Julian Savulescu

In this paper, we argue for the adoption of algorithmic bioethics , which uses algorithms, in the form of decision trees, to articulate, understand, apply, and evaluate bioethical arguments. Algorithms bring many benefits to the study of bioethics: they clearly, precisely, and transparently articulate the decision procedure and facts relevant to an ethical decision; furthermore, they can be used to systematically discover, map out, and compare different ethical decision procedures concerning a practical issue. We then also outline and tackle potential limitations of algorithmic bioethics, and explore future directions for using decision trees in bioethics. Throughout, we illustrate our points using a case study of ethical decision trees, created with the help of large language models, about conscientious objection in medicine.

We have defended the view that there should be no accommodation of conscientious objectors in health care. Numerous critics have advanced arguments designed to defeat this conclusion. We respond here to what we take to be the most serious of the criticisms directed at our analysis. We begin with two distinct consequentialist lines. The first holds that accommodation produces better net outcomes for the healthcare system, even at some cost to individual patients. The second claims that a properly designed accommodation regime need not harm patients at all. Both arguments fail. The supposed consequentialist benefits are empirically speculative, and the costs to patients are not adequately weighed. The compromise position has been tried in many jurisdictions, and the constraints on which it relies are routinely under‐enforced. The unimpeded access promised does not materialise. We then turn to teleological arguments about the supposed nature or traditions of medicine. These accounts presuppose answers to the very normative questions they purport to resolve. And even if granted, they would justify legislative advocacy, not individual clinical refusals at the bedside. We also address the methodological objection that any context‐independent conclusion about conscientious objection is procedurally illegitimate in a pluralistic liberal democracy. This objection does not succeed. Finally, we respond to two direct challenges to our position. The first holds that professional codes and the law are epistemically unreliable. History shows that they have sanctioned abhorrent practices, and conscientious objection has at times served as a check on those failures. The second holds that restricting accommodation degrades the quality of independent clinical judgment and harms patient care. We show why neither objection succeeds. We argue that the patient's entitlement to legally sanctioned care that is derived from an ethically justifiable process is foundational to the obligations of the medical profession. We offer a welfarist consequentialist defense of this position.

Large language models (LLMs) are becoming routine as tools for drafting, editing, and revising scholarly manuscripts, raising difficult questions about how much transparency is needed to assess research integrity, authorship claims, and accountability. In response, some authors have proposed that researchers who use LLMs in manuscript preparation should be required to submit the prompts they used, together with the generated outputs, as supplemental material. We support detailed disclosure when it is needed to assess the scientific validity or reproducibility of a study, for example where LLMs are used in data analysis, coding, literature synthesis, or the production of figures. We argue, however, against Mandatory Full Inclusion (MFI) of prompts and outputs for ordinary writing assistance. Such a policy would, we suggest, be difficult to enforce, easy for dishonest users to evade, burdensome for careful users to follow, and poorly suited to the capacities of editors and peer reviewers. It would also risk discouraging beneficial uses of LLMs, penalizing careful multi-model consultation, distorting how researchers interact with AI systems, and disadvantaging scholars who already face undue or unnecessary barriers in academic publishing, including scholars at under-resourced institutions. More generally, an MFI policy would impose a level of surveillance on AI-assisted writing that has no clear analog in ordinary human collaboration. We propose instead a proportionate prompt-and-output disclosure framework: targeted disclosure where reproducibility or validity requires it; voluntary documentation where authors wish to demonstrate their contribution; and continued reliance on authorship attestation, contributor statements, and misconduct procedures where warranted. This approach better aligns transparency with the risks posed by LLM use in scholarly writing, while also minimizing surveillance-related harms and inefficiencies.

Adolescents sometimes express firm views about high-stakes medical decisions that conflict with those of their caregivers or clinicians, raising difficult questions about how much weight their wishes should carry. A recurring concern is whether such decisions will later be regretted, although the normative significance of future potential regret remains hotly debated. Recent work in adult bioethics has proposed using artificial intelligence–based Personalized Patient Preference Predictors (P4s) to estimate what current treatment incapacitated adults would have chosen based on their known or inferred preferences when last autonomous. This paper explores whether, and with what ethical significance, a parallel ‘Paediatric Personalized Patient Preference Predictor’ (P5) could be applied to children and young people. We analyse how predicted future preferences might bear on questions of substituted judgment or proxy consent, best-interests assessments, resistance to treatment, and adolescents’ own deliberation. We distinguish different groups of paediatric patients based on current and future decision-making capacity, and identify both where a P5 might have greatest promise and where it is least defensible. We identify challenges for a P5 including that of self-reinforcing or self-negating preferences as well as of deciding which future preferences are relevant to current decisions. We argue that, in paediatrics, predicted future preferences cannot ground substituted judgment in the adult sense, but may nonetheless play a supportive role in best-interests reasoning and shared decision-making, subject to significant epistemic and ethical constraints.

Some advances in reproductive technologies raise substantial ethical and psychological challenges, for example, the use of preimplantation genetic testing to select embryos based on non-medical traits. While studies have explored public willingness to use preimplantation genetic testing for medical or non-medical attributes, stated willingness may not reflect implantation decisions when such information is available. This large cross-national study examined public views on polygenic embryo testing. In a US sample ( N = 1,467), participants were more willing to test for medical conditions (for example, heart disease) than non-medical traits (for example, antisocial behaviour or low intelligence), although over half supported testing for non-medical traits. In a forced-choice implantation task, participants used medical and non-medical information to a similar extent when making decisions. Similar patterns were observed in Chinese participants ( N = 623). These findings suggest variability in choices across specific traits and conditions, rather than a uniform distinction between medical conditions and non-medical traits. The Stage 1 protocol for this Registered Report was accepted in principle on 7 October 2025. The protocol, as accepted by the journal, can be found at https://osf.io/vb9c2 .

The article advances the idea of "digital life models" to assist with difficult personal decisions, including those that can change an individual in profound ways. We propose that if personalized AI simulations can model how an individual's life might unfold under different conditions, this could help the individual better understand the likelihoods and expected subjective value of these possible outcomes. It remains to be seen whether these digital life models are technically feasible, given fundamental challenges in modeling psychological complexity, person-environment interactions, and value change. Ethical concerns include data privacy, cultural bias, and the risk of improperly shaping user decisions. Acknowledging these limitations, we argue that digital life models, if successfully developed, could in principle enhance rational and authentic decision-making in the face of potentially transformative change.

Large language models (LLMs) are becoming routine as tools for drafting, editing, and revising scholarly manuscripts, raising difficult questions about how much transparency is needed to assess research integrity, authorship, and accountability. In response, some authors have proposed that researchers who use LLMs in manuscript preparation should be required to submit the prompts they used, together with the generated outputs, as supplemental material. We support detailed disclosure when it is needed to assess the scientific validity or reproducibility of a study, for example where LLMs are used in data analysis, coding, literature synthesis, or the production of figures. We argue, however, against Mandatory Full Inclusion (MFI) of prompts and outputs for ordinary writing assistance. Such a policy would, we suggest, be difficult to enforce, easy for dishonest users to evade, burdensome for careful users to follow, and poorly matched to the capacities of editors and peer reviewers. It would also risk discouraging beneficial uses of LLMs, penalising careful multi-model consultation, distorting how researchers interact with AI systems, and disadvantaging scholars who already face undue or unnecessary barriers in academic publishing, including scholars at under-resourced institutions. More generally, an MFI policy would impose a level of surveillance on AI-assisted writing that has no clear analogue in ordinary human collaboration. We propose instead a proportionate prompt-and-output disclosure framework: targeted disclosure where reproducibility or validity requires it; voluntary documentation where authors wish to demonstrate their contribution; and continued reliance on authorship attestation, contributor statements, and misconduct procedures where warranted. This approach better aligns transparency with the risks posed by LLM use in scholarly writing, while also minimising surveillance-related harms and inefficiencies.