Explainability in Artificial Intelligence

Recent debates over artificial intelligence have focused on its perceived lack of interpretability and explainability. I argue that these notions fail to capture an important aspect of what end-users—as opposed to developers—need from these models: what is needed is systematicity, in a more demanding sense than the compositionality-related sense that has dominated discussions of systematicity in the philosophy of language and cognitive science over the last thirty years. To recover this more demanding notion of systematicity, I distinguish between (i) the (...) systematicity of thinkable contents, (ii) the systematicity of thinking, and (iii) the ideal of systematic thought. I then deploy this distinction to critically evaluate Fodor’s systematicity-based argument for the language of thought hypothesis before recovering the notion of the systematicity of thought as a regulative ideal, which has historically shaped our understanding of what it means for thought to be rational, authoritative, and scientific. To assess how much systematicity we need from AI models, I then argue, we must look to the functions of systematizing thought. To this end, I identify five functions served by systematization, and show how these can be used to arrive at a dynamic understanding of the need to systematize thought that can tell us what kind of systematicity is called for and when. (shrink)

A key assumption fuelling optimism about the progress of Large Language Models (LLMs) in modelling the world is that the truth is systematic: true statements about the world form a whole that is not just consistent, in that it contains no contradictions, but cohesive, in that the truths are inferentially interlinked. This holds out the prospect that LLMs might rely on that systematicity to fill in gaps and correct inaccuracies in the training data: consistency and cohesiveness promise to facilitate progress (...) towards comprehensiveness in an LLM’s representation of the world. However, philosophers have identified reasons to doubt that the truth is systematic across all domains of thought, arguing that in normative domains, in particular, the truth is not necessarily systematic. I argue that insofar as the truth in normative domains is asystematic, this renders it correspondingly harder for LLMs to make progress, because they cannot rely on the consistency and cohesiveness of the truth to work towards comprehensiveness. And the less LLMs can rely on the systematicity of truth, the less we can rely on them to do our practical deliberation for us, as there is correspondingly more of a role for human agency in navigating asystematic normative domains. (shrink)

The increasing use of artificial intelligence (AI) systems in our daily life through various applications, services, and products explains the significance of trust/distrust in AI from a user perspective. AI-driven systems have significantly diffused into various fields of our lives, serving as beneficial tools used by human agents. These systems are also evolving to act as co-assistants or semi-agents in specific domains, potentially influencing human thought, decision-making, and agency. Trust/distrust in AI plays the role of a regulator and could significantly (...) control the level of this diffusion, as trust can increase, and distrust may reduce the rate of adoption of AI. Recently, varieties of studies have paid attention to the variant dimension of trust/distrust in AI, and its relevant considerations. In this systematic literature review, after conceptualization of trust in the current AI literature review, we will investigate trust in different types of human-Machine interaction, and its impact on technology acceptance in different domains. In addition to that, we propose a taxonomy of technical (i.e., safety, accuracy, robustness) and non-technical axiological (i.e., ethical, legal, and mixed) trustworthiness metrics, and some trustworthy measurements. Moreover, we examine some major trust-breakers in AI (e.g., autonomy and dignity threat), and trust makers; and propose some future directions and probable solutions for the transition to a trustworthy AI. (shrink)

There are calls from various sectors, including the popular press, industry, and academia, to incorporate artificial intelligence (AI)-based technologies in general, and large language models (LLMs) (such as ChatGPT and Gemini) in particular, into various spheres of the South African higher education sector. Nonetheless, the implementation of such technologies is not without ethical risks, notably those related to bias, unfairness, privacy violations, misinformation, lack of transparency, and threats to autonomy. This paper gives an overview of the more pertinent ethical concerns (...) that may result from the deployment of various current AI technologies in the South African higher education context. It provides a broad overview of the relevant AI ethics literature and distills a few general AI-ethics principles that can serve as guidelines for the ethical development, adoption, and use of AI systems. Suggestions are made as to how these might be applied to mitigate the ethical concerns in the South African higher education context. Overall, it is argued that AI technologies should only be adopted if they offer demonstrable benefits to stakeholders affected by them and that care should be taken to ensure that any potential harms are adequately addressed. (shrink)

Human-centered explainable AI (HCXAI) advocates for the integration of social aspects into AI explanations. Central to the HCXAI discourse is the Social Transparency (ST) framework, which aims to make the socio-organizational context of AI systems accessible to their users. In this work, we suggest extending the ST framework to address the risks of social misattributions in Large Language Models (LLMs), particularly in sensitive areas like mental health. In fact LLMs, which are remarkably capable of simulating roles and personas, may lead (...) to mismatches between designers' intentions and users' perceptions of social attributes, risking to promote emotional manipulation and dangerous behaviors, cases of epistemic injustice, and unwarranted trust. To address these issues, we propose enhancing the ST framework with a fifth 'W-question' to clarify the specific social attributions assigned to LLMs by its designers and users. This addition aims to bridge the gap between LLM capabilities and user perceptions, promoting the ethically responsible development and use of LLM-based technology. (shrink)

What can contemporary machine learning (ML) models do? Given the proliferation of ML models in society, answering this question matters to a variety of stakeholders, both public and private. The evaluation of models' capabilities is rapidly emerging as a key subfield of modern ML, buoyed by regulatory attention and government grants. Despite this, the notion of an ML model possessing a capability has not been interrogated: what are we saying when we say that a model is able to do something? (...) And what sorts of evidence bear upon this question? -/- In this paper, we aim to answer these questions, using the capabilities of large language models (LLMs) as a running example. Drawing on the large philosophical literature on abilities, we develop an account of ML models' capabilities which can be usefully applied to the nascent science of model evaluation. Our core proposal is a conditional analysis of model abilities (CAMA): crudely, a machine learning model has a capability to X just when it would reliably succeed at doing X if it 'tried'. The main contribution of the paper is making this proposal precise in the context of ML, resulting in an operationalisation of CAMA applicable to LLMs. We then put CAMA to work, showing that it can help make sense of various features of ML model evaluation practice, as well as suggest procedures for performing fair inter-model comparisons. (shrink)

It is commonly accepted that clinicians are ethically obligated to disclose their use of medical machine learning systems to patients, and that failure to do so would amount to a moral fault for which clinicians ought to be held accountable. Call this ‘the disclosure thesis.’ Four main arguments have been, or could be, given to support the disclosure thesis in the ethics literature: the risk-based argument, the rights-based argument, the materiality argument and the autonomy argument. In this article, I argue (...) that each of these four arguments are unconvincing, and therefore, that the disclosure thesis ought to be rejected. I suggest that mandating disclosure may also even risk harming patients by providing stakeholders with a way to avoid accountability for harm that results from improper applications or uses of these systems. (shrink)

In the natural and social sciences, it is common to use toy models—extremely simple and highly idealized representations—to understand complex phenomena. Some of the simple surrogate models used to understand opaque machine learning (ML) models, such as rule lists and sparse decision trees, bear some resemblance to scientific toy models. They allow non-experts to understand how an opaque ML model works globally via a much simpler model that highlights the most relevant features of the input space and their effect on (...) the output. The obvious difference is that the common target of a toy and a full-scale model in the sciences is some phenomenon in the world, while the target of a surrogate model is another model. This essential difference makes toy surrogate models (TSMs) a new object of study for theories of understanding, one that is not easily accommodated under current analyses. This paper provides an account of what it means to understand an opaque ML model globally with the aid of such simple models. (shrink)

For synergistic interactions between humans and artificial intelligence (AI) systems, AI outputs often need to be explainable to people. Explainable AI (XAI) systems are commonly tested in human user studies. However, whether XAI researchers consider potential cultural differences in human explanatory needs remains unexplored. We highlight psychological research that found significant differences in human explanations between many people from Western, commonly individualist countries and people from non-Western, often collectivist countries. We argue that XAI research currently overlooks these variations and that (...) many popular XAI designs implicitly and problematically assume that Western explanatory needs are shared cross-culturally. Additionally, we systematically reviewed over 200 XAI user studies and found that most studies did not consider relevant cultural variations, sampled only Western populations, but drew conclusions about human-XAI interactions more generally. We also analyzed over 30 literature reviews of XAI studies. Most reviews did not mention cultural differences in explanatory needs or flag overly broad cross-cultural extrapolations of XAI user study results. Combined, our analyses provide evidence of a cultural bias toward Western populations in XAI research, highlighting an important knowledge gap regarding how culturally diverse users may respond to widely used XAI systems that future work can and should address. (shrink)

We argue that the trend toward providing users with feasible and actionable explanations of AI decisions—known as recourse explanations—comes with ethical downsides. Specifically, we argue that recourse explanations face several conceptual pitfalls and can lead to problematic explanation hacking, which undermines their ethical status. As an alternative, we advocate that explanations of AI decisions should aim at understanding.

Explainable AI (xAI) methods are important for establishing trust in using black-box models. However, recent criticism has mounted against current xAI methods that they disagree, are necessarily false, and can be manipulated, which has started to undermine the deployment of black-box models. Rudin (2019) goes so far as to say that we should stop using black-box models altogether in high-stakes cases because xAI explanations ‘must be wrong’. However, strict fidelity to the truth is historically not a desideratum in science. Idealizations (...) – the intentional distortions introduced to scientific theories and models – are commonplace in the natural sciences and are seen as a successful scientific tool. Thus, it is not falsehood qua falsehood that is the issue. In this paper, I outline the need for xAI research to engage in idealization evaluation. Drawing on the use of idealizations in the natural sciences and philosophy of science, I introduce a novel framework for evaluating whether xAI methods engage in successful idealizations or deceptive explanations (SIDEs). SIDEs evaluates whether the limitations of xAI methods, and the distortions that they introduce, can be part of a successful idealization or are indeed deceptive distortions as critics suggest. I discuss the role that existing research can play in idealization evaluation and where innovation is necessary. Through a qualitative analysis we find that leading feature importance methods and counterfactual explanations are subject to idealization failure and suggest remedies for ameliorating idealization failure. (shrink)

This paper delves into the use of machine learning-based systems in decision-making processes and its implications for moral responsibility as traditionally defined. It focuses on the emergence of responsibility gaps and examines proposed strategies to address them. The paper aims to provide an introductory and comprehensive overview of the ongoing debate surrounding moral responsibility in automated decision-making. By thoroughly examining these issues, we seek to contribute to a deeper understanding of the implications of AI integration in society.

Artificial intelligence (AI) systems have demonstrated impressive performance across a variety of clinical tasks. However, notoriously, sometimes these systems are 'black boxes'. The initial response in the literature was a demand for 'explainable AI'. However, recently, several authors have suggested that making AI more explainable or 'interpretable' is likely to be at the cost of the accuracy of these systems and that prioritising interpretability in medical AI may constitute a 'lethal prejudice'. In this paper, we defend the value of interpretability (...) in the context of the use of AI in medicine. Clinicians may prefer interpretable systems over more accurate black boxes, which in turn is sufficient to give designers of AI reason to prefer more interpretable systems in order to ensure that AI is adopted and its benefits realised. Moreover, clinicians may be justified in this preference. Achieving the downstream benefits from AI is critically dependent on how the outputs of these systems are interpreted by physicians and patients. A preference for the use of highly accurate black box AI systems, over less accurate but more interpretable systems , may itself constitute a form of lethal prejudice that may diminish the benefits of AI to - and perhaps even harm - patients. (shrink)

Algorithmic risk assessment tools, such as COMPAS, are increasingly used in criminal justice systems to predict the risk of defendants to reoffend in the future. This paper argues that these tools may not only predict recidivism, but may themselves causally induce recidivism through self-fulfilling predictions. We argue that such “performative” effects can yield severe harms both to individuals and to society at large, which raise epistemic-ethical responsibilities on the part of developers and users of risk assessment tools. To meet these (...) responsibilities, we present a novel desideratum on algorithmic tools, called explainability-in-context, which requires clarifying how these tools causally interact with the social, technological, and institutional environments they are embedded in. Risk assessment practices are thus subject to high epistemic standards, which haven't been sufficiently appreciated to date. Explainability-in-context, we contend, is a crucial goal to pursue in addressing the ethical challenges surrounding risk assessment tools. (shrink)

Inchul Yum (2024) argues that the widespread adoption of language agent architectures would likely increase the risk posed by AI by simplifying the process of aligning artificial systems with human values and thereby making it easier for malicious actors to use them to cause a variety of harms. Yum takes this to be an example of a broader phenomenon: progress on the alignment problem is likely to be net safety-negative because it makes artificial systems easier for malicious actors to control. (...) I offer some reasons for skepticism about this surprising and pessimistic conclusion. (shrink)

In recent years there has been an explosion of interest in Artificial Intelligence (AI) both in health care and academic philosophy. This has been due mainly to the rise of effective machine learning and deep learning algorithms, together with increases in data collection and processing power, which have made rapid progress in many areas. However, use of this technology has brought with it philosophical issues and practical problems, in particular, epistemic and ethical. In this paper the authors, with backgrounds in (...) philosophy, maternity care practice and clinical research, draw upon and extend a recent framework for shared decision-making (SDM) that identified a duty of care to the client's knowledge as a necessary condition for SDM. This duty entails the responsibility to acknowledge and overcome epistemic defeaters. This framework is applied to the use of AI in maternity care, in particular, the use of machine learning and deep learning technology to attempt to enhance electronic fetal monitoring (EFM). In doing so, various sub-kinds of epistemic defeater, namely, transparent, opaque, underdetermined, and inherited defeaters are taxonomized and discussed. The authors argue that, although effective current or future AI-enhanced EFM may impose an epistemic obligation on the part of clinicians to rely on such systems' predictions or diagnoses as input to SDM, such obligations may be overridden by inherited defeaters, caused by a form of algorithmic bias. The existence of inherited defeaters implies that the duty of care to the client's knowledge extends to any situation in which a clinician (or anyone else) is involved in producing training data for a system that will be used in SDM. Any future AI must be capable of assessing women individually, taking into account a wide range of factors including women's preferences, to provide a holistic range of evidence for clinical decision-making. (shrink)

Purpose: The method of diagnosing diabetic retinopathy (DR) through artificial intelligence (AI)-based systems has been commercially available since 2018. This introduces new ethical challenges with regard to obtaining informed consent from patients. The purpose of this work is to develop a checklist of items to be disclosed when diagnosing DR with AI systems in a primary care setting. -/- Methods: Two systematic literature searches were conducted in PubMed and Web of Science databases: a narrow search focusing on DR and a (...) broad search on general issues of AI-based diagnosis. An ethics content analysis was conducted inductively to extract two features of included publications: (1) novel information content for AI-aided diagnosis and (2) the ethical justification for its disclosure. -/- Results: The narrow search yielded n = 537 records of which n = 4 met the inclusion criteria. The information process was scarcely addressed for primary care setting. The broad search yielded n = 60 records of which n = 11 were included. In total, eight novel elements were identified to be included in the information process for ethical reasons, all of which stem from the technical specifics of medical AI. -/- Conclusions: Implications for the general practitioner are two-fold: First, doctors need to be better informed about the ethical implications of novel technologies and must understand them to properly inform patients. Second, patient's overconfidence or fears can be countered by communicating the risks, limitations, and potential benefits of diagnostic AI systems. If patients accept and are aware of the limitations of AI-aided diagnosis, they increase their chances of being diagnosed and treated in time. (shrink)