsibilities that have usually been acclaimed to professionals or researchers using the technology (i.e., decision-making, providing users with enough tools to understand and use products, being able to help when needed, etc.) (Gooding and Kariotis, 2021). Scientists and AI developers must not throw caution to the wind regarding the possibility that biased algorithms could be fed to AI models (Ienca and Ignatiadis, 2020). Clinicians will have to tactfully manage to inform patients of the results generated by machine learning (ML) models while considering their risk of error and bias (Jacobson et al., 2020). It is still vague to attribute responsibility to specific actors. However, it is necessary to have different groups work together to tackle the problem (Meszaros and Ho, 2021; Samuel and Gemma, 2021). Some consider that validity, explainability, and open-source AI systems are some of the defining points that lead to responsibility. With the advancement of technologies and its gain of interest, the sense of social responsibility also increased. Indeed, every actor must contribute to making sure that these novel technologies are developed and used in an ethical matter (Nebeker et al., 2019; Aicardi et al., 2020).

An important issue with AIS usually raised is the explainability of results. Deep learning (DL) is another type of ML with more extensive algorithms that encloses data with a broader array of interpretations (Chassang et al., 2021). This makes it harder to explain how DL and AI models reached a particular conclusion (Ienca and Ignatiadis, 2020; Jacobson et al., 2020). This poses transparency issues that are challenging to participants (Grote, 2021).

Since AI is known for its ‘black-box' aspect, where results are difficult to justify, it is difficult to fully validate a model with certainty (Ienca and Ignatiadis, 2020). Deciding to monitor research participants closely could help validate results which, in theory, would bring more accurate results. However, close monitoring could also have a negative effect by influencing participants' decisions based on whether they mind being monitored or not. This event could, as a result, produce more inaccurate results (Jacobson et al., 2020). Furthermore, it could be more challenging in certain contexts to promote validity when journals and funding bodies favor new and innovative studies over ethical research on AI, even if the latter is being promoted (Samuel and Gemma, 2021).

According to the White House Office of Science and Technology Policy (OSTP), transparency would help solve many ethical issues (Cath et al., 2018). Transparency allows research participants to be aware of a study's different outlooks and comprehend them (Sedenberg et al., 2016; Grote, 2021). The same goes for new device users (Chassang et al., 2021). AI models (i.e., products, services, apps, sensor-equipped wearable systems, etc.) produce a great deal of data that does not always come from consenting users (Ienca and Ignatiadis, 2020; Meszaros and Ho, 2021). Furthermore, AI's black-box imposes a challenge to obtain informed consent since the lack of explainability of AI-generated results might not allow participants to have enough information to give out their informed consent (Jacobson et al., 2020; Andreotta et al., 2021). Thus, it is essential to make consent forms easy to understand for the targeted audience (Nebeker et al., 2019).

However, the requirement to get informed consent could lead to other less desirable implications. Some argue that requiring authorization for all data, especially studies that hold a vast set of data, might lead to data bias and a decrease in data quality because it only entices a specific group of people to give out consent which leaves out a significant part of the population (Ford et al., 2020).

While the levels of privacy differ from one scholar to another, the concept of privacy remains a fundamental value to human beings (Andreotta et al., 2021). Through AI and robotics, data can be seen as attractive commodities which could compromise privacy (Cath et al., 2018). Researchers are responsible for keeping participants unidentifiable while using their data (Ford et al., 2020). However, data collected from many sources can induce a higher risk of identifying people. While pursuing their research study, ML researchers still struggle to comply with privacy guidelines.

Some authors argue that the priority might be to consider whether predictions from a specific machine learning model are appropriate for informing decisions about a particular intervention (Jacobson et al., 2020). Others advocate carefully constructing the planned interventions so research participants can understand them (Grote, 2021).

The extent to which researchers should provide extensive information to participants is not evident among stakeholders. So far, research suggests that there is no clear consensus among patients on whether they would want to know this kind of information about themselves (Jacobson et al., 2020). Hence the question remains whether patients want to see if they are at risk, mainly if they cannot be told why, as factors included in machine learning models generally cannot be interpreted as having a causal impact on outcomes (Jacobson et al., 2020). Therefore, sharing information from an uninterpretable model may adversely affect a patient's perception of their illness, confuse them, and immediate concerns about transparency.

The analysis of harms and potential benefits is critical when assessing human research. REBs are well concerned with this assessment to prevent unnecessary risks and promote benefits. Considerations of the potential benefits and harms to patient-participants are necessary for future clinical research, and REBs are optimally positioned to perform this assessment (McCradden et al., 2020c). Additional considerations like benefit/risk ratio or effectiveness and the systematic process described previously are necessary. Risk assessments could have a considerable impact in research involving mobile devices or robotics because preventive action and safety measures may be required in the case of imminent risks. Thus, REB risk assessment seems very important (Jacobson et al., 2020).

Approaching AI research ethics through user-centered design can represent an interesting avenue to understand better how REB can conduct risk/benefices assessment. For researchers, involving users in the design of AI research is likely to promote better research outcomes. Hence, this can be reached by investigating how AI research is actually meeting users' needs and how this may generate intended and unintended impacts on them (Chassang et al., 2021; Gooding and Kariotis, 2021). Indeed there is insufficient reason to believe that AI research will produce positive benefits unless it is evaluated with a focus on patients and situated in the context of clinical decision-making (McCradden et al., 2020c). Consequently, REBs might focus on the broader societal impact of this research (Chassang et al., 2021).

Safety and security are significant concerns for AI and robotics, and their assessment may rely on end-users' perspectives. To address the safety issue, it is not sufficient for robotics researchers to say that their robot is safe based on literature and experimental tests. It is crucial to find out about the perception and opinions of end-users of robots and other stakeholders (Coeckelbergh et al., 2016). Testing technology in real-life scenarios is vital for identifying and adequately assessing technology's risks, anticipating unforeseen problems, and clarifying effective monitoring mechanisms (Cath et al., 2018). On the other hand, there is a potential risk that an AIS misleads the user in realizing a legal act.

Validity is a crucial consideration and one on which there is consensus to appreciate the normative implications of AI technologies. To this end, it is necessary for research ethics that researchers' protocols be explicit about many elements and describe their validation model and performance metrics in a way that allows for assessment of the clinical applicability of their developing technology (McCradden et al., 2020b). In addition, in terms of validity, simulative models have yet to be appropriately compared with standard medical research models (including in vitro, in vivo, and clinical models) to ensure they are correctly validated and effective (Ienca and Ignatiadis, 2020). Considering many red flags raised in recent years, AI systems may not work equally well with all sub-populations (racial, ethnic, etc.). Therefore, AI systems must be validated for different subpopulations of patients (McCradden et al., 2020b).

Demonstration of value is essential to ensure the scientific validity of the claims made for technology but also to attest to the proven effectiveness once deployed in a real-world setting and the social utility of a technology (Nebeker et al., 2019). When conducting a trial for the given AI system, the main interest should be to assess its overall reliability, while the interaction with the clinician might be less critical (Grote, 2021).

Transparency entails understanding how technology behaves and establishing thresholds for permissible (and impermissible) usages of AI-driven health research. Transparency requires clarifying the reasons and rationales for the technology's design, operation, and impacts (Friesen et al., 2021). Identified risks should be accompanied by detailed measures intended to avoid, reduce, or eliminate the risks. The efficiency of such efforts should be assessed upstream and downstream as part of the quality management process. As far as possible, testing methods, data, and assessment results should be public. Transparent communication is essential to make research participants, as well as future users aware of the technology's logic and functioning (Chassang et al., 2021).

The implications presented in Table 4. Seem to encourage REBs to adopt a more collaborative approach to grasp a better sense of reality in different fields. The analysis also showed that data bias is a flagrant problem whether AI is used or not and that this discriminatory component should be taken care of to avoid emphasizing the problem with AI. Informed consent is another value that REBs prioritize and will have to be adapted to AI because new information might have to be disclosed to participants. Safety and security are always essential to consider. However, other measures will be implemented with AI to ensure that participants are not set in danger. One of the main aspects of AI is data sharing and the risk that this might breach participants' privacy. The methods put in place now might not be suitable for AI's fast evolution. The questions of justice, equality, and fairness that have not been resolved in our current society will also have to be instigated in the AI era. Finally, the importance of validity was raised numerous times. Unfortunately, REBs do not have the right tools to evaluate AI. It will be necessary for AI to meet the population's needs. Furthermore, definitions of specific values and principles that REBs usually respond to will have to be reviewed and adapted according to AI.

According to our review, there is a disproportionate focus on consent before other ethical issues. Authors argue that the big piece the REBs ask for relies on consent, not the AI aspect of the project. This finding suggests that narrowing AI research ethics around consent concerns remains problematic. In some stances, the disproportionate focus on consent, along with the importance REBs place on consent forms and participant information sheets, has settled how research ethics is defined, e.g., viewed as a proxy for ethics best practice, or in some cases, as an ethics panacea (Samuel and Gemma, 2021).

Authors report a lack of knowledge for safety review. It appears clear that REBs may not have the experience or expertise to conduct a risk assessment to evaluate the probability or magnitude of potential harm. Similarly, the training data used to inform the algorithm development are often not considered to qualify as human subjects research, which – even in a regulated environment – makes a prospective review for safety potentially unavailable (Nebeker et al., 2019).

On the other hand, REBs lack appropriate assessing processes for assessing whether AI systems are valid, effective, and apposite. The requirement to evaluate the evidence of effectiveness adds to a range of other considerations with which REBs must deal (i.e., the protection of participants and the fairness in the distribution of benefits and burdens). Therefore, there is still much to be done to equip REBs to evaluate the effectiveness of AI technologies, interventions, and research (Friesen et al., 2021).

Researchers point to a disproportionate focus on data privacy and governance before other ethical issues in medical health research with AI tools. Focus on privacy and data governance issues warrants further attention, as privacy issues may overshadow other issues. Indeed it seems problematic and led to a narrowing of ethics and responsibility debates being perpetuated throughout the ethics ecosystem, often at the expense of other ethical issues, such as questions around justice and fairness (Samuel and Gemma, 2021). REBs appear to be less concerned about the results themselves. One explained that when reviewing their AI-associated research ethics applications, REBs focus more on questions of data privacy than other ethical issues, such as those related to the research and the research finding. Others painted a similar picture of how data governance issues were a centralized focus when discussing their interactions with their REB. According to these stakeholders, REBs focus less on the actual algorithm than how the data is handled, and the issue remains around data access and not about the software (Samuel and Gemma, 2021).

Lack of expertise appears to be a significant concern in our results. Indeed, even when there is oversight from a research ethics committee, authors observe that REB members often lack the experience or confidence regarding particular issues associated with digital research (Samuel and Derrick, 2020).

Some authors advocate that ML researchers should complement the membership of REBs since they are better situated to evaluate the specific risks and potential unintended harms linked to the methodology of ML. On the other hand, REBs should be empowered to fulfill their role in protecting the interests of patients and participants and enable the ethical translation of healthcare ML (McCradden et al., 2020c). However, we can notice that researchers expressed different views about REBs' expertise. While most acknowledged a lack of AI-specific proficiency, for many, this remains straightforward because the ethical issues of their AI research were nonexceptional compared to other ethics issues raised by “big data” (Samuel and Gemma, 2021).

Limits of process and regulation are another concern faced by REBs, including a lack of consistency in decision-making within and across REBs, a lack of transparency, poor representation of the participants and public they are meant to represent, insufficient training, and a lack of measures to examine their effectiveness (Friesen et al., 2021). There are several opinions on the need for and the effectiveness of REBs, with critics lamenting excessive bureaucracy, lack of reliability, inefficiency, and, importantly, high variance in outcomes (Prunkl et al., 2021). To address the existing gap of knowledge between different fields, training could be used to help rebalance this and ensure sufficient expertise for all research experts to pursue responsible innovation (Stahl and Coeckelbergh, 2016).

Researchers described the lack of standards and regulations for governing AI at the level of societal impact; the way that ethics committees in institutions work is still acceptable. But there is a need for another level of thinking that combines everything and does not look at one project simultaneously (Samuel and Gemma, 2021).

Finally, researchers have acknowledged the lack of ethical guidance, and some REBs report feeling ill-equipped to keep pace with rapidly changing technologies used in research (Ford et al., 2020).

While AI tools provide new opportunities to enhance medical health research, there is an emerging consensus among stakeholders regarding bias concerns in AI data, particularly in clinical trials. Since bias can worsen pre-existing disparities, researchers should proactively target a wide range of participants to establish sufficient evidence of an AI system's clinical benefit across different populations. To mitigate selection bias, REBs may require randomization in AI clinical trials. To achieve this, researchers must start by collecting more and better data from social minority groups (Grote, 2021). Also, awareness of biases concerns should be taken into account in the validation phase, where the performance of the AI system gets measured in a benchmark data set. Hence it is crucial to test AI systems for different subpopulations. Therefore, affirmative action in recruiting research participants in AI RCTs deems us ethically permissible (Grote, 2021). However, authors reported that stakeholders might encounter challenges accessing needed data in a context where severe legal constraints are imposed on sharing medical data (Grote, 2021).

Vulnerable populations require excellent protection against risks they may face in research.

When involving vulnerable populations, such as those with a mental health diagnosis, in AI medical health research, additional precautions should be considered to ensure that those involved in the study are duly protected from harm – including stigma and economic and legal implications. In addition, it is essential to consider whether access barriers might exclude some people (Nebeker et al., 2019).

Another issue, which needs to be raised when considering critical practices and scope in AI research, relates to fair representation, diversity, and inclusion. According to Grote, one should explore concerns for the distribution of research participants and representatives for the state, country, or even world region in which the AI system gets tested. Here the author advocates if we should instead aim for a parity distribution of different gender, racial and ethnic groups. Hence, he raised several questions to support the reflection of REB on diversity, inclusion, and fairness issues: How should the reference classes for the different subpopulations be determined? Also, what conditions must be met for fair subject selection in AI RCTs? And finally, when, if ever, is it morally justifiable to randomize research participants in medical AI trials? (Grote, 2021).

The aging population and scarcity of health resources are significant challenges healthcare systems face today. Consequently, people with disabilities, especially elders with cognitive and mental impairments, are the most affected. The evolving field of research with assistive robots may be useful in providing care and assistance to these people. However, robotics research requires specific guidance when participants have physical or cognitive impairments. Indeed particular challenges are related to informed consent, confidentiality, and participant rights (Battistuzzi et al., 2021). According to some authors, REBs should ask several questions to address these issues: Is the research project expected to enhance the quality of care for the research participants? What is/are the ethical issue/s illustrated in this study? What are the facts? Is any important information not available in the research? Who are the stakeholders? Which course of action best fits with the recommendations and requirements set out in the “Ethical Considerations” section of the study protocol? How can that course of action be implemented in practice? Could the ethical issue/s presented in the case be prevented? If so, how? (Battistuzzi et al., 2021).

Which ethical and social issues may neurorobotics raise, and are mechanisms currently implemented sufficiently to identify and address these? Is the notion that we may analyze, understand and reproduce what makes us human rooted in something other than reason (Aicardi et al., 2020)?

A good understanding of the process behind AI/ML tools might be of interest to REBs when assessing the risk/benefit ratio of medical research involving AI. However, there seems to be a lack of awareness of how AI researchers gain results. Authors argue that it would not be impossible to induce perception about the external environment in the neuron culture and to interpret the signals from the neuron culture as motor commands without a basic understanding of this neural code (Bentzen, 2017). Indeed, when using digital health technologies, the first step is to ask whether the tools, be they apps or sensors, or AI applied to large data sets, have demonstrated value for outcomes. One should ask whether they are clinically effective, or if they measure what they purport to measure (validity) consistently (reliability), and finally, if these innovations also improve access to those at the highest risk of health disparities (Nebeker et al., 2019).

Indeed, the ethical issues of AI research raise major questions within the literature. What may seem surprising at first sight is that the body of literature is still relatively small and appears to be in an embryonic state regarding the ethics of the development and use of AI (outside the scope of academic research). The literature is thus more concerned with the broad questions of what constitutes research ethics in AI-specific research and with pointing out the gaps in normative guidelines, procedures, and infrastructures adapted to the oversight of responsible and ethical research in AI. Perhaps unsurprisingly, most of the questions related to study within the health sector. This is to be expected given the ascendancy of health in general within the research ethics field (Faden et al., 2013). Thus, most considerations relate to applied health research, the implications for human participants (whether in digital health issues, research protocols, or interactions with different forms of robots), and whether projects should be subject to ethics review.

Specifically in AI-specific research ethics, interestingly, traditional issues of participant protection (including confidentiality, consent, and autonomy in general) and research involving digital technologies intersect and are furthered by the uses of AI. Indeed, as AI requires big data and behaves very distinctly from other technologies, the primary considerations raised by the body of literature studied were predominantly classical in AI ethics but contextualized and exacerbated within research ethics practices. For instance, one of the most prevalent ethical considerations raised and discussed was privacy and the new challenges regarding the massive amount of data collected and its use. If a breach of confidentiality were to happen or data collection would lead to discovering further information, this would raise the possibility of harming individuals (Ford et al., 2020; Jacobson et al., 2020). In addition, informed consent was widely mentioned and focused on transparency and explainability when the issues were AI-specific. Indeed, AI's black-box issue of explainability was raised many times. This is a challenge because it is not always easy to justify the results generated by AI (Jacobson et al., 2020; Andreotta et al., 2021). This then poses a problem with transparency. Indeed, participants expect to have the necessary information relevant to the trial to make an informed and conscious decision regarding their participation. Not having adequate knowledge to share with participants might not align with informed consent.

Furthermore, another principle was brought up many times, which was responsibility. Responsibility is shared chiefly between the researcher and the participant (Gooding and Kariotis, 2021). Now that AI is added to the equation, it has become harder to determine who strictly should be held accountable for the occurrence of certain events (i.e., data error) and in what context (Meszaros and Ho, 2021; Samuel and Gemma, 2021). While shared responsibility is an idea many share and wish to implement, it is not easy. Indeed, as seen in Figure 3, many stakeholders (e.g., lawmakers, AI developers, AI users) may participate in responsibility sharing. However, much work will have to be put into finding a fair way to share responsibility between each party involved.