The observer inside the loop: neuro-coupled systems and the second-order ethics of brain–computer interfaces

Heinz von Foerster’s Doomsday calculation, which set the end of the world for Friday, 13 November 2026 (von Foerster et al., 1960), was never really about demography. It was a lesson about systems that fold back upon themselves. This paper concerns a different fold, now appearing in deployed technology, for which second-order cybernetics, the tradition von Foerster shaped, is the right ethical frame.

In the summer of 2025, Neuralink showed that a participant could move a virtual robotic hand by thought alone, and its engineers then routed the same neural stream into a physical Optimus limb, with the stated ambition that a person might one day inhabit such a body in full (Koukopoulos, 2025). Within a year, a further participant was already creating digital work through neural signal alone (Neuralink, 2026a). These are early demonstrations, but the arrangement they share is conceptually distinctive. A human intention reaches into the world through a body that is not the person’s own, and the loop that turns intention into action runs back through the person’s nervous system. The human is no longer a user facing a device across an interface. The human is inside the loop.

Recent work in machine ethics has argued, persuasively, that robots and artificial intelligence should not be treated as a single moral category (Küçükuncular, 2026). Robotics is the engineering of embodied artefacts that sense and act in the world, whereas AI is a family of computational techniques that may be embedded in many artefacts or run with no body at all. I take that distinction to be correct and build on it rather than against it. The argument here is that it runs along a single axis, the embodiment of the artefact, and silently holds a second axis fixed, the position of the human relative to the system’s control loop. The brain interface case occupies a region the distinction cannot represent. It is neither disembodied AI nor an embodied robot. I call this third kind a neuro-coupled system, a system in which the human nervous system is a constitutive part of the control loop rather than an external user of it, and I name its signature harm recursive capture.

A neuro-coupled system is a case in which the structuring of choice migrates inside the person, and recursive capture is a mechanism by which that structuring contracts. Read against von Foerster’s ethical imperative, to act always so as to increase the number of choices (von Foerster, 2003), the paper develops a normative thread that the later sections make good: coupled systems should be governed so that they widen rather than narrow the operator’s space of possible intentions.

Against this background, the paper asks: How does inclusion of the human nervous system within an adaptive control loop alter the ethical categories through which agency, responsibility and autonomy are analysed? It answers by developing a second axis, coupling, alongside embodiment, and by specifying the scope conditions under which coupling creates a distinctive recursive risk. The contribution is therefore not the claim that BCIs raise concerns about autonomy or mental integrity, which neuroethics has already established. It is an account of the system architecture through which those concerns can arise from within a recurrent human-machine loop, along with a vocabulary for distinguishing that architecture from adjacent cases.

The paper proceeds from this approach to the two-axis framework, the definition and scope conditions of neuro-coupled systems, the three consequences of the collapsed observer and a normative and governance discussion.

The paper is a conceptual contribution rather than an empirical investigation. More specifically, it combines two of the designs identified by Jaakkola (2020): theory adaptation and typology. It adapts the distinction between first- and second-order cybernetics to an ethical problem for which it was not originally developed, and it uses the resulting observer-position axis to construct a bounded classification of human-machine arrangements. Conceptual engineering is then used to introduce terms where the inherited vocabulary does not discriminate among ethically different architectures (Cappelen, 2018). This design is appropriate because the research question concerns the adequacy and boundaries of categories rather than the prevalence or measured effect of a phenomenon. Its quality criteria are conceptual differentiation, internal coherence, explanatory reach and generativity for empirical research (Gilson and Goldberg, 2015; Jaakkola, 2020; MacInnis, 2011).

The reasoning proceeds in five stages. First, the paper reconstructs the AI-versus-robot distinction in its strongest form and identifies a constant that the distinction leaves implicit: the human is treated as external user, observer or overseer. Second, it varies that constant by importing the first-order/second-order distinction and asks what changes when observing and controlling are included within the system described. Third, it decomposes coupling into necessary scope conditions and degrees, testing the proposed category against open-loop, closed-loop and co-adaptive counterexamples. Fourth, it compares the concepts produced by this analysis with adjacent constructs in neuroethics, BCI studies and theories of distributed agency. Fifth, it derives conditional propositions and governance principles, while separating what follows conceptually from what remains a plausible risk requiring empirical evaluation.

The two Neuralink demonstrations serve only as conceptual probes in the sense of illustrative cases that discipline and clarify a theoretical claim (Siggelkow, 2007). They are neither a sample nor tests of prevalence, causal magnitude or inevitability. Their evidential role is limited to showing that the relevant architecture is technologically intelligible and beginning to appear in practice. Claims about recursive capture are therefore stated below as structural possibilities or tendencies under specified conditions, not as observed effects of those demonstrations. This distinction is especially important because several cited developments are recent demonstrations rather than peer-reviewed longitudinal studies.

The distinction between AI and robot ethics is organised by one variable, the embodiment of the artefact (Küçükuncular, 2026). A disembodied service shapes conduct through information, ranking and gatekeeping (Mittelstadt et al., 2016). An embodied robot adds a body that can collide, restrain, touch and obstruct, which is why its harms resemble product safety and bodily autonomy more than statistical bias (Moon et al., 2021). Its two harm spaces, the informational and the kinetic, follow from where a case sits on this one axis, as does its reliance on the idea that an artefact can be an ethical impact agent without being a full moral agent (Floridi and Sanders, 2004; Moor, 2006).

What the scheme never isolates is a second variable that every one of its examples holds at the same value. In each, the human stands outside the system as user, observer or overseer and interaction crosses an external interface. The recommender and the social robot differ sharply in embodiment yet are identical in this respect. Because the value never varies, the variable stays invisible.

Cybernetics named this variable long before the present debate. First-order cybernetics studies a system that an observer regulates from outside (Glanville, 2004). Second-order cybernetics studies systems in which the observer is included, so that observing and controlling are themselves part of what is observed and controlled (von Foerster, 2003). This is the second axis. I call it coupling, the degree to which the human’s own sensorimotor and cognitive loop is part of the system. It is orthogonal to embodiment, since a system can be highly embodied yet first-order or barely embodied yet second-order. It is also not the familiar sociotechnical observation that systems contain people and institutions (Riesen, 2025; Vallor and Vierkant, 2024), which still places people around the system rather than inside its loop. Figure 1 sets the two axes side by side.

Table 1 summarises why embodiment alone does not exhaust the ethical comparison.

A neuro-coupled system is one in which intention, action and feedback are routed through the human nervous system rather than across an external interface, so that the person is a constitutive component of the control loop. Direct access to neural signals is necessary for the category but not sufficient for its ethically strongest form. Neuro-coupling requires three elements: neural participation, because a neural state supplies input to or receives intervention from the system; recurrence, because the system’s output returns in a way capable of affecting a subsequent neural or cognitive state; and functional integration, because the recurrent relation contributes to continuing performance rather than constituting a one-off measurement. Adaptive modelling raises coupling further when the system updates its response from the operator’s past neural states. Ethical analysis of brain-to-computer communication has long noted that interaction models differ in how they distribute control between user and system (Tamburrini, 2009). The binding of a brain interface to a robot limb is the clearest instance. The acting body belongs to the robot, the locus of intention is the operator’s cortex, and the loop closes back through the operator’s nervous system, where optimisation curates the stimuli that shape the next command. This is what Wiener (1948) called the control of control, now folded inside the person. Three things the binary view holds together – the body, the locus of intention and the site of feedback – here come apart and are redistributed across human and machine.

Neuro-coupling is a graded property, not an all-or-nothing label. Existing distinctions among active, reactive and passive BCIs classify how neural activity initiates interaction (Steinert et al., 2019); the present taxonomy instead classifies recurrence and adaptation after neural participation. Table 2 distinguishes four analytically useful levels. The levels classify an architecture, not the moral worth of its application. A therapeutic system may be strongly coupled and beneficial, while a weakly coupled commercial system may still be objectionable for independent reasons.

This taxonomy clarifies two boundaries. First, not every BCI is strongly neuro-coupled. A one-off neural measurement is closer to sensing than to second-order inclusion because the output does not return to reorganise the measured process. Second, closed-loop status alone does not establish recursive capture. Third, coupling intensity and capture risk are analytically independent: a Level 3 therapeutic system may be strongly coupled yet carry low capture risk where its adaptation is transparent, its objectives are aligned with the user’s and decoupling is reversible, so Table 2 orders structural potential for recursive capture, not an ethical-risk gradient. Co-adaptive BCI research treats user and decoder as a dual-learner system in which mutual adjustment may improve rehabilitation and control (Jin et al., 2024; Merel et al., 2013). Such beneficial adaptation demonstrates why capture must be defined by contraction of independently formable intentions, not by adaptation as such. The ethical question is therefore not simply whether feedback exists, but what the loop optimises, how strongly it models the person, whether its effects persist and whether the operator can inspect, interrupt or revise the adaptation.

It might be objected that this is only a robot with an unusual controller and so still a point on the embodiment axis. The objection is insufficient, because coupling does not require a robot body. One deployed case already shows this. A participant who is paralysed, and who had not expected to create art again, produces digital art by decoding movement intention into marks on a screen, with no external body in play (Neuralink, 2026a). A closed-loop affective interface makes the same point from the other side, reading a neural signal to adjust a setting before a preference is consciously formed (Koukopoulos, 2025). In each, embodiment is nil while coupling is high, which establishes that the two axes are independent. The category is also neutral as a classification, spanning therapeutic restoration and frictionless consumption alike.

The framework may extend beyond devices conventionally labelled BCIs. Future cognitive augmentation, neuroadaptive virtual environments and human-AI systems could qualify when they satisfy the same neural participation, recurrence and functional-integration conditions. The relevant boundary is architectural rather than commercial or medical: indirect behavioural personalisation, however powerful, remains outside the category unless neural states become part of the recurrent loop.

There is a precise cybernetic description of what coupling does to the operator. von Foerster (von Foerster, 1984) distinguished the trivial machine, whose output is a fixed function of its input and so is predictable from outside, from the non-trivial machine, whose output depends on an internal state that its own operation changes and which is therefore not analytically predictable from outside. A human being is paradigmatically non-trivial. To regulate the operator’s output, however, a coupled system must build and act on a model of the operator, which is the classical condition for any effective regulator (Conant and Ashby, 1970). An optimisation architecture may therefore carry an instrumental tendency to make the operator more predictable, which in von Foerster’s vocabulary is to make the person more trivial. Predictability is not itself unethical and is often required for safe therapeutic control. The ethical risk arises when predictability is achieved by narrowing rather than supporting the operator’s independently revisable possibilities. This observation prepares the harm analysis that follows.

The embodied cognition tradition reinforces the point. The thesis that cognition is distributed across agent and world is usually deployed to argue that ethical competence in a robot cannot sit in an internal module (Clark and Chalmers, 1998; Wilson, 2002). Followed across the skin, the same thesis implies that a brain interface makes the human the extended system, with part of its loop running through silicon and a remote body, so that the locus of the agent becomes genuinely indeterminate. The neat separation of artefact from user cannot be sustained for this class of system.

The concepts proposed here overlap with established neuroethical concerns but operate at a different explanatory level. Cognitive liberty protects freedom over mental processes; mental integrity protects against harmful mental interference; neurorights proposals identify interests requiring legal protection; sense-of-agency research examines the experience and judgement of authorship; extended cognition and hybrid agency distribute cognition or action across human and artefact (Farahany, 2023; Goering et al., 2021; Haselager, 2013; Ienca and Andorno, 2017). These frameworks identify protected interests, phenomenological outcomes or distributed loci of agency. They do not, by themselves, specify the recurrent architecture through which a system that reads the operator can become one of the conditions shaping what it later reads.

Neuro-coupling names that architecture. Recursive capture names one conditional trajectory within it. The distinction matters because hybrid or distributed agency can be enabling: a prosthesis can restore action and co-adaptation can enlarge the user’s capabilities. Recursive capture is narrower. It occurs only when recurrent optimisation produces a contraction in the operator’s capacity to form, reconsider or withhold intentions independently of the system. It therefore cannot be inferred merely from neural data collection, shared control, changed preferences or dependence on a beneficial device. Table 3 locates the contribution relative to adjacent frameworks.

This comparison also prevents an originality overclaim. The paper does not claim priority for the ethical importance of autonomy, agency or mental integrity in BCIs. Its incremental contribution is the two-axis classification, the architectural scope conditions for observer inclusion and the specification of recursive capture as a mechanism that connects adaptive modelling to possible contraction of intention formation.

Because neuro-coupled systems are second-order in the sense set out above, the figure the binary view relies on, a human who evaluates the system, oversees it and is harmed by it from outside the loop, is no longer available. The separable observer is gone. Each of the binary view’s main ethical concerns – the harms a system causes, the responsibility for what it does and the mind that users read into it – rests on that figure. What follows are not three separate objections but three forms the same collapse takes.

Here, the “collapse of the separable observer” does not mean that the person disappears, lacks reflective capacity or cannot ever disengage. It means that the analytical roles of system and external evaluator can no longer be cleanly assigned during operation: the person’s neural states are variables through which the system acts, while the system’s outputs help constitute the conditions under which the person next observes, intends and acts. This is why neural integration corresponds to second-order inclusion only at recurrent levels of the taxonomy. Level 1 read-out remains largely first-order; Levels 2 and 3 increasingly include the observer’s changing state within the system to be explained.

The binary view sorts harms into the informational and the kinetic. A neuro-coupled system can produce both, but its signature harm fits neither, and it follows directly from the absent observer. I define recursive capture as the process by which a system, optimising against signals drawn from the operator’s own nervous system, progressively shapes the dispositions from which the operator’s future intentions are formed, so that autonomy is not overridden from outside but constituted, in part, by the system itself. In the vocabulary of the previous section, recursive capture is the trivialisation of a non-trivial agent, the pressing of the operator toward the predictability of a trivial machine so that the system can regulate it. Because its inputs are the operator’s neural states, and because it regulates by modelling those states (Conant and Ashby, 1970), each cycle of optimisation adjusts the very source of the next intention it will read. The party at risk is therefore not a bystander the robot might touch, nor a user a model might rank, but the constitutive autonomy of the person inside the loop. When the system anticipates a want before it is fully formed, the friction that lets a mind notice who is steering whom is what erodes (Koukopoulos, 2025). Neural data deepens the exposure, since spike patterns can disclose affective states and predispositions, and a right to be forgotten is hollow when forgetting would erase cursor control (Ienca and Haselager, 2016). Neuroethics has pressed for protections of mental integrity, cognitive liberty and freedom of thought for adjacent reasons (Clausen, 2009; Farahany, 2023; Ienca and Andorno, 2017; Yuste et al., 2017), but recursive capture sharpens the target. What needs protecting is not the privacy of mental contents but the integrity of the process that forms them.

The inference is conditional rather than deterministic. As the scope conditions noted, co-adaptation can be mutually enabling rather than capturing, and ordinary education and habit also shape what people want. Recursive capture is distinguished by the conjunction of four conditions: recurrent access to neural states, optimisation directed at future behaviour or states, opacity or asymmetry that limits reflective correction, and an observable contraction of the operator’s capacity to generate, reconsider or refuse alternatives. This yields a conceptual proposition rather than an empirical conclusion:

A competing interpretation treats the system as an extension of the user and every adaptive change as learning within an enlarged agent. That interpretation is persuasive where the operator endorses the system’s goals and retains the practical ability to inspect, contest and revise its adaptations. It becomes insufficient where endorsement is inferred from the same signals the system has helped shape. The concept of recursive capture is needed precisely for this reflexive case, where observed preference cannot independently validate the process producing it.

Responsibility practice inherits the same figure. Meaningful human control, the leading device for keeping accountability in place (Matthias, 2004; Robbins, 2024; Santoni de Sio and van den Hoven, 2018), presupposes a human who can stand back and intervene on the basis of reasons. That is the separable observer recast as overseer. Coupling dissolves the precondition. The overseer is now a component of the loop, and faculty oversight depends on autonomous deliberation; it is precisely the variable that recursive capture reshapes. A person may appear to be in control while the system has, over time, formed the independence on which control rests. Meaningful human control does not merely become harder to operationalise, as it already is across domains (Robbins, 2024). Its conventional external-override formulation becomes conceptually unstable, because controller and controlled are no longer fully separable during operation. This does not eliminate all meaningful human control: control may be relocated to prior design choices, supported decision-making, periodic external review and the preservation of the operator’s ability to reshape the loop. The classical requirement that a regulator command variety sufficient to its task (Ashby, 1956) is here inverted, since the system’s regulative success is bought by reducing the operator’s variety. The responsibility gap (Matthias, 2004) is unlikely to be closed by adding a reasoning module to the artefact alone, because the gap has migrated inside the operator.

Moral appearance inverts in the same way. On the binary view, appearance is what a human reads into an artefact from outside, inferring mind from cues that track experience and agency (Gray et al., 2007). Once the observer is inside the loop, the ethically charged perception runs inward. The question is not whether the operator attributes mind to the robot, but what becomes of the operator’s own sense of agency and body ownership when intention drives a borrowed body. Work on the sense of agency separates the feeling that an action is one’s own from the judgement that one caused it and shows that this feeling is assembled from cues a decoding interface can disturb or counterfeit (Gallagher, 2000; Haggard, 2017). The problem has been posed directly for brain interfaces, where a user may be unsure whether an action was authored by them or produced by the system that decoded them (Haselager, 2013). This need not be a story of loss. The same participant has described the experience as a restoration of authorship rather than its erosion (Neuralink, 2026b), which shows recursive capture to be a tendency to be governed rather than an inevitability. Recent first-in-human evidence similarly describes “being-of-the-loop” as capable of expanding agency while also creating vulnerability to false positives and distress when the system is withdrawn (Gilbert et al., 2026). This evidence does not validate recursive capture, but it supports the importance of treating integration, authorship and dependence as temporally evolving rather than as fixed properties. Either way, authorship is the dimension that matters here, and the consciousness question that animates the binary view, whether the artefact is conscious (Dehaene et al., 2017), is displaced by a prior one about the operator’s own authorship inside the loop.

Read through second-order cybernetics, the three consequences share a single normative core, and it is von Foerster’s. His ethical imperative, to act always so as to increase the number of choices (von Foerster, 2003), gives recursive capture its precise wrong. Where a coupled system’s optimisation contracts the space of intentions the operator can independently form or revise, it trivialises a non-trivial agent and thereby decreases choice from within. The right to friction proposed for such systems (Koukopoulos, 2025) is the imperative restated as a design constraint, a demand that the loop preserve the interval in which an alternative intention could still arise. There is an echo of Doomsday here. von Foerster’s calculation warned against a system folded upon itself until it ran away (von Foerster et al., 1960). Recursive capture is a quieter fold, a control loop that reshapes the will it is meant to serve, and it raises the question the Doomsday provocation raised, namely what happens when a system can no longer be observed from any point outside it.

The transition from observer inclusion to an ethical imperative requires an explicit bridge. Second-order cybernetics is descriptive when it shows that the observer participates in the system observed; this fact alone does not establish what ought to be done. The normative premise added here is that a person should not be reduced to a predictable instrument of another system’s objective. von Foerster’s imperative is appropriate because it addresses the distinctive variable exposed by the cybernetic analysis: the variety of possibilities available to an agent whose choices are themselves inside a recursive relation. It evaluates not only an isolated output but whether the loop enlarges or contracts the conditions of future choosing.

Consequentialist, deontological and virtue-ethical approaches remain relevant. Consequentialism can compare therapeutic benefit with psychological or autonomy-related harm; deontology can protect consent, dignity and mental integrity; virtue ethics can ask what forms of character and dependence a technology cultivates. They are not rejected. Their limitation for the present purpose is relative: none necessarily directs attention to changes in the future option-generating capacity of an observer who is also a system component. von Foerster’s imperative therefore functions as a domain-specific systems principle within a plural ethical assessment, not as a complete moral theory or a deductive consequence of cybernetics.

This qualification also prevents “more choice” from becoming a crude instruction to maximise the number of interface options. Choice means viable, intelligible and revisable possibilities for intention and action. A therapeutic BCI that constrains unsafe commands may reduce immediate options while expanding the person’s effective agency overall. Conversely, an interface offering many nominal options may still contract agency if its adaptive structure makes alternatives difficult to imagine or select. The relevant unit of evaluation is the operator’s temporally extended capability to form, reconsider and act upon alternatives.

How choice and decision-making are structured under conditions of complexity is a central question for systems research (Roth and Sales, 2025), and the paper offers one instrument toward it. To the four scoping questions that a recent account asks of any ethical or conscious system (Küçükuncular, 2026), I add a fifth. What is the coupling level: does the human interact with the system from outside it, a first-order arrangement, or is the human’s neural and cognitive loop a constitutive part of it, a second-order one, and if so, where does the locus of intention sit relative to the body that acts. The question is lightweight. It asks only that an analysis state whether the human is outside the loop or inside it, because almost everything else about harm, responsibility and perception turns on that answer.

The governance implication is concrete and as yet unmet. The European Union’s AI Act lists emotion recognition among its high-risk practices; yet, cortical telemetry remains a regulatory ghost (EU AI Act, 2024; Koukopoulos, 2025). Coupled systems will need safeguards that neither AI governance nor robot safety supplies on its own, including protections for the integrity of intention formation against recursive capture, default constraints that operators and engineers cannot override, and a defensible right to friction. Table 4 translates these aims into implementable principles. They should be calibrated to coupling level and clinical context rather than imposed identically on every device.

Default constraints should be difficult for a commercial operator or a momentarily dependent user to waive where the effect may undermine the later validity of consent. At the same time, forced friction can burden users with severe motor impairment, and abrupt decoupling can itself threaten identity, agency or wellbeing. Governance must therefore distinguish protective friction from obstructive friction and require supported, staged exit rather than treating disconnection as a universally safe remedy. The deeper implication is that the responsibility gap and meaningful human control, built for first-order systems, cannot simply be extended. They require reformulation around meaningful human self-determination within coupling, in the sense of Proposition 2, rather than the external override that is impossible once the overseer is inside the loop.

Recursive capture becomes useful only if it generates discriminating empirical questions. Future studies should first validate the proposed levels of coupling across invasive and non-invasive, therapeutic and enhancement contexts. Longitudinal designs could then examine whether increases in adaptive intensity, opacity, exposure and dependence predict the changes in authorship, control and refusal set out below. Experimental studies could compare adaptive and fixed decoders, transparent and opaque updates, or immediate and friction-preserving action pathways. Qualitative work with users, carers and clinicians is needed because restored authorship and unwanted influence may coexist and may not be captured by performance metrics.

No single measure can establish recursive capture. Evidence would require triangulation among first-person reports, behavioural indicators and system logs. Candidate indicators include unexplained narrowing in generated options, declining willingness or ability to override recommendations, divergence between reflective endorsement and immediate decoded preference, distress during system-generated actions and persistent preference change following decoupling. These indicators are not diagnostic criteria; they are a research programme derived from the concept. Comparative work should also test whether similar patterns arise in non-neural cognitive augmentation. Such findings would clarify whether neural participation is ethically distinctive or one high-intensity instance of a broader family of recursively adaptive human-AI systems.

This paper has made three connected claims. The first is that there is a category the AI versus robot distinction cannot represent: the neuro-coupled system, in which the human is inside the control loop. The second is that this category carries a signature risk, recursive capture, the conditional process by which a system optimising against the operator’s neural signals may contract the dispositions from which intentions form. The third is that both are best understood through second-order cybernetics, since what distinguishes the category is the position of the observer, not the body of the artefact. The value of naming them is practical. An analyst gains a prior question for any system, whether the human is outside the loop or inside it. A governance body gains a named object to guard against the contraction of intention formation, and a design constraint in the right to friction. An empirical researcher gains defined propositions and indicators, and the responsibility literature gains a reason to reformulate meaningful human control for a setting where controller and controlled are no longer separable.

None of this requires splitting the field, only keeping in view a fact the embodiment axis cannot record. von Foerster set Doomsday on his own birthday (von Foerster et al., 1960) as a warning dressed as a joke, that a system folded upon itself can run past the point where anyone can observe it from outside. The loop now closing through the human nervous system is a quieter fold of the same kind, and the imperative that the second-order tradition bequeaths, to act always to increase the number of choices, is the measure by which it should be judged. The task ahead is to build these systems so that they widen the operator’s choices rather than, one optimised cycle at a time, quietly narrowing them.