香京julia种子在线播放

Modern artificial intelligence (AI) is the cornerstone of the fourth industrial revolution. Successes in data availability, algorithm design, and processing power (Craglia et al., 2018) have empowered AI systems to make dramatic impacts in disparate sectors including transportation, education, agriculture, public services, finance, and healthcare (Artificial Intelligence for Africa: An Opportunity for Growth, Development, and Democratisation, 2018).

The varying degrees of autonomy with which AI systems can operate distinguish it from other emerging technologies. The advantage of AI lies in its ability to process massive amounts of varied information, and thereby perform valuable functions or draw useful conclusions inspired by its interpretation of the information. However, the essence of its usefulness is also its most challenging feature. For example, machine learning is a common approach to AI system design in medicine. Instead of programming the system for all possible scenarios with specific instructions, when using machine learning, developers set a broad goal which the system uses to form its own instructions to achieve the goal through repeated experiments and self-research (Rachum-Twaig, 2020). As it processes information, the AI system adjusts the parameters by which it judges inputs to produce more accurate outputs, effectively programming itself (Townsend, 2020). These approaches usually produce more accurate systems and they also require less human control (Grimm et al., 2021). Alarmingly, this and similar approaches to AI system design remove the human element at key stages of development in a way which may complicate inquiries into the attribution of responsibility and liability. This becomes especially pronounced where the AI system is so complex that its operations become inscrutable to humans. These so-called “black-box” algorithms lack the transparency to fully audit how they came to the conclusions they did. In response to this issue, some developers have endeavoured to design and create ‘explainable’ AI systems and ways of ensuring transparency which would foster an environment of accountability and responsibility and create better evidence when determining liability (Ali et al., 2023).

Determining responsibility will be important in dealing with the social challenges of AI integration. Perc et al. (2019) investigates how AI systems will likely have to choose between acting in favour of one party’s interest over another in certain contexts and how this may influence how the technology evolves. Developers may be incentivised to produce systems which favour owners’ interests above users’ in order to drive sales. The solution may be to require that AI systems act in the interests of the broader community; however, this policy may create its own issues in that it will potentially disincentivise people from buying AI systems which will not protect their interests outright and therefore lead to a lower adoption and investment in AI systems overall. This approach may then fail to fully realise the safety gains which can be had by increased AI usage. Of course, as Perc et al. (2019) consider, another approach may be to leave such decisions for the AI system to decide itself, or simply leave it to chance. This approach, however, suffers from a lack of clear answers to questions of responsibility and liability for the outcomes of decisions. Robust regulation and thoughtful juristic approaches to AI challenges will be necessary to provide adequate responses to responsibility for actions in these cases. This will be vital to supporting the benefits of AI integration whilst properly addressing the risks of the technology. Specifically, in healthcare, AI systems show impressive potential to increase the overall efficiency of healthcare systems and to manage disease outbreaks (Owoyemi et al., 2020). Furthermore, these systems can increase the reach of initiatives, while supplementing an already overburdened sector (Pepper and Slabbert, 2011). However, healthcare institutions deal with patients who are at their most vulnerable, where an incorrect decision could prove fatal. In addition, healthcare practitioners are required to abide by particularly high ethical and legal standards which AI systems may not easily conform to. In particular, the black box nature of some algorithms may prevent physicians from providing enough information to their patients about their treatments to satisfy requirements of informed consent, the emergent abilities of AI systems also raise questions as to how they will be considered in relation to the usual standard of care expected of physicians, and medical liability may need to be redefined for AI use.

Many jurisdictions already have laws and regulations which would encompass AI technologies; however, the specific challenges of AI may mean that these regulations do not provide desirable results when they are relied upon. As a response to this, many jurisdictions outside Africa have begun drafting specific AI law and regulations (Sallstrom et al., 2019). Providing a proper response to the issues posed by AI use in healthcare is essential to providing legal certainty to all stakeholders. This will allow them to order their interactions with AI systems and create an environment of trust in relation to AI use (European Commission, Directorate-General of Communications Networks, Content and Technology, 2019). This trust will be important for the future of AI as a lack of trust could permanently harm the reputation of AI in healthcare, or lead to additional costs through inefficient regulation or repeated amendment (Floridi et al., 2018).

The aim of this article is to set the stage for legal development and policy initiatives in Africa by exploring the legal concepts relevant to the attribution of liability for AI harm. First, we begin by describing current developments and the use of AI in healthcare in Africa in Section 2. Then we discuss the concept of liability broadly in Section 3. In Section 4, we describe how AI presents novel challenges to liability determination, particularly the concept of personal liability. In Section 5, we review the different approaches to determining liability. We provide our concluding thoughts in Section 6.

AI systems in healthcare can perform tasks normally requiring human physicians (Joshi. and Morley, 2019). Most current uses are in diagnosis and screening; however, future systems could scan images, discover new drugs, optimise care pathways, predict positive treatment outcomes, and provide preventative advice (Joshi and Morley, 2019). Increased use of AI allows physicians to focus on tasks where, given the current state of technology, they cannot be replaced. Furthermore, AI could further broaden public health initiatives by increasing access and tracking disease outbreaks, while lowering the cost of care (Joshi and Morley, 2019).

For example, DeepMind’s AlphaFold is an AI system which accurately predicted the protein structures of the COVID-19 virus, being an important aspect of creating a vaccine (Jumper et al., 2020). This use could greatly reduce vaccine response times in the future. IBM’s Watson for oncology is another system which has been able to analyse genomic data of patients in light of medical data from vastly more journals than a person could process, so providing more personalised treatments with high accuracy rates (Chung and Zink, 2018).

While other jurisdictions are considering policy-level AI implementation in healthcare systems (Joshi,and Morley, 2019), Africa has had relatively little meaningful interaction with AI in healthcare both academically (Tran et al., 2019) and clinically (Owoyemi et al., 2020) and, currently, African countries are at a nascent stage in their AI regulatory policies (Townsend et al., 2023). This is despite AI’s utility in developing countries where AI systems could lead to better utilisation of resources and enable new, effective treatments and treatment management systems (Sallstrom et al., 2019). Furthermore, AI systems can provide overarching and effective treatment options that improve standards of living, improve direct patient care, maximise supply-chain efficiencies, reduce administrative tasks, and streamline and improve compliance measures (Sallstrom et al., 2019).

Even though relatively limited, there has been some AI system use in Africa. In South Africa, Vantage, a machine learning-based system developed by BroadReach Healthcare, was used to assess clinics’ performance and provide staffing and operational recommendations in HIV clinics in KwaZulu-Natal (Singh, 2020). Further, DrConnect, an application by Discovery Health, provides personal assessments of medical symptoms and advice and remote support using AI technology, by using information from wearable devices such as smartwatches, to give medical and lifestyle advice (Singh, 2020). In Ghana, MinoHealth AI labs have used AI systems for automated diagnostics, forecasts, and prognostics. Also, BareApp is using specialised AI technology to diagnose skin disease and suggest treatments (Eke et al., 2023). In Uganda, AI is being merged with other technologies to develop a specialised system in the management of female chronic diseases (Eke et al., 2023). In Nigeria, Ubenwa is using AI to improve the diagnosis of birth asphyxia in low-resource settings (Owoyemi et al., 2020). Also in Nigeria, AI is proving effective in the identification of fake drugs (Owoyemi et al., 2020).

These examples illustrate the growing use and development of AI systems in Africa. However, as this use grows, it will be vital that African countries position themselves to take full advantage of AI’s benefits. Legal regulation will be especially important in directing AI system use and development by providing legal certainty by the formation of proper policies and regulations. A main concern though will be the determination of liability for AI harm.

The nature of emerging technologies is that we need time to understand them and develop policies and regulations which will encourage equitable use (Calo, 2015). AI in healthcare is no different. While AI has the potential to positively influence healthcare, its implementation must necessarily be coupled with appropriate safeguards to minimise risks of harm (European Commission, Directorate-General for Justice and Consumers, 2019). Specific to AI, unforeseeable risks may still arise in apparently well-trained systems where performance is being improved (World Health Organisation, 2021). As it currently stands, when risks arise, our existing policies and regulations will be the basis of determining who is responsible and liable for the harm caused. Assessing whether these policies and regulations are sufficient to properly determine responsibility will be important, as the determination of responsibility plays an important role in determining the basis of legal liability for AI conduct and garnering trust in AI usage more broadly. Currently, this will largely depend on civil liability rules.

Generally, civil liability provides the dual purpose of providing a means for victims of harm to be compensated, while also providing an economic incentive for those held liable to avoid continuing harmful conduct (Buiten et al., 2021). Accordingly, these rules are an important means of protecting patients and providing clarity to businesses on how they may innovate and operate their products (Buiten et al., 2021). However, the varying complexity of AI systems, system updates, algorithms which change from environmental input, and cyber-security concerns may make it difficult to justify claims for compensation and to provide clear pathways for victims to bring claims (European Commission, Directorate-General for Justice and Consumers, 2019). It is also unclear whether the rationale behind current liability regimes will be effective in dealing with AI harm. For example, where AI systems make decisions, it may be difficult for a plaintiff to find a suitable defendant or for a court to determine the standard of care to be expected from an AI system. Therefore, it is currently unclear how current liability regimes will consider AI harm in healthcare.

Proper liability policy formation will consider the outcomes of current liability rules but, in addition, it must necessarily consider the impact which the policy will have on the development and use of AI in the future. This means tailoring policy towards managing AI-specific risks while encouraging positive uses. For example, a lack of legal certainty and fear of unreasonable legal penalties for relying on AI recommendations may discourage healthcare practitioners from using AI systems as active participants in treatment, relegating AI systems’ role to the mere confirmation of decisions made by healthcare practitioners (World Health Organisation, 2021). On the contrary, removing penalties may encourage AI systems use; however, this position may be tenable only where existing issues of accountability and responsibility are properly considered.

Of particular concern in healthcare should be determining how an AI system will form part of the standard of care. Such a determination will be essential for providing sufficient information for physicians and patients to make decisions about relying on the technology (World Health Organisation, 2021). The decision of the physician is important as he/she will also likely be responsible for the proper operation, monitoring and maintenance of the technology (Bertolini and Episcopo, 2021), and their decision could be consequential for their employer through vicarious liability (World Health Organisation, 2021).

A concern specific to Africa is that many policy frameworks which would guide the development of AI systems are created in environments outside of Africa. Moreover, a lack of access to high quality data sets and limitations in infrastructure could lead to the use of algorithms which are predominantly developed outside of Africa. These could be potentially prejudicial as they may not be properly designed to work in low-resource environments (World Health Organisation, 2021). Therefore, liability policies will need to consider that developers are situated outside of Africa, and that algorithms are adapted for, rather than designed for, the African context.

The role of an AI policy framework should be to prevent AI harm and to promote AI innovation, following a risk-based, rights-preserving, agile, adaptive, and innovation-supporting regulatory approach (Townsend et al., 2023). Robust and effective regulation will provide important guiding principles for the development and implementation of AI systems in healthcare in Africa (World Health Organisation, 2021). Legal certainty will provide routes for compensation for patients and ensure accountability and responsibility through integration and innovation in the healthcare system.

AI systems’ successful imitation of qualities normally associated with humans has bolstered the inquiry into AI personhood (Abbott and Sarch, 2019). A crucial development in support of AI personhood has been the ability to program generalised goals into AI systems. This approach is markedly different to traditional software as the AI system is programmed to decide what steps it would take to achieve its goal, instead of being programmed with specific, step-by-step instructions (Bostrom and Yudkowsky, 2014). This goal-directed behaviour is what powered IBM’s Deep Blue chess robot. Programmers surpassed their own chess skills by encoding the rules of the game into Deep Blue and relying on its superior processing power to find ways of “winning” which the developer would not be able to do (Bostrom and Yudkowsky, 2014). Should this be enough to draw the necessary philosophical conclusions on AI personhood, it is clear that the legal implications would be substantial (Solum, 1992). Where AI systems are considered persons, even in limited form, they may be held responsible for their actions in their own capacity.

However, the utility of recognising AI personhood should not replace thoughtful policy formation. An AI system fulfilling roles normally delegated to humans does not mean that personhood necessarily follows (Thaldar and Naidoo, 2021). This may be illustrated by the recent granting of a patent in South Africa where the sole inventor was AI. Although some would consider “inventing” to be a human characteristic, without the ability to fully contain human emotion and capacity to engage in relationships, it is difficult to see such an AI system as more than a “special species of legal object that has the ability to invent” (Thaldar and Naidoo, 2021). As AI becomes more autonomous, legal rules can be developed to allow for special treatment of AI systems, which would be comparable to the legal rules that provide for the special treatment of animals (Thaldar and Naidoo, 2021).

While it is generally agreed that current AI systems are not capable of being considered legal persons, more sophisticated, generalised, and autonomous systems may change this assumption (Solum, 1992). Current systems can be changed, created, or completely deleted like any other software, but where AI systems enjoy a degree of personhood, our relationship with them may become far more complicated. Legally, the granting of AI personhood would aid plaintiffs of AI harm in that they could gather evidence from the AI system through its examination as a witness (Chung and Zink, 2018). However, this benefit may be somewhat limited in systems that lack transparent reasoning.

More definitively, some scholars insist that a separate legal personality for AI systems will never be necessary (European Commission, Directorate-General for Justice and Consumers, 2019). They contend that even fully autonomous systems’ actions are better attributed to individuals or other legal persons than to the system itself (European Commission, Directorate-General of Communications Networks, Content and Technology, 2019).

An important consideration is that AI systems’ lack of abstract thought limits their comparison to human personhood and decision-making, particularly in healthcare. Whereas human decision-making in healthcare is largely justified by morality, AI systems lack moral input in decision-making (Chung and Zink, 2018). Moral considerations become vitally important in healthcare and resource-scarce environments where circumstances require difficult decisions to be justified, usually with reference to moral ideals. Therefore, we suggest that in lacking moral capacity, AI systems could be limited in how they could be held accountable if they were considered persons or could lack prerequisites to make decisions in moral contexts.

For scholars who consider AI more than a tool, the lack of moral input is an issue they contend with (Bashayreh et al., 2021). Dignum (2017) suggests that even AI systems acting as assistants may inherit a moral framework for decision-making through incorporating the values of their engineers. However, a mere copy of an engineer’s morals may not necessarily lead to satisfactory results as AI systems may not apply moral lessons to their environments in the same way as humans (Bostrom and Yudkowsky, 2014). Dignum contends that identifying and analysing these imbued values will nevertheless improve system performance (Dignum, 2017). This would also ensure that incorporated morals are interpreted in an acceptable way, meaning that, as these systems become more autonomous and powerful, moral assessment may become an essential component of their decision-making, especially in a field such as healthcare (Dignum, 2017).

Accordingly, there is some possibility of future AI systems bearing some form of personhood (Solum, 1992). However, conferring even a limited form of personhood on AI systems presents further practical difficulties. For example, as is commonly suggested, a limited form of personhood may be imbued on AI systems through the extension of the principal-agent relationship. In determining responsibility, however, the standards which apply when adjudicating AI system conduct and under what circumstances AI systems would be considered liable for their conduct would remain unclear. This will be discussed further in the section on the principal-agent relationship below.

Further, a final practical issue of attributing liability directly to AI systems is that it leaves no clear pathways for compensation of victims (Bashayreh et al., 2021). As AI systems are currently incapable of ownership, there are no assets that a victim could claim. To remedy this situation, some scholars have suggested the introduction of an insurance scheme funded by developers from which victims may claim (European Commission, Directorate-General of Communications Networks, Content and Technology, 2019). However, such a scheme may not adequately replace clear and fair liability rules and could lead to high administrative costs, so defeating the cost-saving benefits of a clear claim process (European Commission, Directorate-General of Communications Networks, Content and Technology, 2019). Furthermore, there is a lack of guidance on the value of AI insurance policies as there are no standards against which to assess risk or begin a cost analysis (Bertolini and Episcopo, 2021).

The subsections below discuss the main approaches to the attribution of liability for harm caused by AI systems in healthcare. Section 5.1 broadly considers the extension of the principal–agent relationship to include AI systems and the consequences of such an extension. Section 5.2 deals with AI as a product and how consumer protection law standards may be applied to AI system harm. We then comment on current fault-based liability regimes as they apply to AI systems in Section 5.3. This leads to a discussion of efforts to use strict liability to attribute liability for AI harm in Section 5.4. In Section 5.5, we consider an approach to AI harm focusing on improving AI system use in healthcare through reconciliatory forums.

The assimilation of AI technologies in the African healthcare sector is an unprecedented juncture in the continent’s journey towards equitable and advanced medical care. As AI solutions make inroads into African medical establishments, they bring along a multitude of autonomy and opacity issues, challenging the longstanding ethical pillars and legal norms ingrained in the diverse cultures of the continent. The quintessential medico-legal principle of informed consent is now juxtaposed against the intricate algorithms of AI, challenging the very essence of transparency and patient understanding. Similarly, the increasing autonomy of AI systems amplifies the intricacies of liability, pushing the boundaries of traditional legal frameworks.

In this article, we tried to provide the reader with an overview of the legal concepts relevant to the issue of AI and liability in healthcare. We started with the contemplation of AI personhood. While captivating, we suggest that it poses substantial challenges in an African context, particularly when addressing tangible redress mechanisms for AI-induced mishaps. Next, the principal-agent framework, although providing a modicum of accountability, could inadvertently stifle the AI adoption rate by placing considerable responsibilities upon local medical practitioners. While product liability law offers another plausible approach, it struggles to categorise the continually evolving nature of AI in the static confines of conventional product definitions. Alternative strategies, such as risk-based liability may offer clearer paths in contexts where fault determination proves onerous. Yet, they too grapple with ensuring specificity and justice. Strict liability, offering more transparent compensation mechanisms, raises concerns about economic implications, reputational risks and, most critically, the challenge of harmonising such policies across Africa’s diverse legal landscapes.

An approach based on reconciliation rather than liability potentially provides the best environment for a regulatory sandbox; however, reconciliation in the context of AI-driven harm in the healthcare context lacks the same level of scholarship as the approaches based on liability. We suggest that reconciliation offers much potential and deserves more academic attention.

In distilling these insights, it is evident that Africa’s AI journey in healthcare is not solely a scientific or medical transition. It also requires profound legal reflection and evolution.