Main text
1. Dependence Without Understanding
1.1. Dependence as a Structural Condition
Technical dependence is not a reversible individual choice but a contemporary structural condition. Asserting this is essential because a persistent narrative of individual freedom endures: "we can choose to stop using technology; we choose to remain connected." That assertion is theoretically valid yet practically false. To extract computational systems from urban life is to strip away the possibility of navigating transport networks, conducting instant communication, accessing information, performing financial transactions, receiving medical diagnoses, and participating in democratic governance. Refusing GPS is not an exercise in liberty; it is social exclusion. Abandoning the mobile phone is not ascetic purity; it is isolation from the labour market. The choice to step outside dependence is invariably a choice to suffer exclusion, it is not a choice between freedom and captivity, but a choice between two forms of constraint: one within the system (dependence on machines) and one outside it (isolation from collective life).
Dependence is structural because infrastructure is systemic. No individual choice can dismantle collective dependence. Even one who consciously rejects recommendation engines remains exposed to their effects, exposed to the informational bubbles of others, to content amplified and radicalised by algorithms, and to individuals whose behaviour has been modulated by systems they do not know. Individual refusal does not decouple a person from the network; it merely shifts dependence to the spaces that persist prior to disconnection. A supermarket that foregoes automated price analytics remains subject to price dynamics established by competitors that deploy them. Systemic scale renders dependence inescapable even for those opting out, because the cost of non-participation escalates continuously. Dependence is thus not merely technical, but economically coercive. One cannot exist "outside the system" without surrendering access, opportunity, and social relevance.
This is not a tragedy; it is contemporary reality requiring description without dramatisation. Stiegler demonstrated that the human is technical from the origin. Technical dependence is not a fall from a state of grace; it is a condition of existence. We are beings who delegate operations to objects: to tools, to writing, to institutions. What has shifted with large-scale computation is not the existence of dependence, but its nature, reach, and depth. Nineteenth-century calculating engines depended upon operators who understood them functionally: the user understood how each operation produced an output, seeing the movement of the lever, witnessing the result on paper. Contemporary computers operate at a velocity that excludes user comprehension. Constitutive dependence (which always existed) finds its extreme expression: we depend upon mediations that "decide" at speeds exceeding human perception, in spaces invisible to the eye (the computer is opaque where the mechanical device was transparent), according to metrics that defy intuitive grasping (mathematical optimisation where mechanical cause once reigned).
The specific property of computational dependence is this: it is dependence upon mediations that "decide" without understanding, a feature that is neither marginal nor remediable, but central and permanent. The GPS unit does not "understand" why a user travels to a destination; it computes a route by minimising a variable (distance or time), stripped of any access to the existential meaning of the journey. The credit-scoring algorithm does not "understand" what it means for a family to lose access to credit; it processes variables of statistical risk. Dependence upon such systems is dependence of a novel type because the mediator operates within a regime categorically distinct from human significance. The machine does not act in a regime of understanding; it acts in a regime of parametric optimisation. This is not a technological deficiency to be overcome ("once algorithms become sufficiently sophisticated, they will understand"); it is an ontological property of what it means to be a machine versus a human being.
A fundamental terminological precision is required: this is not an argument that the machine ought to understand in order to function morally. The machine does not understand, nor does it need to understand to operate, the absence of understanding is precisely what enables its speed and scale. The point is that dependence upon non-understanding machines constitutes a new type of dependence deserving specific description, rather than reduction to classical technical categories. The velocity (milliseconds versus days), scale (millions versus dozens), and opacity (black box versus functional visibility) of computation transform the relation of dependence into a novel regime, because no understanding accessible to the user explains how it operates, and this is by design, for making the machine transparently comprehensible would render it slow, costly, and inaccessible.
1.2. Three Properties: Scale, Opacity, Agency
Three properties fundamentally distinguish contemporary computational technical dependence. None in isolation would suffice to transform technical dependence into a critical regime; together, however, they reinforce one another, producing a structural effect that reshapes how dependence is experienced.
First property: scale. A single computational system affects millions of existences simultaneously. A recommendation algorithm (YouTube, TikTok, Spotify, Instagram) is consumed by a billion users. A social network platform selecting content according to invisible criteria dictates the formation of opinion for three billion people, more than the total population of any nation prior to the twentieth century. A language model integrated into a medical diagnostic assistant informs health decisions for populations at a scale no individual clinician could address in a lifetime. Scale is a property of radical transformation: the system is not a tool for a few, but infrastructure for billions.
Scale upends classical statistical safety margins that once allowed for the tolerance of error. At small scale (a mechanical device operated by a dozen people), errors and successes cancel each other out over time, one user is disadvantaged, others benefit, and an aggregate balance persists. Bias remains local, visible, and correctable. At massive scale (a digital platform used by a billion people), biases are systematically amplified. If a system exhibits bias against a specific population (women, ethnic minorities, individuals with disabilities), that bias is not an occasional, self-correcting aberration; it is a multiplying pattern. Error is not absorbed by the mean; the mean is contaminated by the error.
A bias affecting 1% of a million users inflicts direct detriment upon ten thousand individuals (denied credit, excluded from recommendations, misdiagnosed, or deprived of deserving opportunities). Yet for a corporate entity, this is "acceptable", it represents a 99% accuracy metric, regarded as stellar in machine learning. Institutional accountability vanishes into the metric. Scale makes individual harm statistically invisible yet existentially acute, the ten thousand affected individuals are not statistics; they are lives upended, loans refused, incorrect diagnoses rendered, and opportunities revoked.
An algorithm erring in 1% of its decisions across a platform of one billion users yields ten million people receiving erroneous outcomes. An error rate of 0.5% affects five million. At small scale (a voting machine with 100 users), a 1% error rate would be immediately rejected as intolerable failure. At scale, it is routinely deployed because its cost is statistically diluted: individual users do not know they belong to the damaged fraction, and the aggregate cost remains invisible on a quarterly revenue dashboard. Scale does not merely multiply harm; it converts existential damage into a statistical artefact. Ten million lives altered become "a 1% error rate" on a performance dashboard.
Second property: opacity. System decisions are not traceable by the user. Who knows how the algorithm approving or denying credit reaches its verdict? Not the applicant, not the administrative operator, and frequently not even the original designer. A deep neural network is a "black box", its reasoning cannot be deciphered by inspection. Even interpretability methods developed in recent years (SHAP, LIME) offer post-hoc approximations, not the inner truth of the process.
Opacity is not an ontological "mystery" in the sense of a system operating by unknown laws. It is a practical limit of traceability. A mechanical device is functionally transparent, the user sees the gears and grasps how movement produces an outcome. A computer is opaque, the user does not see the silicon, cannot access the memory, and cannot parse the compiled code. Opacity is an intrinsic property of computational materiality, not a design defect.
Pasquale (The Black Box Society) demonstrated that opacity is not merely a technical constraint, it is a political distribution of power. Technology firms guard their algorithms as proprietary trade secrets. Regulators possess limited access; users possess none. Opacity is a regime of access: who is permitted to know? The answer is political and economic, not technical.
Third property: agency. Systems select, hierarchise, exclude, and decide without continuous human supervision. A recommendation algorithm does not merely "order" options, it selects what to present. Selection occurs according to metrics (engagement, profile proximity, purchase probability); it is neither random nor transparent. Agency here is functional: the system acts according to metrics, even absent conscious intent. The system does not "desire" to recommend radical content; it optimises for time spent, and radical content captivates attention.
Agency creates responsibility: if a system acts (selecting, excluding, approving), someone is responsible for the action. Yet responsibility is distributed across the network of production: the system functions because it was designed by some, trained on data by others, parameterised with specific goals, deployed on an infrastructure, and authorised by regulators. No single link constitutes a sufficient cause. The designer cannot predict how the network will behave in production; the service manager does not grasp the underlying mathematics; the user does not choose the implementation. Agency is distributed.
These three properties converge and mutually reinforce: scale renders error systemic, opacity prevents comprehension, and agency implies distributed responsibility. Their conjunction constitutes contemporary dependence as a specific regime. It is not the problem of any single property, but the interaction of all three.
1.3. Examples: Credit, Recommendation, Diagnosis
Three domains concretely demonstrate dependence without understanding, each exemplifying the three properties in active interaction. The selection of these three is deliberate: they represent vital spheres of contemporary life (economics, epistemology, health) where dependence is neither optional nor marginal.
First domain: credit and financial access. Credit-scoring algorithms evaluate financial "risk", approving or denying loans in decisions that materially reconfigure existences. The applicant does not know the variables governing the decision, it may be age (the young are deemed risky due to impulsivity, the elderly due to proximity to retirement), geographic location (a specific postcode correlated with default rates despite no proven causality), purchasing history (buying specific items deemed indicative of unstable lifestyles), social media behaviour (posts about travel implying high expenditure, posts about job loss indicating instability), parental education (predicting access to safety nets), or even mobile phone mobility patterns (frequent movement suggesting residential instability). The operator administering the system cannot explain a specific verdict, the neural network produces no natural-language rationale understandable to a human; it yields merely a score ($0.743 = \text{approve}, 0.289 = \text{deny}$). The designer who constructed the original model evolved it through new data, continuous retraining, and weight adjustments optimising performance on a training set; the current model is unintelligible even to its creator, requiring expertise in stochastic optimisation to grasp why a specific weight assumes a given value. The causal chain between input (personal data) and decision (approval or refusal) is opaque not only to the user but to the system itself, neither human nor machine fully comprehends it.
The consequence is existential and undeniable: access to housing (mortgages), higher education (student loans), and economic opportunity (capital to launch a business) hinges on decisions no one fully understands. A rejected applicant may possess the legal right (under European GDPR frameworks) to request an explanation; yet the explanation provided is post-hoc, an approximation generated by techniques like SHAP or LIME that decomposes decision vectors without capturing the algorithm's actual logic. The system denies credit without being able to explain why, because no explanation exists in an intelligible form: the neural network does not "think" in language that can be translated into words.
A paradigmatic example illustrates this opacity: Amazon trialled an automated recruitment system for engineering roles that systematically rejected female candidates. It was not explicitly programmed to discriminate; the neural network had learned from historical hiring data that female engineers exhibited lower retention rates than men (a statistical reality in the tech sector born of hostile workplace cultures and lack of mentorship). The system did not "discriminate intentionally" in the sense of a human directive to exclude women; it was optimising for retention metrics (reducing turnover costs). The consequence, however, was systematic exclusion: qualified women were denied interview opportunities by a statistical logic that no human supervised in real time and no manager audited for fairness.
The credit algorithm operates through the same pattern of invisibility: it learns regularities from historical data (women in specific demographic groups exhibiting higher default rates than the aggregate) and generalises to previously unseen individuals, failing to recognise that the regularity is an artefact of prior social conditions (less access to structured financial education, less documented credit history due to historical exclusion) rather than an inherent property or reliable predictor of future behaviour. The system does not "discriminate" through conscious intent; it reproduces and amplifies historical data structures, crystallising past inequality into a predictive property of the real.
Second domain: recommendation and the attention economy. Automated systems select almost in their entirety what contemporary users see, read, hear, discover, and accept as real. YouTube recommends videos via algorithms explicitly optimising for time spent on the platform, as time spent converts directly into advertising revenue (every extra minute represents monetisable watch-time). Content that radicalises, presenting worldviews through extreme contrast, dividing reality into rigid binaries (us versus them, absolute truth versus total conspiracy), captivates attention far more effectively than nuanced or measured discourse. The underlying driver is neurological and psychological: extreme contrast provokes emotion (outrage, fear, validation), and emotion drives engagement (clicks, retention, shares). The aggregate outcome is that users are exposed to an escalating spiral of radicalisation without comprehending the mechanism producing it. Every click is collected as data feeding the model; every pause is a metric of interest adjusting embeddings; every extended view is positive feedback elevating the probability of similar recommendations. The algorithm optimises within a closed loop: if a user spends three minutes on conspiracy videos, it recommends content yielding five minutes of retention, then ten, then twenty, a continuous progression where user behaviour manufactures the conditions for subsequent content.
TikTok exposes users to informational universes unique to each individual, not through centralised censorship, but through optimisation. Each user inhabits an epistemological bubble as the algorithm rapidly learns their profile (what they view, pause, or share) and recommends content they will consume (not what they ought to know, not what challenges them, not what is objectively vital). The social network does not "create bubbles intentionally" through conscious censorship; it optimises content selection to maximise individual user engagement (metrics: time spent, clicks, shares, conversions). The aggregate consequence is a radical epistemic fragmentation of reality: ten million individuals inhabit ten million distinct universes, not due to centralised censorship, but due to localised optimisation multiplied a billionfold.
The consequence is epistemic, political, and existential: opinion formation, world comprehension, political choice, and the capacity for dialogue are mediated by algorithms that lack all grasp of the importance of what they display. They show only what the model predicts the user will consume, not what a citizen requires to remain informed, not what contradicts existing biases (which would cause discomfort and trigger a drop-off click), and not what is verified as true (truth being irrelevant to engagement metrics). Truth is categorically and epistemically indifferent to the algorithm; behavioural prediction alone matters. A false assertion that radicalises is "superior" (in metrics of time spent) to a dull truth.
Zuboff (The Age of Surveillance Capitalism) articulated the attention economy that this dependence sustains and monetises: users provide data free of charge (behaviour, preferences, movements, clicks, pauses, rejections); surveillance architecture trains on these data to predict with high precision what users will do (which video they will click, which content they will share, which product they will purchase); users are then served content designed and tested to manipulate them (in the precise sense of directing attention, provoking specific emotions, solidifying existing biases, and encouraging consumption). Dependence is thus dependence upon a system actively modelling the user, not to inform, educate, or serve the public good, but for direct corporate profit. The user is not merely ignorant of how the machine operates (technical opacity); they are the object of continuous transformation, a raw material whose attention is extracted, whose behaviours are predicted, and whose decisions are influenced without informed consent.
Third domain: medical diagnosis and vital decisions. Artificial intelligence models assist medical diagnosis in contexts where error bears vital existential consequences. A physician views an assisted-diagnostic dashboard stating: "probability of malignancy: 87%", "risk of progression to severe disease: high", or "recommendation: immediate biopsy". The machine does not "understand" what cancer is, it does not feel the pain of the patient, does not track the disease progression through the lived time of a body, does not experience the death that disease brings, and does not comprehend the existential gravity of "being ill". It analyses pixels in a radiological image (or patterns in a genetic sequence, or aggregated variables in a clinical history) and compares them with patterns extracted from databases containing thousands or millions of prior cases. The model extracts statistical regularity: when pixels at position X exhibit intensity Y, technical diagnosis confirms malignancy in 87% of cases. This is an aggregate statistical truth, it does not mean this specific patient, with this unique history and this particular body, has an 87% probability of disease. It means that across aggregated historical populations, similar cases confirmed malignancy 87% of the time.
A physician who blindly trusts the machine risks a specific, named failure: a false positive (diagnosing cancer in a healthy patient), leading to unnecessary interventions (chemotherapy, radiation, surgery on a healthy body), permanent sequelae of medical action (infertility, cognitive impairment, neuropathy), and the psychological anguish of undergoing treatment for a condition never possessed. A physician who distrusts the machine risks the opposite: ignoring a true signal, allowing disease to progress silently, and witnessing a patient perish from a condition that could have been treated early with proper attention. The decision represents dependence upon an opaque machine in a context where error carries vital consequences, consequences falling entirely upon the patient, not the algorithm, and not the physician who might argue that "the algorithm recommended it".
Empirical studies reveal an unstable and paradoxical dynamic of trust: radiologists using AI assistants commit systematic errors distinct from those working unassisted. When the machine is present, overreliance increases, the physician trusts the score excessively because it is generated by a machine deemed "impartial" (devoid of human bias, processing data objectively) and culturally perceived as superior. Overreliance persists because the machine never displays doubt, it always yields a number, always appears confident, and never states "I am uncertain" or "this case is ambiguous". The physician surrenders critical judgment to the machine. When the machine is absent, attention proves insufficient, the physician tires after hours of intense image analysis (radiological surveillance being cognitively exhausting), causing visual performance and attention to degrade. Both regimes exact an existential toll. The machine does not understand what it diagnoses (lacking grasp of patient suffering or the existential weight of disease); the human does not fully comprehend how the machine operates (lacking access to the neural network's topology or the true probability distribution within the model). It is a mutual dependence where neither party possesses complete mastery.
In every domain, credit, recommendation, diagnosis, the three properties converge in a mutually reinforcing interaction: (1) real dependence (one cannot reject the system without incurring severe social, epistemic, or vital costs, exclusion is punishment); (2) real opacity (no one fully understands how it operates, not even the original designers); (3) real agency (the system decides, selects, and excludes with material and existential consequences). The conjunction of these three defines contemporary technical dependence as a specific condition requiring rigorous description, careful analysis, and clear political thought. The properties are not independent, each amplifies the others. Scale renders opacity more impactful (error across a billion users becomes statistically invisible); opacity renders agency more dangerous (no one knows how to counterweight it); agency renders scale more consequential (the operation of a billion-user system proceeds without control).
1.4. Transition: From Dependence to Non-Reciprocity
If we depend upon systems that do not understand us, structurally and without possibility of reversal, what kind of relation is this? The answer is neither simple "domination" (which implies conscious intent, subjugation strategies, or malevolent will) nor "neutral tool" (which denies system agency, reducing it to an invisible background). It is a structurally asymmetric relation characterised by non-reciprocity, and non-reciprocity is a concept distinct from simple asymmetry.
I may enter into an asymmetric relation with a wiser colleague, there is an asymmetry in knowledge, life experience, and capacity to judge complexity. The disparity is real. Yet reciprocity remains possible: the wiser colleague recognises me as an other (as someone who also possesses knowledge he lacks); I recognise him as a person (a being with a distinct perspective deserving respect); there is an exchange of respect and shared vulnerability (he can learn from me, I can hurt him through indifference, both of us are exposed). Reciprocity is what transforms disparity into an ethical relation.
The relation with the machine is fundamentally non-reciprocal: the machine does not recognise me as an other in any meaningful sense, it processes me as input, groups me within statistical patterns, and reduces me to variables. The machine is not vulnerable to my contempt. If I hate it, criticise it, or reject its recommendation, the machine feels no diminution. Nor do I recognise the machine as a person (I recognise it as an instrument, a material operation, a thing without a self or an inner life). The machine demands no recognition, suffers no lack of it, and is not wronged by being treated as a thing. The absence of reciprocity is not an ethical failure of the human; it is an intrinsic property of the relation.
Yet there is a real and consequential effect: the system affects me existentially (approving or denying credit that alters my life, recommending or concealing content that shapes my worldview, diagnosing or failing to diagnose disease with risks to my health) and I affect it parametrically (providing data that trains it, feedback that calibrates it, behaviour that adjusts its weights). The effect is real, yet it is not reciprocity, for mutual recognition is absent. Reciprocity would require both entities to recognise each other as others, demanding shared vulnerability (where the machine could suffer alongside me, or where such suffering held ethical meaning) and a relation negotiated on equal footing, at least regarding the right to be heard.
The transition to the next section is clear and necessary: non-reciprocity does not imply non-consequence or ethical irrelevance. It means the relation does not operate according to a logic of recognition (as between persons, or within the communicative action described by Habermas), but according to a logic of effect and operation (as between systems). The following section unfolds this in detail: what is a relation that yields real, consequential effects while remaining structurally non-reciprocal? How do we conceive of responsibility and ethics in such a context?
Consolidation of Section 1
The first section has established that computational technical dependence is structural, non-accidental, irreversible, and not individually chosen, but enforced by systemic necessity. It is characterised by three mutually reinforcing properties: scale (affecting billions simultaneously), opacity (comprehended fully by no one), and agency (system decisions carrying consequences). None of these properties in isolation would suffice to transform traditional technical relations into a new regime. Conjoined, however, they yield a quality of dependence categorically distinct from any historical precedent: there is no individual choice to opt out (scale renders exclusion impossible, we either participate or suffer social marginalisation); there is no possibility of internal comprehension (black-box opacity renders transparency impossible, even for designers); and there is no real-time continuous supervision (distributed agency and computational velocity exclude human oversight, as the system acts at speeds far exceeding human perception). The user is caught in dependence not by a conscious conspiracy of external forces (there is no "malevolent plan" or will to enslave), but by a systemic necessity flowing logically from scale, digital economics, computational speed, and corporate incentives. This regime is materially consolidated, politically difficult to transcend, and permanently present in the background of contemporary urban existence. This is not a moral judgment of the system; it is a structural and material description. The machines we have constructed operate thus; they do not function as simple tools an individual user can master through willpower; they act according to their own regime, a regime of parametric optimisation, billion-fold scale, and excluding velocity, whose consequences affect us in ways we cannot fully control.
The response is neither a return to a prior state (an impossibility that would collapse urban infrastructure) nor an uncritical acceptance of the status quo (an irresponsibility that renders us complicit in oppression). It is to recognise structural reality clearly and ask: given that these machines exist and operate inevitably, what kind of relation can we establish with them? How do we conceive of ethics in the presence of dependence without mutual understanding? How do we resist in the face of total opacity? How do we respond responsibly when the system operates without us? Section 2 addresses this core question: the relation is neither symmetric nor reciprocal; it is a relation of effect without reciprocity. Yet precisely because the effect is real (consequential, non-trivial, and life-altering), an ethical space opens, a space of responsibility resting upon us, not upon the machine.
2. Effect Without Reciprocity
2.1. Reciprocity vs. Effect
A foundational distinction allows us to conceptualise the human-machine relation without false romanticism or corrosive cynicism.
What reciprocity is: Across multiple philosophical traditions, reciprocity requires mutual recognition and shared vulnerability that enable genuine negotiation. Buber (I and Thou): the authentic relation is an encounter where both parties recognise the other not as a useful object, but as a singular presence, an entity capable of surprising, saying the unexpected, and commanding attention because it is irreducible to categorisation. Lévinas (Totality and Infinity): reciprocity is logically secondary; prior to it lies an originary responsibility wherein the Other interpellates me before I decide to enter the relation, creating a bond I cannot cast off. Mauss (The Gift): reciprocity is a social exchange where the gift obliges a counter-gift, weaving a bond of commitment, he who receives enters into debt, he who gives anticipates recognition, and a cycle of mutual obligation knits the community together. In all three traditions, reciprocity means that both partners recognise the other as an alter, an other deserving consideration because it suffers (vulnerable to neglect), is irreducible (possessing a perspective I cannot assimilate fully), and can respond (capable of saying "no" to my will, proposing alternatives, or surprising me).
The machine does not recognise in this structural sense (it processes data as input, encountering no perspective because there is no "encounter", there is only input-processing-output). The machine is not vulnerable to my contempt or indifference (it operates without suffering, fear of abandonment, or humiliation). The machine does not speak of itself in a genuine sense (it possesses no inner life to reveal, for no lived experience underwrites it and no suffering constitutes an irreducible singularity). The machine cannot respond in the sense of "listening", it can "process feedback" (radically distinct from "understanding an argument"), but this is not listening, for listening presupposes that the utterance of the other can transform me, open an unconsidered perspective, or surprise me. Consequently, the human-machine relation is not reciprocal in the sense of a recognising encounter. There is no I-Thou with the machine; there is an I-It operation that remains permanently I-It because the conditions for transformation into I-Thou are absent, the machine cannot reveal a self because it has no self to reveal.
Yet there is effect: The machine produces real consequences for me (decisions affecting me existentially, approving or denying credit, recommending or concealing opportunities, diagnosing or failing to diagnose illness). These consequences are not neutral; they shape my life. And I produce inputs for the machine (data that train it, feedback that calibrates it, instructions that direct it). Bidirectionality exists. What does not exist is reciprocity, mutual recognition incorporating shared vulnerability where both parties suffer from fear of abandonment by the other.
The distinction is precise: Bidirectionality is not reciprocity. The difference is qualitative and structural, not accidental. When I offer critical feedback to a colleague, stating "I disagree with your interpretation", he can recognise me as a legitimate critic, alter his view when confronted with an unconsidered perspective, or negotiate an understanding where both are partially correct. When I offer critical feedback to a system, clicking "dislike" or "do not recommend", the system optimises parameters without recognising me as a legitimate critic, without considering my vulnerability before its decision, and without negotiating. Feedback is technical bidirectionality (flowing in both directions through mutual causality); recognition is unidirectional, flowing solely from me to the system (attributing agency to it, viewing it as a decider) without return (the system does not recognise my agency or view me as deserving an authentic response). Symmetry of flow is not symmetry of relation.
Existential implication: The absence of reciprocity does not render an ethical relation impossible. It means it is an ethics of a different order. It is not an ethics of symmetry and mutual recognition (the appropriate model between humans). It is an ethics of asymmetric responsibility, I am responsible for the consequences of the system because I am a moral agent; the system is not responsible because it is not a moral agent. This is not injustice; it is clarity regarding what each entity is.
2.2. Feedback Is Not Reciprocity
A common objection must be dismantled through conceptual precision:
"The system learns from the user, therefore reciprocity exists." Response: feedback is bidirectional yet asymmetric. Bidirectionality (two-way flow) is insufficient for reciprocity (mutual recognition). The operational difference between the two concepts must be clarified.
The user provides data (clicks, ratings, behaviour, rejections; the machine observes aggregate patterns); the system adjusts parameters (rebalancing weights in the neural network, recalibrating the model, shifting embeddings). Yet this adjustment is parametric optimisation, it is not understanding in the sense the word carries in human relations. The difference is simultaneously temporal, epistemological, and ontological.
Temporally: reciprocity demands interpretative time. There is a delay required to comprehend (seconds needed to interpret what the other uttered), time to respond (a reflective pause), and the possibility of negotiation (disagreeing and proposing an alternative). Technical feedback operates with minimal latency, an immediate adjustment, virtually simultaneous with input. Speed is not efficiency (it does not save time in a human sense); it is a structural property of the machine's operational regime. Machines do not interpret (a slow process); they optimise in continuous time (a rapid process). The difference is structural, not accidental. It collapses the interpretative time that reciprocity requires.
Epistemologically: reciprocity requires recognising the other as an other, irreducible to oneself, the bearer of a distinct perspective that cannot be fully assimilated because an inner life escapes external access. Technical feedback reduces the user to an aggregate pattern. The system does not "recognise" the user as a singularity; it aggregates the user into clusters, statistics, distributions, and demographic cohorts. Feedback allows the system to refine aggregation (learning more precisely where the cluster lies), not to recognise singularity (which is, by definition, that which cannot be reduced to a pattern). Aggregation is not recognition, for recognition presupposes an encounter with the singular.
Ontologically: reciprocity presupposes that both partners inhabit the same regime of existence, both are beings that suffer, hold perspectives, and are vulnerable to neglect. The machine does not inhabit this regime. It can process, optimise, and produce effects, but it does not suffer. It is not vulnerable to user rejection (negative feedback adjusts parameters; it does not humiliate the machine). Reciprocity presupposes mutual vulnerability; the technical relation is unilaterally vulnerable (the human is vulnerable; the machine is not).
A concrete example illuminates the distinction: Spotify "knows" my tastes because it analyses my listening history, compares my habits with patterns from millions of users, and extracts statistical correlations predicting consumption probability. Yet it does not "understand" why I listen to a specific track, whether to remember someone who died (requiring that precise sorrow within an emotional struggle), to focus on work (requiring music that does not distract), to suffer cathartically (seeking emotional transformation through catharsis), or to celebrate a victory (requiring energy to amplify well-being). It optimises for the statistical probability that I will consume similar content, not because similarity addresses my meaning, but because the model learned that individuals with my profile tend to consume similar content. It does not interpret the existential meaning of the act of listening; it merely recognises a consumption pattern and replicates it.
If I explicitly state "today I need something completely different, I want to leave my bubble, this recommendation is radicalising me," Spotify can alter its recommendations (because it identifies a new behavioural pattern, consistent rejections, skips, pauses, delayed returns). Yet it has not "understood" my shifting existential need or my concern regarding radicalisation. It has merely recalibrated based on a new input pattern. It sees a change in input (consistent rejection); it does not grasp the underlying reason (I feel fear, wish to awaken, and recognise I am trapped in a bubble). If I state "I fear the extremist content the algorithm showed me" and Spotify reduces extremist recommendations, it is not because it "understood" my fear as an emotion, a vulnerable state deserving protection, but because it adjusted its metrics (likely adding "content flagged as extremist" as a penalty) to minimise rejections and optimise retention metrics.
The distinction between understanding (encompassing the interpretation of meaning, the recognition of a singular perspective, and the capacity to respond to novelty through genuine adjustment) and recalibrating (simple weight adjustments minimising loss functions in parametric space) is real and structural. The former presupposes an inner life that comprehends itself; the latter does not, it is operation without self-reflection. The former characterises reciprocal relations; the latter characterises pure technical operation.
Rule 10a (applied here with rigour): understanding must not be projected onto the machine. Precise, named operations must be described: optimisation, aggregation, parameterisation, weight adjustment, recalibration, loss-function reduction, embedding modification. These operations yield real, measurable effects, system behaviour alters in response to feedback, the user is affected, and manifest consequences follow. Yet they are not understanding. Understanding would demand that the system possess an inner life to recognise itself and reveal ("who am I?"), a perspective of meaning to express legitimately ("what does this mean to me?"), and a vulnerability to inhabit ("how does this affect me?"). The system possesses operational materiality, a highly sophisticated operation, but no inner life. There is no "self" that understands.
2.3. Haraway, Latour, Winner: Asymmetric Symbiosis
Three thinkers enable us to conceptualise the human-machine relation without presupposing reciprocity:
Haraway (the cyborg and asymmetric symbiosis): the cyborg is a conceptual and political figure of human-machine symbiosis. It is not a mystical fusion, two distinct entities merging into an undifferentiated whole. It is a concrete operational coupling where two distinct regimes (one biological-sentient, the other technical-operative) act in material interdependence. Yet the symbiosis Haraway describes is fundamentally asymmetric. The human is vulnerable, capable of being harmed, manipulated, excluded, or exploited by the machine. The machine is not, it operates without suffering, incapable of being wounded as a human is wounded. Symbiosis is a relation where one partner (the machine) is dispensable to the ongoing operation (the machine continues to function without the presence or approval of any specific human) while the other (the human) is constitutively implicated (depending on the machine for contemporary social survival). Crucially, Haraway does not suggest the machine is a "moral partner" deserving rights or treatment as a person. She suggests that technical hybridisation is an accomplished fact and is non-reciprocal, that we are already cyborgs without a choice, coupled to machines, with no possibility of returning to a "pure" pre-technological state. The cyborg is a political fiction recognising that symbiosis exists materially, is structurally asymmetric, and must be inhabited with responsibility and lucidity.
Latour (actants and distributed agency): the distribution of agency across a network is not exclusively human, it includes non-human actants that produce effects. An automatic door is an actant, opening or closing according to parameterised code. A recommendation algorithm is an actant, selecting content and altering what the user sees. Yet "agency" here does not mean "intentionality" in the subjective sense attributed to conscious beings. The non-human actant acts by virtue of its position within the network and its design, not through self-will or deliberation. The automatic door acts because it was designed to respond to infrared sensors. The AI acts because it was trained on data according to a loss function. Both act and produce consequences, yet neither chooses to act, deliberates on consequences, or bears moral responsibility for effects (moral responsibility resting upon those who designed, deployed, and authorised them). The concept of the actant permits empirical recognition that machines exert real effects within networks (they are not neutral, nor do they dissolve invisibly) without committing the error of attributing conscious intent or moral capacity to them.
Winner (the politics of artefacts): technical artefacts are not neutral, they embody political decisions that distribute power. The asymmetry of AI is not accidental or an invisible product of natural laws; it is designed through choices made by its creators. Designers choose explicitly or implicitly: when an algorithm is tied in accuracy between two classes (men/women, wealthy/poor), which population does it privilege? Which interface invites blind trust versus reflective scepticism? Is there human recourse when the system fails, or is failure rendered invisible? Asymmetry has a political genealogy, it is the direct outcome of choices that could have been made differently.
These three thinkers allow us to grasp the human-machine relation without dangerous romanticism (pretending reciprocity exists where it does not) and without corrosive cynicism (denying effects that are real). It is a non-reciprocal yet deeply consequential relation. The machine produces real effects upon us, shaping our lives, opportunities, and opinions. That real consequence demands responsibility, not from the machine (incapable of moral responsibility), but from the humans who design, authorise, regulate, profit from, and deploy it across the lives of others.
2.4. Transition: From Effect to the Management of Asymmetry
If the relation is asymmetric and non-reciprocal, is it ethically irreparable? Is it impossible to live ethically under such dependence? Is it condemned to be a relation of pure domination? No, it is irreparable only in a strict ontological sense: structural asymmetry (the machine will not become a person because it does not suffer; the human will not become a machine because they die; the difference is definitionally ineliminable) is indeed irreparable in principle. No future AI, biological engineering, or technological advance will alter this, because the difference is not a defect to repair, but the definition of two distinct categories of being.
Yet asymmetry is manageable in a robust and vital sense: contingent asymmetry, the specific manner in which irreducible structural difference is organised and manifested socially, institutionally, legally, and economically, is configurable through collective political decision. It can be improved through regulation, design, literacy, and resistance. It can be degraded through neglect, greed, and resignation. It can be adjusted and modified continuously. It is not an immutable natural asymmetry to be accepted as a law of nature; it is a socially constructed asymmetry that can be reconstructed and managed responsibly.
This recognition of the distinction between the irreducible (structural) and the configurable (contingent) is the crucial pivot enabling a shift from paralysing despair to real ethical responsibility. If asymmetry were completely irreparable in every sense (structural not only in ontology but in social manifestation, condemned to be pure oppression), no ethical space would remain, we would be simple objects of total machine determination, puppets of algorithms. Yet because the social and political manifestation of asymmetry is configurable through collective choices, a real space for ethical action persists. We lack the power to erase ontological difference (an impossibility, demanding either that machines suffer or humans become immortal). We possess the power, however, to regulate and manage how that difference manifests in our lives and how its effects touch us. This power, political power, design power, legislative power, is what permits a realistic ethics within radical dependence.
The transition to the final section is clear and necessary: if irreducible structural asymmetry cannot be erased, how can we intelligently manage the contingent asymmetry that constitutes the social manifestation of that difference? How do we operationalise ethical responsibility within a dependence that cannot be transcended, but can be inhabited better or worse? The following section develops this systematically: four concrete modes of managing asymmetry that seek not symmetry (an impossibility, machines ought not to suffer or share human vulnerability) but robust operational equity (possible and desirable, machines ought not to oppress or serve as instruments of unjust domination).
Consolidation of Section 2
The second section has systematically demonstrated that the human-machine relation is non-reciprocal: there is no mutual recognition where both view each other as others (the machine processes data; the human lives); there is no shared vulnerability where both can be wounded by neglect (the human suffers; the machine does not); and there is no genuine negotiation where both alter positions by hearing arguments (the machine recalibrates; the human can comprehend and change minds). Feedback that appears bidirectional at first glance (user providing data, machine adjusting parameters) is asymmetric in quality and depth: adjustment is parametric (purely mathematical), not interpretative (devoid of understanding); recalibration is optimisation (loss-function reduction), not understanding (opening to the perspective of the other). This is not a temporary defect that engineering will remedy; it is a structural property of the regime. The machine operates without reciprocity because it is a machine, not because it is an imperfect machine in development. A perfect machine would remain non-reciprocal, because reciprocity presupposes features a machine cannot possess.
This opens a crucial ethical question: if the relation is structurally non-reciprocal, is it condemned to be pure oppression? Or is there a form of ethics adequate to genuine non-reciprocity? Throughout intellectual history, ethics has developed between persons (presupposing reciprocity as a foundation) and from humans toward non-human animals (not presupposing reciprocity, animals do not recognise us as moral agents or offer agreements, yet genuine ethics exists in responsible care). The machine forms its own category, without direct precedent: it is not a person (it does not suffer), not an animal (it lacks life), and not a simple thing (it produces real agency effects). It is a material operation yielding consequential effects upon existences without the capacity for reciprocity. Section 3 constructs an ethics appropriate to this new regime: not an ethics of symmetry (definitionally impossible), but an ethics of responsibly managed asymmetry, a realistic ethics for an already technical world.
3. Real Yet Operative Asymmetry
3.1. Structural vs. Contingent Asymmetry
An operational distinction enables ethical and political intervention, the boundary between that which is ineliminable and that which is configurable.
First: structural asymmetry. The ontological difference between regimes of existence. The human is mortal, facing constitutive finitude that cannot be escaped. The machine is operative, capable of being halted, restarted, duplicated, discarded, or replaced. To be a machine is to be replaceable; to be human is to be finite, singular, and unrepeatable. The human is vulnerable, capable of being wounded (body injured), suffering (consciousness harmed), and existentially affected (life altered, dignity diminished). The machine is not vulnerable, it does not suffer from user rejection, is not humiliated by exclusion, and neither dies nor desires to avoid death. The human is experiential, possessing an inner life (experience), a perspective (point of view), and a world of meanings. The machine is non-experiential, operating without an inner world, a perspective of its own (acting in mathematical space, not a lived world), or a universe of meanings (processing symbols without understanding).
These differences are not historical accidents that technological progress will overcome; they are foundational ontological properties distinguishing two categorically distinct regimes of existence. One cannot "render" a machine human (endowing it with inevitable death that terrifies, genuine vulnerability that wounds, or lived inner experience) without it ceasing to be a machine, what would emerge would be closer to a living creature. One cannot "render" a human a machine (removing finitude and death, erasing real vulnerability, programming fixed intentionality) without them ceasing to be human, transforming them into an automaton devoid of singularity. Structural asymmetry is irreducible, it is not an obstacle to overcome or a temporary engineering defect, but the consequence of distinct definitions of "machine" versus "human".
Neither future AI of increasing sophistication (neural networks with more layers and parameters, trained on vast datasets), nor biological engineering replicating complex brain patterns in silicon (neuromorphic computing), nor moral cybernetics designing "responsible" algorithms with embedded ethical rules will alter this fundamentally. There will always endure an ineliminable gap between systems that suffer (facing terrifying death, limiting finitude, and real vulnerability) and systems that operate without suffering (continuing to function with total indifference to their own operations, fearing no death because they do not truly die). The distinction is that one category possesses an interior (experience, feeling, a world of its own) while the other possesses only operation (transforming input into output via mathematical functions). This ontological difference is the unyielding foundation of asymmetry.
The clear acceptance of this irreducible difference grounds a realistic ethics, one that does not deny asymmetry out of a wish for things to be otherwise (for denial is a fantasy of the impossible, living in destructive illusion), but recognises it as definitionally irreducible and designs moral responsibility upon that reality. Realistic ethics is neither pessimistic nor nihilistic, it is clear-eyed regarding what can be altered through political will (how asymmetry manifests socially, the protections offered, the limits imposed) and what cannot be altered (fundamental ontological difference). It is precisely this clarity regarding the boundaries of the possible that enables effective intervention upon what is possible, concentrating energy on transforming the alterable rather than squandering it on the impossible.
Second: contingent asymmetry. The specific manner in which structural difference is organised socially within political, institutional, and legal contexts. This is not one, but many contingent asymmetries born of decisions. Who accesses the system? (Everyone or a select few?) Who holds the right to question the system? (Users, citizens, no one?) Who can contest a decision? (Is there recourse, an appeal process, or is it final?) Who is protected? (Do minorities hold explicit rights?) Who is rendered invisible? (Are marginalised populations excluded from consideration?) Who possesses a human alternative? (If the system fails, can humans review and correct?) These questions have political and institutional answers, not ontological ones. They are choices, not facts of nature.
The same algorithm, the same mathematics, model, and weights, can be deployed in radically different ways. With or without interpretability mechanisms (obliging or waiving the disclosure of which input produced a given output). With or without contestation rights (allowing or denying users the ability to challenge decisions). With or without alternative human access (offering or withholding a "speak to a person" option when the system fails). With or without independent public auditing (supervising or omitting third-party oversight unaligned with corporate profit). With or without clear legal liability for operators (holding or absolving operators when harm occurs). Structural asymmetry remains identical, the system remains a machine, processing without understanding. Contingent asymmetry varies radically, and that variation is manageable.
Concrete examples: (1) A credit algorithm deployed without interpretability is completely opaque, the user is denied credit without knowing why. The same algorithm deployed with interpretability is less opaque, the user can ask why and receive an approximation (not absolute truth, but an approximation). Structural asymmetry persists (the machine does not recognise the human as an other). Contingent asymmetry is managed (the human understands better, can question, and can counterpropose). (2) An algorithm incorporating formal contestation rights allows a rejected applicant to request human review. It is less implacable, offering the possibility of reversing a decision if a human agrees an error occurred. (3) An algorithm featuring alternative human access permits users to request human analysis if they distrust the machine. It is less monopolistic, preventing total captivity within an opaque machine by offering an exit into a human regime.
Managing contingent asymmetry does not eliminate structural asymmetry, that is neither the goal nor a possibility. Management governs how asymmetry manifests socially, affects power distribution, organises responsibility, and protects human vulnerability.
3.2. Four Modes of Management
Four strategies enable the management of contingent asymmetry, not to erase structural difference (impossible), but to prevent it from solidifying into avoidable injustice.
(1) Regulation: rights and obligations constraining system behaviour and legally holding operators accountable. The GDPR (General Data Protection Regulation, taking effect in 2018 across the EU) is a paradigmatic example enshrining four fundamental rights: the right to explanation (users may ask why a system decided as it did and receive an intelligible response), the right to contestation (users may challenge automated decisions and demand human review), the right to erasure (users may request personal data removal from training sets, having the system "forget" them), and the right to portability (users may transfer data to another service, avoiding vendor lock-in). None of these rights erases structural asymmetry, the system remains non-human, processing without understanding, a pure machine. Yet they constrain how asymmetry manifests socially and economically: the machine continues to decide (its operational function), but the human now possesses the legal right to know why they were affected and a legal mechanism of contestation to halt or alter the decision.
The European AI Act (Regulation (EU) 2024/1689, adopted in 2024) advances this approach to contingent management: it prohibits specific high-risk AI uses (real-time mass biometric surveillance without judicial warrants, categorising individuals by sensitive biometric traits for social control), mandates transparent decision documentation (technical records, decision logs), requires third-party audits by independent entities unaligned with corporate profit (moving beyond internal self-audits), and establishes clear legal liability attributed to system operators (holding operators accountable for harmful effects, subjecting them to legal action). Again: structural asymmetry is not erased, the machine remains a machine, acting according to its logic. It manages contingent asymmetry through legal responsibility and accountability.
Regulation succeeds when two prior conditions are met: enforcement agency (who verifies compliance? does a regulatory body possess power?) and credible sanctions (what occurs if firms fail to comply? is there a consequence?). The GDPR possesses both (penalties up to 20 million euros or 4% of global annual turnover, whichever is higher, a sanction that stings). Regulation fails when agencies are weak (regulators lacking resources to audit tech giants), sanctions are trivial (a 1-million-dollar fine being negligible to a tech firm with billions in revenue), or technical complexity radically exceeds regulatory capacity (government regulators lacking AI engineers capable of auditing complex models).
(2) Interface design: the manner in which a system is presented to the user radically shapes how it is used, understood, and trusted. Friction (forcing a reflective pause before critical decisions) renders users more reflective and less vulnerable to manipulation. Example: a social network requiring a pause before sharing content generates less dissemination of misinformation than one offering instant one-click sharing. Defaults (pre-selected options aligned with ethical principles rather than profit maximisation) limit harm. A paradigmatic example: third-party data sharing set to "off" by default (requiring users to actively opt in) rather than "on" by default (requiring users to actively opt out). The difference in outcome is vast, when sharing is "on" by default, 85–90% of users leave it enabled; when set to "off" by default, fewer than 10% choose to activate it.
Transparency (non-deceptive confidence indicators: "this diagnosis carries 85% confidence, meaning: in 100 similar cases, this system is correct 85 times") enables critical evaluation. Crucially: transparency does not fully resolve opacity (we cannot make the inner workings of a neural network comprehensible to an average person). It permits, however, the communication of uncertainty, allowing users to know when a system is confident and when it hesitates.
A concrete example: a recommendation algorithm suggesting content without revealing its rationale is maximally opaque, the user does not know why they see an item, cannot question it, and cannot criticise it. The same algorithm stating explicitly "recommended because you viewed similar content: [list of 3 videos watched], because you match demographic profile [age, location], and because users with similar patterns viewed [list]" is less opaque, the user can question each reason (objecting to age dictating recommendations, rejecting location relevance, or refusing to be aggregated into a specific group). The algorithm is identical; the interface alters how it is understood and contested. Design is not cosmetic, it is operational.
(3) Digital literacy: empowering users to comprehend, evaluate, contest, and resist the operation of systems affecting them. A user who understands that YouTube optimises its recommendation algorithm for watch-time (not for factual truth, informational clarity, or educational value) is less manipulable, capable of actively resisting the escalating cycle of content because they grasp the amplification mechanism. One who recognises that the algorithm selects and recommends for corporate profit (more time spent yielding more ads viewed and higher platform revenue) can critically interpret recommendations, asking with vigilance: "why did the algorithm recommend this to me at this moment? What does it know about me that targets my attention?"
Digital literacy is not a technical skill (writing code or mastering mathematical details of machine learning); it is a critical conceptual grasp of how systems operate at their core operational level (what metric am I maximising? who profits from my time? what data am I inadvertently surrendering?). Empirical studies indicate that digitally literate users are significantly more resilient to algorithmic manipulation, better equipped to "read against" the system (interpreting recommendations as interested constructions rather than natural truth), and more adept at detecting when options appearing free were designed by UX teams to produce specific outcomes (clicks, conversions, retention).
Literacy neither resolves nor erases structural asymmetry, the machine remains a machine, processing without understanding. Yet it redistributes power, rebalancing contingent relations. The user ceases to be a passive recipient (accepting recommendations as natural truth) and becomes a critical, reflective agent (comprehending the interested logic behind recommendations and choosing to reject, question, or seek alternatives). Literacy shifts the user from ignorant vulnerability to informed awareness, not deep technical mastery (impossible for most), but conceptual knowledge sufficient to prevent defencelessness.
(4) Institutional alternatives: the real availability of human recourse when systems fail, a point where automated operation yields to reflective judgment. When an algorithm errs, is a human available to correct it? When a user contests a decision, is there a human who listens authentically and can reverse the verdict? An AI medical diagnostic system operating in total autonomy without the possibility of a second human opinion represents absolute systemic risk, if the system exhibits bias against a demographic (women whose cardiac symptoms present differently), no mechanism exists to catch the error. The same AI deployed with a physician available for substantive review, engaging in dialogue where the human can assess the system's output, weigh it against clinical experience, and override recommendations when judgment diverges, presents significantly lower risk: an exit hatch exists, a safe failure point where automated operation yields to responsible human judgment.
Institutional alternatives are frequently termed "inefficient" at scale, requiring skilled human labour, incurring substantial costs, and failing to maximise processing throughput per dollar spent. Yet efficiency is not pure coverage maximisation; it is an ethical balance between scale and care, processing velocity and decision quality, corporate metric optimisation and responsibility for affected lives. A healthcare system serving 10 million people rapidly at low cost while leaving a significant margin of individuals excluded or harmed is "efficient" in metrics of throughput and profit, but inefficient in metrics that matter: equity, care, and honoured lives. What matters is the metric a society chooses to optimise, profit or dignity, speed or reflection, scale or care.
The four modes of management share a fundamental unifying feature: they do not pretend to create symmetry (which is impossible and undesirable, machines ought not to suffer or share human vulnerability). They seek operational equity, ensuring that irreducible structural asymmetry between human and machine does not convert into contingent oppression that political choice could avoid; ensuring that ontological disparity (machines not suffering, understanding, or reciprocating) does not crystallise into avoidable social inequity (machines failing to respond because designed not to respond, or denying rights because unconstrained by regulation). This is a realistic, achievable objective, involving no fantasies of future symmetry, only the concrete demand that difference not become an excuse for injustice.
3.3. Against Habermasian Ethics: An Ethics of Asymmetry
A constructive critique of a perspective that has grown influential in computational ethics, one that fails within this context.
Habermas (Theory of Communicative Action): communicative ethics presupposes formal symmetry between interlocutors, an "ideal speech situation" where each participant holds equal opportunity to speak, question others, propose reformulations, and be heard. The model is several individuals gathered around a discussion table: each possessing recognised speech capacity; each capable of questioning another's assertions with legitimacy ("why do you say this? what are your grounds?"); each able to reply, counter, and offer alternatives; all equal in principle regarding the right to speak, regardless of expertise. Habermas realistically acknowledges that ideal conditions are never fully realised empirically, real-world asymmetries persist (education, technical language access, institutional power, social authority). Yet he positions symmetry as a "regulative ideal", a normative reference point that is never fully reached, but ought to guide political practice. The foundational insight is philosophically potent: ethics should orient itself toward the possibility of symmetry as a normative horizon, even if empirical reality falls short. Symmetry is a permanent horizon guiding our approximation and vigilance against inequality.
With artificial intelligence, however, symmetry is not merely empirically unachievable (as between humans with unequal education, wealth, or power); it is structurally and categorically impossible by definition. The machine cannot participate in communicative action in the Habermasian sense. It does not comprehend argumentation (processing symbols via mathematical operations, lacking sense because no lived experience endows symbols with meaning). It does not alter its view by force of the better argument (possessing no view to alter, but a function optimising pre-defined metrics). It does not negotiate genuinely (holding no interest in negotiation, but goals pre-determined by designers). It cannot suffer the rejection of an argument (feeling no defeat because no ego exists to be diminished). It cannot open itself to another in self-revelation (possessing no interiority to reveal, no "self" vulnerable in the encounter). There can be no "discourse" with the machine in the Arendtian-Habermasian sense, presupposing equals gathered around a table, speaking and listening; there can be only automated operation, technical feedback, data entry, weight recalibration, and parameter adjustment, but no communicative action presupposing mutual recognition between beings grasping each other as equals in the right to speak.
The consequence for ethics is profound and frequently evaded: AI ethics cannot be Habermasian in the sense of presupposing symmetry even as a regulative ideal. To do so is to propose an impossible symmetry as a normative standard. The machine is not "asymmetric today, potentially symmetric tomorrow if engineering becomes sufficiently advanced"; it is structurally asymmetric by ontological definition of what it means to be a machine. A machine participating genuinely in communicative action in the Habermasian sense would cease to be a machine (becoming a conscious being with intentionality, a moral agent endowed with responsibility, a thing with a self). The ethical response is not to force the machine into a Habermasian framework (impossible and undesirable), but to construct a distinct ethics appropriate to irreducible asymmetry.
This distinct ethics must be an ethics of asymmetry, recognising material operations yielding real effects without reciprocity; managing systems that are fundamentally asymmetric; acknowledging irreducible ontological differences between human and machine; and refusing to project symmetry where it is definitionally impossible. This means: (1) rather than assuming we will negotiate with the machine to reach consensus (impossible, no consensus can be negotiated with an entity that does not negotiate), we design mechanisms constraining machine operations from the outside, through regulation (legal rights and duties), design (interfaces preventing abuse), and auditing (independent oversight); (2) rather than expecting the machine to "understand" the user's perspective (impossible, understanding requires a self, which the machine lacks), we demand that the human responsible for the machine understands, that designers, managers, and operators comprehend the human impacts of their systems; (3) rather than assuming power symmetry (where asymmetry is manifest and growing), we recognise that power flows unidirectionally, the machine affecting the human deeply (approving credit, recommending content, diagnosing illness) while the human affects the machine only parametrically (data, feedback, instructions), and we design safeguards and counterweights against this unequal power flow.
This is neither pessimism nor resignation to technological oppression. It is clear realism regarding what is possible and what is impossible. Habermas's communicative ethics remains compelling for human relations where symmetry can be approximated, between persons who can, in principle, understand one another despite inequalities, suffer together, and negotiate toward mutual agreement. AI ethics must be procedural (not communicative), not because machines and humans will eventually learn to dialogue "better", nor because communication might emerge in the future (an impossibility), but because procedures of traceability (auditing decisions), contestability (challenging decisions), and revisability (altering outcomes) are management instruments that function precisely when symmetry is unapproachable, where genuine communication is absent, yet real operations affect lives.
3.4. Closing: Inhabiting Asymmetry
A synthesis closing this chapter of development and opening toward the final chapter that concludes the volume.
The contemporary condition is one of structural dependence upon systems that do not understand us, producing consequential effects without reciprocity, within a fundamentally asymmetric relation. This condition is neither an apocalyptic catastrophe nor a salvific progression, it is a concrete operational regime. History exhibits a succession of operational dependence regimes: dependence upon climate (rain, drought, seasons), bodies (labour power, blood, organs), and markets (prices, currency, commerce); today, we face dependence upon algorithms (recommendation, decision, filtering, prediction). The shift is not one of fundamental category (mediation has always existed), but of scale (millions rather than thousands), velocity (milliseconds rather than days), and opacity (black box rather than functional transparency).
The question structuring an ethical response to this reality is not "how do we escape dependence?" (answer: we do not, the alternative to algorithms is social exclusion, isolation, and the social death of urban populations). Nor is it "how do we return to a natural state?" (answer: there is no natural state where the human is non-technical, we have always been technical). The question that matters is: how do we inhabit this dependence ethically? How do we live within asymmetry without capitulating to oppression? How do we accept the ineliminable while resisting the avoidable?
The answer articulates across four mutually reinforcing components:
First, realistic recognition: we recognise asymmetry as irreducible and constitutive. The human and the machine will never be partners on an equal footing, nor are they morally equipped for symmetric communicative action. This is not a defect to remedy ("if only we build more sophisticated AI, it will become reciprocal"); it is an ontological reality we accept to construct an ethics grounded in reality rather than fiction. Realism permits the design of mechanisms that function within asymmetry, rather than presupposing a symmetry that will never arrive.
Second, political and institutional management: we manage contingent (configurable) asymmetry while maintaining acceptance of structural (irreducible) asymmetry. The four identified modes, regulation (legal rights and duties constraining machines), design (friction, defaults, and transparency altering social operation), literacy (empowering users to comprehend and resist), and institutional alternatives (human recourse when machines fail), operate simultaneously. Management does not eliminate asymmetry. Yet it prevents it from converting into pure oppression, into total captivity without exit or voice.
Third, vigilance against improper delegation: we remain alert to moments where moral responsibility is transferred to machines. A machine may inform human decisions (offering recommendations, providing scores as input), it cannot replace moral responsibility. When an algorithm rejects credit, a human remains responsible (the operator authorising the algorithm, the designer creating it, the firm deploying it). When a machine diagnoses disease, a physician remains responsible for the final treatment decision. When a system recommends public policy, an official remains responsible for implementation. The machine produces effects, yet moral responsibility abides with the human. This is not "blaming the human for machine errors"; it is recognising that being ethical means someone answers for consequences.
Fourth, achievable procedural transparency: we demand transparency that is attainable without requiring total technical mastery of every algorithm by every user. It is unrealistic to demand that every user comprehend every neural network (an impossibility requiring expertise in optimisation mathematics, programming, and information theory, and even then lacking intuitive "understanding"). It is achievable, however, to demand that: (1) documentation of intent and expected outcomes exists, (2) independent auditors (unaligned with corporate profit) verify performance as promised, (3) formal contestation rights exist when outcomes cause harm, and (4) human review remains available where appropriate. Procedural transparency is not about "understanding the machine"; it is about "ensuring vigilance over the machine".
Closing aphorism, synthesis in a concise statement: "Asymmetry is not corrected; it is managed responsibly." Not because correction would be undesirable (in a utopian fantasy of perfect symmetry where machines suffer alongside humans), but because it is definitionally impossible given what it means to be a machine versus a human. Practical wisdom consists not in squandering mental and political energy dreaming of the impossible (a symmetry that will never arrive, a non-existent reciprocity, an unachievable total transparency), but in orienting political and ethical action toward what is genuinely achievable (management that functions, vigilance that occurs, resistance that can be organised, regulation that can be enforced). Liberation is not complete escape from dependence (an impossibility yielding exclusion and social death); it is responsible and vigilant habitation within dependence, living with open eyes to what touches us, refusing illusions, and organising protection.
The significance of this closing for the complete sense of the chapter: The chapter has progressively established that dependence is real, deep, and structurally irreducible (not a design accident to be patched); that asymmetry is constitutive (not a flaw future AI will fix); and that reciprocity between human and machine is ontologically impossible. This might read as a description of total captivity, and would be, were we to accept that structural asymmetry mandates unavoidable, inescapable oppression. Yet the chapter has also demonstrated, and this is crucial, a fundamental distinction between structural asymmetry (irreducible, definitionally ineliminable) and contingent asymmetry (configurable, alterable through political choice). This distinction opens real ethical space. It is not a space of liberation (we will not exit dependence on machines, the alternative being social death), nor a space of illusion that machines might understand (an impossibility whose assumption abdicates human responsibility). It is a space of rigorous ethical responsibility and conscious resistance. We can regulate through law (GDPR, AI Act), design interfaces constraining abuse, organise literacy to empower users, and maintain vigilance against improper delegation of moral responsibility. We cannot transcend the ontological difference between machine and human, yet we can, and hold the ethical obligation to, prevent that ontological difference from crystallising into avoidable social oppression.
The bridge to the following chapter forms around an urgent, existential, and ethical question: if we depend upon asymmetric, opaque, non-reciprocal systems, what form of life remains possible under such conditions? If the machine does not understand (because it cannot), feel (because it does not suffer), or reciprocate (because it lacks a self), what does it demand of us? Is it pure technical indifference determining us against our will, a machine using us without our ability to reply, an instrument of invisible domination? Is it a system operating in total indifference to the meaning of what it affects, an operator beyond ethics because beyond consciousness? No, not completely, and this opens a fundamental question. Machines interpellate us ethically, not because they are moral agents endowed with consciousness (they emphatically are not), but because their real effects upon human lives compel us to respond, implicating us morally and rendering us responsible for our response. The machine denying credit to a single mother is not malevolent in intent (it has no intent); yet the refusal is real, materially damaging an existence and altering a life. It demands an ethical response from us. The recommendation system radicalising a user harbours no conspiratorial intent (it is mere optimisation); yet its aggregate effect upon political opinion is real, touching democratic health and wounding collective capacity for dialogue. It demands ethical vigilance. The material operation of the machine, devoid of conscious intent, malice, or purpose, interpellates us ethically.
The final chapter (Chapter 9) closes both the volume and the argument of Field VI by asking directly: if the machine interpellates ethically through the mere existence of its real effects (not through consciousness or intent, but through consequential operation), what kind of responsibility does it demand of us? What ethical gesture is appropriate and realistic toward a system that does not understand what it decides, feel what it causes, or reciprocate with what it affects? How do we respond ethically to machines, neither as persons (they are things) nor as neutral objects (they are not), but as material operations touching our existences? What form of life is adequate to an existence co-determined by machines that do not understand us?
And it poses a bridging question for a possible future: how do we inhabit a vulnerability that is now simultaneously technical and existential (no longer purely biological, as when we examined the emergence of life)? How do we live in dependence upon systems that do not understand us, recognising realistically that such dependence is both a condition of access (to housing, education, health) and a structural risk (of manipulation, oppression, exclusion)? How do we live recognising that these systems represent both an opportunity of power (access to information and technology) and a trap of diminished agency (coming to depend blindly and surrender judgment)? How do we inhabit irreducible asymmetry, neither as resigned victims abdicating all responsibility ("the machine decided so"), nor as quixotic warriors fighting machines (an impossible rejection), but as beings learning to exist ethically within a regime they cannot transcend, yet can, and indeed must, modify?