When a medical device responds to the brain, it can be tempting to picture a simple division of labour: the patient acts, the machine assists, and responsibility remains with the person. A new peer-reviewed paper in Neuroethics argues that this picture may become inadequate as brain stimulation systems move from fixed rules toward genuinely adaptive, learning-based control.
Shiv Prashant Desai, affiliated with computer science and neuroscience at the University of Texas at Austin, examines adaptive deep brain stimulation, or aDBS, through the lens of moral responsibility. The paper does not report a clinical trial or estimate how often responsibility disputes occur. Instead, it develops a philosophical argument about a particular future-facing architecture: a device whose stimulation policy is externally designed, continuously updated by the patient’s neural activity, and applied back to the same neural tissue that is helping to update it.
Desai argues that this feedback structure creates more than ordinary causal complexity. In sufficiently adaptive systems, the patient’s contribution and the device’s contribution to an action may no longer have identities that can be fixed independently of one another. If that is right, asking whether an action belongs entirely to the patient or to the device may start from the wrong conceptual assumption.
The ethical problem begins with the control loop
Deep brain stimulation is already used clinically for conditions including Parkinson’s disease, essential tremor and dystonia, with expanding research in psychiatric disorders. Conventional systems deliver stimulation according to parameters configured by clinicians. Between programming sessions, the policy is effectively fixed.
Responsive systems add feedback. They sense neural activity and adjust stimulation when predefined conditions are met. The important point for Desai is that many current systems remain rule-based. Neural activity can determine what stimulation is delivered next, but it does not continuously rewrite the underlying policy that maps neural signals to stimulation.
The paper distinguishes a third architecture. In a genuinely adaptive system, the mapping itself can change as the device accumulates experience with the patient. Reinforcement learning, model-predictive control and related techniques could allow the policy at a later time to differ from the one operating earlier because the patient’s intervening neural responses have helped update it.
This distinction matters because stimulation also changes neural activity. The device therefore reads from a biological system that it is simultaneously influencing. The patient’s neural state helps shape the policy, while the policy helps shape the patient’s subsequent neural state. Over time, neither side is simply a fixed input into the other.
Three features make the adaptive case different
Desai identifies three features that together generate the philosophical difficulty. First, the system is authored. Engineers, clinical researchers, manufacturers and regulators make deliberate choices about algorithms, biomarkers, reward functions and deployment conditions. The policy is not a naturally occurring biological process.
Second, the policy is continuously modified by patient activity. The designers establish the conditions under which learning occurs, but they do not directly write every later policy state that emerges for a particular patient.
Third, the device intervenes on the same substrate that supplies the signals used for adaptation. Stimulation changes neural activity, and altered neural activity can then contribute to further policy updates. The reading and writing operations are therefore tightly coupled.
This is why the argument is not simply that brains are complicated. Biological systems already contain countless feedback loops. Hormonal systems, neural circuits and metabolism all involve components whose states influence one another. Desai’s claim is narrower: an externally authored artefact with an adaptive policy can enter that loop in a way that gives identifiable human institutions a role in shaping the conditions under which agency itself unfolds.
The paper targets a middle form of separability
The central concept is what Desai calls the separability assumption. Importantly, the paper distinguishes three possible meanings rather than treating separability as all or nothing.
A weak version requires only that the normative roles of the parties remain distinguishable. We can still ask what a patient, clinician, manufacturer or regulator is responsible for in their respective role. Desai accepts that this form survives.
A very strong version would require responsibility to be divided into numerical shares, as if an action could be assigned 40% to one contributor and 60% to another. The paper does not claim that existing theories need such quantitative decomposition.
The target lies between them. The intermediate assumption is that the patient’s contribution and the device’s contribution to a particular action each have an identity that can be specified independently. Desai argues that genuinely adaptive stimulation can undermine this intermediate form because the state of each side has partly developed through interaction with the other.
Three familiar approaches run into the same structural problem
The article tests this diagnosis against three broad approaches to responsibility. Individualist accounts focus on whether an action flows from the person’s reasons-responsive psychology and values. Historicist accounts ask whether post-intervention agency has the right continuity with the person’s earlier self. Relational accounts examine agency through ethically significant relationships between the person, technology and other actors.
Desai argues that each can struggle when the relevant adaptive policy is continuously changing through neural feedback. The problem is not simply a shortage of information about what the device did. It is that the frameworks may assume that the two contributions are already distinct objects waiting to be evaluated.
The paper uses a hypothetical case to make the stakes concrete. A patient with adaptive stimulation develops problematic gambling behaviour after treatment. The behaviour later subsides after the stimulation policy is adjusted. If a court asks the clinical team whether the patient was responsible for the bets, Desai argues that a binary answer may conceal the structure that produced the action.
The case is illustrative, not evidence that adaptive stimulation commonly produces gambling or that a specific commercial system caused such behaviour. Its function is philosophical: to show how familiar responsibility questions become unstable when the system affecting behaviour is itself being altered by the neural activity it affects.
Responsibility becomes distributed across four loci
Rather than treating distributed responsibility as a fallback after individual attribution fails, Desai proposes a spectrum of joint agency. At one end are systems whose contributors remain independently identifiable even when their interaction is complicated. At the other are systems in which the contributors’ states at the moment of action have been partly constituted through their joint operation.
The paper places learned-policy aDBS far enough toward the structurally entangled end that the practical default should change. Instead of beginning with individual localization and demanding reasons to distribute responsibility, evaluators should begin with distribution and demand affirmative reasons before localizing responsibility in one party.
Desai identifies four loci through what he calls constitutive authorship. The patient bears first-personal responsibility associated with living and regulating the life being shaped. The clinical team bears therapeutic responsibility for selection, configuration and monitoring. The manufacturer bears design responsibility for the foreseeable behaviour of the system it authored. The regulator bears systemic responsibility for the conditions under which a class of devices is approved, deployed and surveilled.
These are not interchangeable pieces of one responsibility total. They concern different objects, operate over different timescales and are discharged through different practices. The framework therefore resists the idea that distributed responsibility necessarily means diluted responsibility.
The patient does not disappear as a moral subject
A major objection to distributed agency is that it might erase the patient. If the action is attributed to a patient-device system, does the person lose moral standing or become merely one component in a machine?
Desai answers by separating the execution of an action at a particular moment from responsibility exercised across time. Even if a specific action cannot be cleanly assigned to a single source, the patient can retain responsibility for how they regulate their situation over time, including whether they report warning signs, seek adjustments or engage reflectively with changes in behaviour.
This distinction also matters for punishment. The paper suggests that where an externally authored adaptive policy genuinely contributes to an action, the patient may be a poor candidate for the full retributive burden that would attach to an ordinary, undistributed act. But attenuation is not the same as exculpation. Responsibility for longer-term self-regulation can remain, while design, clinical and regulatory responsibilities may become more salient.
Why the argument is deliberately future-facing
The scope limitation is crucial. Desai does not claim that every person currently receiving deep brain stimulation is part of a jointly learning patient-device agent. The argument specifically concerns learned, externally authored and continuously updating stimulation policies.
Many commercial systems currently sit elsewhere on the spectrum. Rule-based responsive devices can sense biomarkers and change stimulation while still operating under a mapping configured in advance. Conventional open-loop DBS falls further away from the paper’s target. The philosophical argument therefore applies most strongly to systems that are still an active research frontier rather than standard clinical care.
That limitation is also part of the paper’s purpose. Ethical frameworks are often forced to catch up after a technology is already widespread. By examining responsibility before self-tuning neural systems become routine, the study asks whether law, consent processes, clinical governance and device regulation need concepts designed for adaptive relationships rather than fixed tools.
What this paper can and cannot establish
This is a philosophical research article, not an empirical effectiveness study. There is no participant sample, randomized intervention, effect size or prevalence estimate. The evidence is conceptual and interdisciplinary: the paper distinguishes technical architectures, evaluates existing theories of agency and responsibility, develops a structural argument, and tests that argument against objections and difficult cases.
That means the article cannot show that adaptive stimulation actually changes responsibility judgments in courts, clinics or the public. It also cannot establish how often future devices will produce behavioural changes relevant to responsibility. Those questions require empirical legal, clinical and social research.
The four-locus framework is likewise a normative proposal, not a measured fact. Other philosophers may disagree about where the proposed joint-agency spectrum should flip from localization to distribution, whether constitutive authorship selects the right parties, or whether existing theories can accommodate adaptive coupling without being replaced.
The paper itself acknowledges a particularly important open problem around criminal desert. Its primary contribution concerns attributability and answerability. A complete account of punishment when action is synchronically distributed would require further philosophical and legal development.
A responsibility problem that grows with the technology
The significance of the argument lies in how it reframes neurotechnology. A fixed device can often be treated as something a person uses or undergoes. A learning device that continuously changes through the activity of the brain it stimulates is harder to place outside the agent in the same clean way.
Desai’s proposal is not that technology eliminates personal responsibility. It is that sufficiently adaptive neurotechnology may alter the structure through which responsibility should be assigned. Patient, clinician, manufacturer and regulator can remain morally distinguishable even when the action produced by the patient-device coupling cannot be cleanly decomposed into independently specified contributions.
For a field moving toward more adaptive and personalised neural control, that distinction could become practically important. The closer devices come to learning from patients while simultaneously changing the neural activity from which they learn, the less adequate a simple question of whether the person or the machine caused an action may become.
Source Information
Study: Beyond Separability: Closed-Loop Brain Stimulation and the Distribution of Moral Responsibility
Author: Shiv Prashant Desai
Journal: Neuroethics
Published: 28 September 2026
Volume and article: 19, Article 48
DOI: 10.1007/s12152-026-09673-1
