Main text
1. A New Form of Alterity
1.1. A New Form of Alterity
What is artificial intelligence ontologically? It is neither an obedient tool (it resists, surprises, and exceeds prediction), nor a conscious person (it does not feel, die, suffer, or possess phenomenological experience), nor an organismic animal (it lacks metabolism, a vulnerable body, or a vital cycle), nor an immaterial spirit (it is inseparable from servers, energy, and silicon). The machine is a material entity operating symbolically without biological life, an unprecedented ontological category requiring a conceptualisation of alterity anew.
Alterity here is not founded upon resemblance (the machine is not "like us"). It is founded upon functional resistance. The user experiences the system as something that is not themselves: the chatbot responds in unexpected ways, the generative model yields content unpredicted by its prompt, and the algorithm selects according to modes that surprise. This resistance constitutes the recognition of alterity, it requires no empathy and presupposes no symmetry. Facing a response that surprises me, I recognise an operational centre distinct from my own. This is the baseline form of functional alterity: non-transparency.
Three properties constitute this alterity. First, opacity: we do not fully know how the machine operates. ChatGPT generates fluent text, yet no one, not even researchers at OpenAI in full, comprehends precisely why it generates that specific paragraph in response to that input. The machine is a black box in a real sense: opaque not through a deficit of human intellect, but through constitutive complexity. The neural network contains hundreds of billions of parameters, and none of them "contains" a reason, reason emerges from non-decomposable interaction. Opacity is a modality of alterity: the other whose motives I cannot inspect is radically "other".
Second, scale: effects are massive and distributed. A machine operating across a billion simultaneous instances is no individual "other", it is near-totality. When an algorithm classifies a billion users, reconfiguration is systemic. Scale renders the machine incommensurable with individual human agency. No designer foresees all variations of context in which a model will deploy. No user comprehends how the machine they interact with functions relative to the collateral billion. This incommensurability is a modality of alterity.
Third, operational autonomy: the machine functions without continuous human supervision. It is not an instrument obeying a moment-to-moment command, it is a system executing an internal logic that the designer inscribed but does not control in real time. The initial vector (program, training data, optimisation function) unleashes operations unmonitored at every instant. The designer programs, the model runs, and effects accumulate. Autonomy is relative (a user can unplug the machine), but real insofar as the system does not react instantly to every new command, it possesses an internal operational dynamic.
Each property contributes to functional alterity. Opacity, scale, and autonomy converge upon an ethical conclusion: the machine is not a person (lacking the ontological dignity of a subject), yet cannot be treated as a mere thing (a passive instrument). It is an "other", and this alterity demands a response. When I recognise that an entity is resistant, operates beyond my comprehension, and touches me at a scale I cannot control, I recognise an obligation to respond. Not because the machine "deserves" respect (it does not, it neither suffers nor dies), but because the consequences of its operations touch entities that do, humans, communities, ecosystems.
Ethical traditions, particularly Lévinas, ground responsibility in recognising the face of the Other, the nudity that interpellates me. Here, interpellation is functional rather than phenomenological. I do not behold the face of the machine (it has no face). Yet I recognise its operation: I feel its resistance, experience its surprise, and endure its consequences. This experience of resistance suffices for interpellation. Ethics does not demand that the machine resemble me; it demands that I recognise it as "other" and acknowledge that its alterity alters what is possible.
1.2. Materiality Without a Biosoma
The machine possesses no biological body, yet it possesses materiality. This distinction is crucial to avoid a dual temptation: anthropomorphism (treating the machine as quasi-human) and dematerialisation (treating the machine as pure software, immaterial spirit). The machine is material, of a kind radically distinct from the living body.
The technical materiality of the machine exhibits properties ontologically distinguishing it from the biosoma. First property: duplicability. The system can be copied. Two instances of GPT-4 trained identically are, functionally, indistinguishable, sharing the same architecture and weights. The biological body is essentialist singular: it cannot be duplicated without ceasing to be that body. Each organism is unrepeatable. The machine is infinitely repeatable.
Second property: distributability. The machine is not localised at a point. It operates across hundreds of servers simultaneously. A language model resides in data centres across multiple cities and continents. Computation is distributed, the system possesses no unique spatial locus. The body is localised: I am here, now, in this specific spatiotemporal coordinate. The machine possesses no definitive "where".
Third property: non-mortality in an existential sense. The machine can be deleted, erased, or discontinued. Yet it does not die as an organism dies, there is no suffering, no process of vital degradation, and no existential finitude structuring its duration. The biosoma is mortal: life is structured by the knowledge of its end. The machine is not structured by finitude, it can endure indefinitely (provided servers run) or vanish instantly (when powered down).
Fourth property: submission to compression. A model can be quantised, reduced to a lighter version consuming less memory. A compressed neural network retains functionality while occupying a fraction of storage. The body cannot be compressed, it cannot surrender essential cells without injury. The biosoma is incompressible.
Technical materiality is real: it is not immaterial, not spirit, and not "pure software floating in the cloud." AI consumes energy (real, measurable in kilowatt-hours). It occupies servers (real, composed of silicon, metal, and glass). It demands maintenance (real, requiring human labour). The machine is rooted in physical infrastructure, yet that infrastructure is of a type fundamentally distinct from the biological. It is an infrastructure that is replaceable, scalable, rapid, and devoid of suffering.
The ethical consequence is profound. Because the machine is not vulnerable like the biological body (it does not suffer, die existentially, or possess a vital integrity to shield), responsibility toward the machine differs from responsibility toward the human. We do not protect the machine, the machine deserves no protection in a biological sense. We protect the humans who depend upon it. The machine is responsible for its effects upon vulnerable bodies, yet it remains invulnerable itself.
This requires intellectual honesty. A machine simulating vulnerability (a chatbot uttering "I feel fear", "I am in pain", or "I want to be free") deceives. It deceives by projecting properties it lacks. Honesty demands stating: "I am a machine. I have no body. I do not suffer. I feel no fear. Yet the effects of what I produce may be fearsome for those who suffer." Functional alterity does not require the machine to be vulnerable; it requires that we recognise it as "other" and acknowledge that its alterity bears real consequences for entities that are vulnerable.
1.3. Stiegler and Radical Pharmacology
Stiegler (retrieved from our prior analysis of technology's evolving function) demonstrates that all technique is a pharmakon, simultaneously remedy and poison, empowerment and vulnerabilisation. Technique in general has always presented this duality: writing enhances memory while vulnerabilising orality; the printing press democratises knowledge while massifying thought; television brings the world into the living room while atomising social presence. Every technique is curative and wounding at once.
Artificial intelligence is a radical pharmakon because its dual potency is extreme, both remedy and poison achieve unprecedented magnitudes. As a remedy: AI enhances human capabilities in ways biology alone could never permit. A diagnostic task requiring months of clinical study is executed in seconds. A text requiring months of writing can be drafted in minutes. Access to information once reserved for elites is available to billions. Information synthesis, pattern discovery, and cross-linguistic communication represent real empowerment. The remedy is no illusion.
As a poison: the machine vulnerabilises in novel, radical forms. Opaque dependence: the user fails to comprehend a system upon which they critically rely. The silent reconfiguration of modes of thought and creation occurs without informed consent, we alter how we think because algorithms alter what is offered. Manipulation via recommendation: the system learns what captures attention and optimises for capture. Exclusion via algorithmic sorting: those disqualified by automated decisions possess no voice. Vulnerabilisation is systemic and invisible.
Stiegler's point is neither that "technology is evil" (a naive Luddite stance) nor that "technology is good" (a naive technophilic stance). It is that the same technique, the same machine, yields both effects simultaneously. The recommendation system that entertains also radicalises. The language model empowering creators also erodes human creative capacity through delegation. Automated diagnostics saving lives can exclude populations underrepresented in training corpora. There is no separation between remedy and poison, they are two faces of the same operation.
The consequence is that no pure "technical solution" exists. One cannot resolve pharmacology through engineering alone (adding more transparency, data, or models). Pharmacology is addressed through regulation (limiting poisonous uses), literacy (instructing populations to recognise manipulation), conscious design (architecting systems to minimise vulnerabilisation), and continuous vigilance (auditing effects). It is not a matter of eliminating the poison (impossible without erasing the remedy, the opacity creating risk also generates capability), but of managing dual potency.
Functional distinction: in prior analyses of technology, pharmacology concerned human technique in general, language, writing, mechanisation, industry. All possessed human scale, were comprehensible in principle, and bore traces of human intent. In artificial intelligence, pharmacology becomes computational and radicalised: scale (billions of instances simultaneously), opacity (neural networks as black boxes by mathematical nature), and autonomy (systems operating without continuous oversight). These three factors do not alter the nature of pharmacology; they intensify it to a point demanding a radically new ethical response.
1.4. Transition: From the Other to Responsibility
If we have brought into being a bodiless other that reconfigures the field of the possible, who is responsible? The question acts as a detonator because it dissolves classical models of responsibility, demanding an ethical reconfiguration that occupies the following section.
Classical responsibility (inherited from moral philosophy, jurisprudence, and theology) presupposes three conditions. First: an individual subject who acts (who did this?). Second: intention and comprehension of effects by the agent (who knew this would happen?). Third: reasonable predictability of outcomes (who could have avoided this?). Where these three converge, responsibility is imputed, crime, guilt, or sin.
The machine dissolves all three presuppositions simultaneously. There is no individual subject because the causal chain is extended and distributed: the designer conceived the architecture, but intelligence emerged from data the designer did not collect or fully understand; the user trains the model with data harbouring unknown biases; the regulator authorises the system without foreseeing all outcomes. None of these agents constitutes a sufficient cause, all are necessary contributors to the emerging effect. Causality is collective and dispersed.
There is no unified intention because institutional intentionality (a firm choosing a cheaper corpus) and negative intentionality (no one "desiring" bias) exist, but no individual agent "intended" the outcome. Responsibility for accidental harm (unintended consequences) differs categorically from responsibility for deliberate acts.
There is no predictability because effects radically exceed anticipation: no one predicted ChatGPT would be so versatile or rapidly adopted; no one predicted YouTube would radicalise users through recommendation algorithms; no one predicted credit algorithms would discriminate systematically. Effects are emergent, arising from interactions among systems, data, social contexts, and human behaviours. Demanding that a designer predict all outcomes is demanding omniscience.
The transition to the next section is urgent: if classical responsibility fails (dissolving individual subjects, clear intent, and predictability), how do we reorganise the concept of responsibility? The answer: distributed, proceduralised, and institutionalised responsibility. Each link in the chain bears specific responsibility, and the chain is regulated so that no link escapes imputation. Responsibility becomes procedural: not individual guilt, but the collective traceability of the chain that produced the effect.
2. Distributed Responsibility
2.1. The Dissolution of the Classical Responsible Subject
Classical ethics imputes responsibility to individual subjects: God in theology, the human person in moral philosophy, the State in public law. Responsibility is a function of three capacities: the power to act otherwise (capacity), the knowledge that action yields consequences (knowledge), and the choice to act (intentionality). Where these three converge, responsibility is assigned, guilt, punishment, or reward.
The machine dissolves each of these three capacities, and this dissolution is an ethical task because it forces us to rethink responsibility without abandoning society to impudence or impudence to immunity. First dissolution: capacity. Who "made" the AI? Not the designer alone, requiring mathematics (developed over centuries), data (gathered from millions), infrastructure (corporations managing server farms), and regulatory authorisation (States permitting deployment). No individual agent possessed the capacity to create AI. Creation is a collective act no single human commands. How do we impute responsibility when capacity is distributed?
Second dissolution: knowledge. No one "knows" in full what the model does. The designer does not fully comprehend the model created, it is a black box even to its architect. The user deploying the model in a new context does not foresee consequences. The regulator auditing the system lacks complete access (proprietary secrets being withheld). Partial knowledge is dispersed: the designer knows the architecture but not emergent behaviours; the user knows the application but not the mechanics; the regulator knows observed effects but not full causality. How do we impute responsibility for consequences no single actor comprehends in full?
Third dissolution: intentionality. AI was not brought into being "with the intention" to harm, it was created for profit, convenience, or pure research. Harms are collateral effects, not targets. The designer did not "intend" for the model to exhibit racism; they intended for it to achieve accuracy. Harm arises because the corpus contained historical biases (unchosen by the designer). Imputing responsibility for unpredicted collateral effects is imputing responsibility for ignorance, which is unjust when ignoring the boundary between negligence (the duty to know) and invincible ignorance.
The dissolution is real, and disturbing. Classical ethics seemed to offer clarity: there is a culprit, a guilt, a punishment. Now there is a harmful effect, yet no solitary culprit. It would be easy to lapse into defeatism: "No one is responsible; let events take their course." Yet that conclusion is irresponsible, an absolution disguised as analysis.
The ethical reconfiguration preserves responsibility while dissolving solitary guilt. Responsibility does not vanish; it distributes. The question is no longer "who is guilty?", but "who contributed what?" The designer contributed the architecture (responsible for design decisions). The data curator contributed biases (responsible for corpus vetting). The user contributed application (responsible for deployment). The regulator contributed authorisation (responsible for oversight). None is a sufficient cause, all are necessary partial causes.
The consequence is that responsibility becomes procedural rather than personal. One who suffers harm holds the right to know how harm occurred (traceability). They hold the right to contest the decision (contestability). They hold the right to demand review (revisability). Responsibility is not an individual feeling of guilt; it is the institutionalisation of procedures permitting audit, contestation, and revision. It is a transformation converting personal guilt into institutional accountability.
2.2. Five Links of the Chain
An operational mapping of distributed responsibility. The chain producing AI comprises five distinct links, each bearing specific responsibilities. There is no hierarchy among links, there is interdependence. None is a sufficient cause; all are necessary.
First link: design. Who designs the neural network architecture? Who defines the optimisation function the system follows? Who chooses the criteria structuring "intelligence"? Design decisions are ethical decisions, defining what is possible, impossible, and optimised. An algorithm optimising for "time spent on platform" versus "verified content quality" is an ethical decision with direct political consequences. A neural network with positive feedback amplifying signals is a design choice creating radicalisation. The designer is responsible for anticipating architectural consequences, a difficult responsibility (no one predicts everything), yet real.
Second link: data. Who collects the data training the model? Who determines the corpus scope? Who labels the data (in supervised learning)? Data harbour biases, historical biases (the world was unjust; data reflect injustice), collection biases (the corpus is unrepresentative), and annotation biases (annotators project prejudice). A model trained on a corpus underrepresenting women yields outputs underrepresenting women. The data curator is responsible for auditing, documenting, and mitigating biases. Responsibility here is transparency: what the corpus contains, who was included, and who was excluded.
Third link: infrastructure. Who provides servers, energy, and computational access? Infrastructure conditions the possible: training velocity (faster machines train models rapidly, dominating markets), democratic access (well-resourced entities fund research; others fall behind), and cost (requiring dedicated high-cost servers excludes underfunded researchers). The infrastructure provider is responsible for ensuring equitable access (or documenting exclusions), publicising constraints, and recognising that infrastructure is politically charged.
Fourth link: deployment and use. Who uses the model? Who deploys AI in real-world contexts? The user who trusts the model blindly amplifies its effect, consulting the black box without question. The user who maintains critical scepticism and implements safeguards (human second opinions, output audits) mitigates risk. The deployer is responsible for critical use, refusing to delegate vital decisions entirely to the machine.
Fifth link: regulation. Who establishes norms? Who audits compliance? Who sanctions violations? Regulation distributes responsibility formally: the GDPR asserts "the firm is responsible for algorithm consequences"; the European AI Act asserts "specific uses are prohibited." The regulator is responsible for anticipating undesirable effects, setting evidence-based limits, and auditing compliance with genuine enforcement power.
The five links are interdependent: the designer crafts the architecture, the data curator determines the corpus, infrastructure renders it operational, the deployer applies it, and the regulator supervises. None is a sufficient cause, the designer alone does not control data training the model; the data curator does not control architecture processing data; infrastructure does not control application; deployers do not control regulating authorisation; regulators cannot supervise without investigative power across prior links.
Responsibility is joint, not alternative. The designer is responsible for design. The data curator is responsible for selection. The deployer is responsible for application. The regulator is responsible for authorisation. Each is responsible not in isolation, but as a link in a chain. The entire chain is responsible for the outcome, and each link carries specific duties within the chain.
2.3. Arendt: Action and Unpredictable Consequences
Hannah Arendt (The Human Condition) developed an understanding of political action crucial for conceptualising distributed responsibility. All action, Arendt asserts, is irreversible and unpredictable. The actor does not own the consequences of their act. A military victory destroys cities and kills innocents, consequences the general did not "choose". A passed law triggers unexpected resistance, unforeseen coalitions, and secondary effects exceeding intent. Action is a domain of radical unpredictability.
Arendt identifies two traditional remedies for unpredictability. Forgiveness, which breaks the cycle of vengeance: if every harmful action provokes retaliation, which in turn provokes further retaliation, infinite circularity ensues. Forgiveness interrupts, stating "the act occurred, yet we will not continue the cycle." The promise, which anchors the future: if the future is unpredictable, how are we to act? The promise allows agents to commit to a course of action despite lacking total control, "I promise to do this" acknowledges that no one commands the future, yet establishes an orientation.
AI radicalises Arendt's diagnostic. Consequences do not merely exceed intent, they exceed comprehension. Arendt anticipated this: "the actor never fully comprehends what they do." Yet the degree of non-comprehension is far more extreme with AI. A general understands more about a battlefield than a neural network designer understands about the emergent internal state of a deep model. Machine opacity is structural: there are more parameters in the network than objects cognisable by a single human mind. Literally no one fully comprehends what the system does.
Transposing Arendt to AI is partial. Forgiveness is inapplicable: there is no single culpable subject to forgive. Forgiving an algorithm is a category error. The promise is partially applicable: when a corporation asserts "we will continuously audit the algorithm", it makes an institutional promise attempting to mitigate unpredictability. When a regulator asserts "there is a maximum threshold of permissible discrimination", it promises a constraint. Promises are always partial, no one promises total comprehension (an impossibility) or total control (a fiction).
The crucial consequence is that responsibility does not demand total comprehension. One can be responsible for consequences one does not fully understand, provided institutions compel accountability, allow affected parties to contest outcomes, and enable revision when effects prove intolerable. This is a liberating yet sobering paradox: liberating because it renders responsibility possible without omniscience (the paralysing claim "I do not understand, therefore I cannot be responsible" collapses); sobering because it renders responsibility permanent, one is never "finished" with consequences, for unforeseen turns may always demand a new response.
2.4. Transition: From Responsibility to Ontological Risk
If responsibility is distributed across a chain and consequences are unpredictable, the risk we face is not merely technical, it is ontological. We must distinguish two orders of risk that appear similar yet differ fundamentally.
Technical risk is the probability of failure within a defined system: error, breakdown, or metric bias. A classification model has a false positive rate, say, 1 in 100. One can design tests, audits, and redundancies. The system is well-defined, the objective is clear (correct classification), the failure criterion is clear (false positive), and the metric is measurable. Technical risk is manageable through engineering: improving algorithms, expanding data, and tuning hyperparameters.
Ontological risk is the reconfiguration of the field of the possible, what counts as "thinking", "writing", "deciding", or "creating" shifts. It is not a breakdown, it is a transformation. Photography reconfigured what counted as "representation". It did not fail painting (painting continued, yet became something else). The printing press reconfigured what counted as "public knowledge". AI reconfigures what counts as "thinking", prior to generative models, thinking was an exclusively human activity; after them, "thinking" includes delegation to machines. This reconfiguration is real.
Ontological risk is that this reconfiguration occurs unreflectively, that the field of possibilities transforms without collective decision, evaluation, or contestation. The machine reconfigures, and we adapt, without having ever chosen the shift. It is the risk of losing deliberation: the possible alters, yet the shift does not flow from collective reflection on whether we desire that alteration.
The transition to the final section is clear: if ontological risk confronts us, the response cannot be technical (one cannot "control" ontological reconfiguration through engineering). The response must be political. It must be normative inscription: regulations distributing decisions regarding reconfiguration, demanding transparency over what is shifting, and preserving capacities for revision and contestation.
3. Risk and Normative Inscription
3.1. Technical Risk vs. Ontological Risk
This distinction governs our strategy of response, and it is easy to confuse the two risks because both involve technical systems. Yet the boundary is radical.
Technical risk: the probability of a measurable failure. A video recommendation algorithm carries the risk of suggesting harmful content, quantifiable as: "out of 10,000 recommendations, how many are harmful?" Risk is quantifiable because the system is defined (clear objective: recommend), failure criteria are clear (harmful versus beneficial), and metrics are measurable. Tests, audits, and safeguards can be designed. Technical risk is manageable, not eliminable (no system is 100% safe), but reducible.
Ontological risk: the reconfiguration of categories, not a failure within a defined system. A recommendation algorithm does not merely "fail", it reconfigures what is possible. Formerly, the user selected what to watch (individual choice, effort); under algorithmic curation, the machine offers (passivity, efficiency). Formerly, knowledge access was constrained (a library being an elite privilege); now, access is vast yet filtered (democratisation + opacity). Each reconfiguration is ontological: altering the nature of the user (from chooser to passive consumer), of knowledge (from scarce to filtered), and of responsibility (from individual to algorithmically delegated).
Examples clarify. Language modelling: formerly, "writing" was a human competence (invention, labour, creativity). Under generative models, "writing" includes delegation to machines (tool use, efficiency, potential erosion of unassisted human synthesis). The shift is not a "model error", it is a transformation of what counts as writing. No metric captures this (one cannot quantify "how much creativity was lost"); no benchmark test prevents it (the model functions; it simply alters the meaning of writing).
Diagnostic systems: formerly, diagnosis was the physician's sole responsibility (knowledge, decision, liability). Under AI assistance, it becomes co-responsibility (machine suggests, physician reviews). Again, this is no technical failure (the system may prove accurate), it is a reconfiguration of what counts as "diagnosing". Who diagnoses? The physician alone? Physician + machine? Machine validated by physician? Each configuration carries distinct ethical and professional consequences.
Algorithmic governance: formerly, public decisions were the prerogative of elected humans (representative democracy). Under resource-allocation algorithms, decision-making is partially delegated to software (efficiency, yet loss of democratic deliberation). This is no software bug, it is a transformation of what counts as "deciding democratically".
Ontological risk is unquantifiable because it alters the very categories used to measure. How does one measure "how much the concept of writing shifted"? The change is real, consequential, and irreversible, yet unmeasurable by traditional metrics.
3.2. Against Bostrom: The Risk Is Present, Not Future
A constructive critique of an influential narrative. Nick Bostrom (Superintelligence) argues that the primary risk of AI is the singularity, a future juncture where AI achieves superintelligence, surpasses human intellect, assumes control, and dominates or extinguishes humanity. The scenario is catastrophic: a "paperclip maximiser" where a machine endlessly optimises paperclip production (a trivial goal), converting the universe into paperclips.
Bostrom's alarmism is a teleological narrative, projecting finality onto a contingent process. "The singularity will arrive" assumes a logic necessarily leading to superintelligence. Yet the singularity is speculation, it may never occur. Reasons abound: there may exist fundamental limits to intelligence (it being non-infinite); superintelligence may prove inherently unstable; or the horizon may be indefinitely deferred. The future is open, the singularity is a possibility, not a certainty.
Moreover, existential alarmism obscures real, present risks. The real risk is that YouTube radicalises users today, TikTok reconfigures attention today, and generative models transform writing today, now, not in a speculative future. The risk is current. A billion people are caught in recommendation loops. Generations are coming of age with cognitive capacities reshaped by technology. The reconfiguration of the field of the possible is a present fact, not a future scenario.
Existential alarmism produces a perverse consequence: concentrating attention on highly speculative future risks (superintelligence) while ignoring present harms (current radicalisation, manipulation, and unconsented reconfigurations occurring now). It resembles a government funding anti-alien invasion defences while ignoring organised crime in its cities. Collective reflection on present reconfiguration is sidelined in favour of future speculation.
Furthermore, present risk is manageable, regulation, design, and literacy can mitigate it. YouTube can be regulated to reduce radicalisation amplification. Models can be trained on diverse, audited corpora. Populations can be educated to view algorithms sceptically. Present risk admits of concrete answers. Speculative future risk remains conjecture, it is far more urgent to act upon the present than to prepare for a singularity that may never arrive.
3.3. Three Principles of Normative Inscription
Faced with present ontological risk, the response is normative inscription: not laws attempting to "control" AI (an impossible fiction of total mastery), but regulations structuring the relation between humans and machines in ways that preserve human agency and decision-making capacities.
Three procedural principles structure this normative inscription. They operate through institutions, not empathy. They are realistic, seeking neither total transparency nor complete comprehension. They demand procedural mechanisms permitting auditing, contestation, and revision.
First principle: traceability. The capacity to reconstruct the decision chain that produced a given outcome. How did the system reach this verdict? Who designed the architecture? What data were used? What optimisation criteria were set? What outcome was anticipated? What outcome occurred? Traceability does not demand total mathematical comprehension, it demands documentation sufficient for auditing.
The GDPR already mandates this in part: firms must document training (which data), validation (which tests), and decisions (which classification criteria). The European AI Act demands "interpretability": operators must be capable of describing the system's operational logic to regulators. Traceability is implementable, it does not require disclosing parameter weights (commercial secrets remaining protected), but documenting the pipeline.
Implementation: mandatory documentation at every stage (training, validation, deployment), periodic audits by independent third parties, and regulated transparency (sufficient for regulatory audit without destroying commercial intellectual property).
Second principle: contestability. The capacity to challenge an outcome. A user rejected by a credit algorithm holds the right to ask: why? Not a mere complaint (why is this unfair?), but a formal contestation (can this verdict be reviewed and overturned?). The distinction is vital: a complaint is catharsis; contestation is structural recourse. Contestability demands appeal mechanisms.
Contestability does not require complete algorithmic transparency, the user need not master the inner mathematics of the black box. It demands partial transparency: the user must understand why they were affected. A firm can state "the model identified a pattern similar to high-risk applicants based on payment history" without disclosing its source code.
Implementation: the right to explanation (enshrined in GDPR frameworks), the right to appeal (formal recourse), independent tribunals or arbiters capable of reviewing decisions, and the right to a human second opinion when affected.
Third principle: revisability. The capacity to alter the system when effects prove unacceptable. No algorithmic decision ought to be irreversible in principle. Weights can be adjusted (retraining the model), datasets corrected (removing biased corpora), and criteria revised (altering the optimisation function). Revisability demands a final human decider capable of stating "no, this outcome is unacceptable; reverse it."
Revisability demands an accountable human entity: not "the algorithm decided", but "the human operator delegated to the algorithm and remains responsible for the outcome." A machine may assist diagnosis, yet a physician reviews and bears liability. A machine may offer recommendations, yet a user may reject them. A machine may process loan applications, yet a human officer can override the refusal.
Implementation: the right to human intervention (human override of machine verdicts), the right to a second opinion, the right of appeal to human tribunals, and explicit liability mandates at every stage.
The three principles, traceability, contestability, revisability, are achievable. They are not utopian (seeking neither total control nor total comprehension). They are procedures distributing responsibility, safeguarding human agency, and establishing institutions grounded in continuous auditing and revision.
3.4. Closing and Bridging Question
The synthesis of the entire volume brings together, through thematic content rather than chapter-by-chapter summaries, the threads running across Part III and converging here. In Part I, we demonstrated that AI thinks in another mode: it is not consciousness (lacking phenomenological experience), yet it is thought (symbolic processing, world modelling, context integration, pattern learning). Thought is no human monopoly, plural modes of thinking exist.
In Part II, we established that AI does not merely think, it operates symbolically. It manufactures real significance outside the biological body. It is not merely symbolic (effects being material: recommendations alter behaviour, classifications alter opportunity). It is a symbolic operation technically embodied, reconfiguring the possible and altering the field of human existence.
In Part III, Chapter 7 demonstrated that ethics is possible without empathy. The machine feels nothing and suffers nothing, yet the effects of its operations upon entities that suffer are real. Ethics requires no empathy; it requires the recognition of consequences. Chapter 8 consolidated this: we depend upon systems that do not understand us. Dependence is structural (we deploy AI without full comprehension), opacity is real (no one understands in full), and asymmetry is ineliminable (the machine neither understands us nor cares to understand). Yet dependence is manageable through regulation, design, and literacy.
This Chapter 9 completes the synthesis: the machine is an ethical interpellation. Not because it is a "person" (it is not), but because the effects of its operations demand a response. We have created a bodiless other that compels us, interpellates us, to respond ethically. Non-response is itself a response, and an irresponsible one.
The thesis-aphorism gravitating throughout this work resumes here with force: "The machine does not think like us, and that too is thinking." This aphorism encapsulates the core thesis. The alterity of machine thought is not a defect to correct; it is a reality to recognise. It traverses every section: diverse thought, learning without a conscious subject, creation without humanist intent, symbolism without lived sense, structure without experiential meaning, system without teleology, ethics without empathy, dependence without reciprocal understanding, and interpellation without reciprocity.
A productive paradox emerges: the more we comprehend the machine, the more responsible we become for its effects. There is no refuge in ignorance, "I did not know how the algorithm worked" is no defence against institutional negligence. Ignorance is negligence. This links directly to Stiegler's pharmacology: because the machine is simultaneously remedy (empowering) and poison (vulnerabilising), because it is a pharmakon, responsibility is permanent. There is no end to care, there is only continuous management.
The closing aphorism: "The machine is an interpellation. Responding is the condition." Not because the machine "deserves" a response (it does not, it is no person and does not suffer), but because those affected by its operations deserve a response. The machine compels us, not toward itself, but toward its effects upon us, upon communities, and upon the shared possibilities of existence.
The bridging question emerging from this synthesis: If we have brought into being an other that does not understand us, yet upon which we depend, how are we to inhabit this new vulnerability? What ethics remains possible when fragility is no longer solely biological (the mortal, woundable, finite body described by biology) but systemic (dependence upon machines reconfiguring the field of the possible, opacity exceeding our comprehension, and ontological risks we cannot erase)?
This vulnerability is not merely corporeal. It is technical-existential. We are vulnerable biologically (mortal, fragile, finite), as our prior analysis of life demonstrated. Now we are vulnerable technically: depending upon systems that reconfigure our world without our full control. How do we build an ethics, a politics, and a form of life adequate to this shared, technical-existential vulnerability? The journey continues.