Technē without poiesis:

rethinking craft beyond human

I treat craft’s waning as an entropic drift: a non-reversible tendency in computationally reticulated milieus (AI and beyond). Once poiesis is re-encoded as statistics and served from aligned infrastructures, hysteresis locks timings, interfaces, and value extraction. What re-enters is post-selection, not authorial negotiation. I first define a minimal vocabulary—H, R, and OL—to name how duration is lost when making is encoded into substrates.

Definition — Time-at-Risk and Ontological Lag (OL).
Let H denote histories of making with time-at-risk (formative, path-dependent duration). Let R denote a retentional substrate (weights, protocols, standards, market rules). Let E: H → R be the encoding of H into R. This transfer is lossy for duration: there is no feasible Ψ: R → H that recovers the time that did the work without doing it again.

Ontological Lag (OL) names that non-invertible remainder of duration at the H↔R interface. OL ≠ latency, ≠ noise, ≠ missing data: latency can be budgeted, noise filtered, data interpolated; OL is lost duration—a qualitative gap no amount of data can fill.

What “re-risking time” means

Re-risking time = re-entering a formative interval whose outcome isn’t fixed ex ante, so that the time-dependencies of making (order, pacing, material coupling, real hazard) are instantiated again, not merely replayed.

Why this is required in principle:

Because R (weights, logs, specs) doesn’t carry the causal micro-temporal dependencies that did the work. To get those dependencies back you must instantiate a new history H′—a fresh negotiation in time. Simulation, enforced delay, or stylistic smoothing recover the look, not the duration.

Method note — H/R/OL as structural/operational ontology.

H/R/OL is a structural/operational ontology, not a grand metaphysical thesis. It is a substrate-agnostic analytic lens for how any encoding E ⁣: ⁣H→R strips temporal thickness (“duration”) from lived making. The claim is diagnostic, not speculative: whenever a lived history HHH is captured into a retentional substrate RRR, some duration is non-invertibly lost (OL). This holds irrespective of cosmological commitments; it describes a structural fact about mediation, not a metaphysics.

Formatività as Time-at-Risk

Philosophies of making framed as Formatività treat form-giving as a temporally extended negotiation with matter—time at risk in which responsibility and material resistance co-determine outcome (Ribault, 2022). My claim is not that this formative interval was illusory, but that in the AI stack it is pre-posed in data/manifolds and post-posed as curation. Thus, the Formatività horizon clarifies what AI subtracts: not skill as such, but the pacing in which skill becomes world-disclosing. OL names what that subtraction cannot return: a duration remainder at H↔R that no retentional substrate can invert. What remains is not poiesis, but residual duration—OL at the H→R interface.

If Formatività names the formative interval, Tanizaki’s “shadow” names a practice of withholding that interval; today’s AI exposure pipelines flatten both. OL is precisely that non-invertible remainder of duration at H↔R. It is not latency, noise, or missing data; it is duration that cannot be reconstructed from the substrate. Thus, the Formatività horizon clarifies what the substrate subtracts: not skill, but the pacing in which skill becomes world-disclosing—and OL measures what that subtraction cannot return.

Across the history of making, OL’s location and convertibility shift, not its existence. Read with H (histories of making as time-at-risk) and R (the retentional substrate), the seam H↔R migrates as follows:

Phase I — Pre-industrial workshop (technē with poiesis).
R: tools, rules of thumb, guild memory. H stays close to matter; feedback loops are immediate. OL remains near the hand and is often redeemable—folded back into H during apprenticeship, repair, and patination. Poiesis is local; the remainder is spent inside making.

Phase II — Industrial factory (the performance turn).
R: standards, jigs, machine setup, quality-control protocols, shift cadence. R thickens and externalises competence; H is segmented across tasks. OL relocates to setup, calibration, and exceptions; some poiesis persists in on-line adjustment and line deviation. What looks like preserved craft is largely performance choice riding atop an increasingly externalised R.

Phase III — Artificial-intelligence computing (statistical/alignment phase).
R: weights, datasets, prompts, inference graphs, protocols. Competence migrates upstream into R; downstream, H reappears mostly as post-selection. In a monotechnic recursion, variance collapses into parameters; restoring formative dependencies requires doing the time again—instantiating a new H′. OL becomes a detached surplus: legible as seam or glitch, not reinjectable into competence without re-risking time. Smoothing is cheap for the model; recovering duration is not.

Tanizaki’s “shadow” is not an aesthetic palette but a timing rule, a refusal to saturate appearance (Tanizaki, 1977/1933). AI’s recursions reverse that rule; pipelines treat underexposure, grain, and seam as error—producing parametric chiaroscuro while maximising recall and leaving OL intact: the non-invertible remainder of time-at-risk. Once shadow is parameterized, the temporal reservation that grounded authorial negotiation migrates upstream; this is the sense in which authorship collapses into infrastructural pre-composition.

If authorship (here) names the locus of kairotic negotiation (H coupled to matter under hazard), the curator model confines the human downstream to selection over a pre-composed retentional field R—a shift from shaping histories (H) to sampling substrates (R).

Within the current generative-AI milieu—where design emerges from recursive pattern recognition rather than intentional form-making—authorship has become structurally obsolete. The figure of the architect as author—as synthesiser of form, meaning, and intention—has dissolved into the procedural logics of alignment, reinforcement learning, and post-curation (Wen, 2024b). Design propositions are generated upstream, in data manifolds and model architectures, before any human enters the frame. The human role, at best, is reduced to that of curator—a post-selection agent, a symbolic validator of outputs already shaped by infrastructures of computation.

Definition (Authorship): Authorship (here) = the locus of kairotic negotiation, not merely selection among pre-composed options. In this sense, “co-making” names parametric sufficiency, not poietic negotiation. On this reading, curation is a restricted sub-case of authorship that operates within R (post-selection), whereas authorship-proper spans the H↔R seam and entails re-risking time to constitute a new H′.

In this curator model, embodied negotiation with material—formerly central to craft—arrives only after form has congealed in data. This reversal is not neutral; it shifts value production from situated making to infrastructure maintenance, inaugurating a politics of databases over bodies and relocating OL upstream (from situated making H to retentional substrates R). Hui’s account of digital individuation clarifies why poiesis migrates upstream: once form arises as a reticulated milieu of models, protocols, and data repositories—what the human encounters downstream is already a stabilised phase of infrastructural becoming (Hui, 2016a, 2019, 2021).

Within the alignment apparatus as currently configured, disclosure is pre-posed upstream as tertiary retentions in models and protocols (Stiegler, 1998). The hand returns downstream as post-selection—curation after the fact (Wen, 2024b). To the substrate, craft becomes commensurable feature-data under technical codes that naturalise timing defaults as neutrality (Feenberg, 1999). Gestures are scraped, trained, reproduced—perfectly indexed without presence (Baudrillard, 1994). Retrieval persists; duration does not (Bergson, 1911). Hui names a horizon where co-creation remains carbon poiesis—form arising through human–technical relation rather than an isolated hand (Hui, 2016b). To mark how that horizon is presently deformed, I borrow Chomsky’s distinction: competence as generative licensing, performance as graded stream (Chomsky, 1965). Today, competence is largely silicon-side—licensed in models and protocols and, crucially, often executed there as well (Stiegler, 1998). Carbon re-enters as performance after the fact—post-selection, caretaking, symbolic validation (Wen, 2024b). What passes for technodiversity are intra-silicon parameter turns within a single licensing configuration: monotechnics-in-practice (descriptive, and predicted by the H ⁣→R duration-loss invariance), with cosmotechnics retained as rhetoric in this paper’s frame. In Heidegger’s register, this rearrangement yields technē without poiesis downstream—capability shorn of a risked formative interval (Heidegger, 1977).

Temporal grammar: Chronos and Kairos.

This split also clarifies the temporal grammar. Chronos is prediction/control time; Kairos is the cultivated interval of openness where order, pacing, and hazard cannot be pre-timed (Trogemann, 2025). In the present stack, silicon competence runs on Chronos; carbon performance tends Chronos as curation. Poiesis, however, indexes to Kairos. Hence the limit-case: a silicon system could become poietic only where it re-risks time—open-ended, path-dependent formation under real hazard with order/pacing sensitivity. Strange-loop recursion can furnish self-representation on the silicon side, a necessary precondition, but without timed coupling to hazard it remains representational, not Kairotic (Hofstadter, 2007). Speed-minus-one is a kairotic damping that preserves OL without claiming to restore poiesis. Custodial craft (ζ term clarified): Speed-minus-one is a damping tactic: introduce a small ζ > 0 drag into production so duration remains audible without changing the attractor. It preserves OL as a kept interval rather than pretending to restore pre-AI poiesis (→ Custodial craft / ζ).

In this configuration, OL is structural: for any encoding E: H → R, OL names the non-invertible remainder of duration at the H↔R interface, substrate-invariant across carbon (human) and silicon (AI). Even if a silicon stack were to satisfy Kairos—re-risking time under hazard—once experience is deposited into retentional supports, every H→R transfer is lossy for duration and the remainder relocates with the seam.

Cosmotechnics (bracketed). Cosmotechnical plurality may be invoked rhetorically as a horizon of possible ontologies; in this paper’s H/R/OL frame—a structural/operational ontology, not a grand metaphysical thesis—it is bracketed, because the claim that any H ⁣→ R encoding is lossy for duration does not depend on plural cosmologies. “Monotechnics” is used here descriptively for the actually existing pipeline (cf. Hui, 2016b).

Alignment functions as a timing policy, compressing heterogeneous durations into commensurable latencies (Feenberg, 1999). Even where a silicon stack were to satisfy Kairos—re-risking time under hazard—OL would not merely persist; it would be necessary: once experience is deposited into retentional supports, any E:H→RE: H \to RE:H→R is lossy for duration and the remainder relocates with the seam (Stiegler, 1998). Hence the present diagnosis—technē without poiesis—and the accompanying claim: OL is the meter that keeps time across substrates, future-proof even under operational strange loops.

Relay: Simondon → Marx → Heidegger → Serres

Each of these four frames an unassimilable remainder—technical, economic, ontological, parasitic—that the statistical stack radicalizes temporally as OL. Simondon identifies a technical residue that individuation cannot absorb; Marx politicises this residue as dead labour—life sedimented into fixed capital; Heidegger ontologises both as Bestand (standing-reserve), once abstraction totalises availability; Serres then hears the leftover as parasitic noise—present, enabling, unassimilable (Simondon, 2017/1958; Marx, 1973; Heidegger, 1977/1954; Serres, 2007/1980). Function in argument: Simondon names the residue technically; Marx names its political economy; Heidegger names its ontological mode of revealing; Serres names its structural remainder.

In Simondon’s terms, each technical ensemble individuates by internalising earlier techniques yet always leaves behind disparate, unassimilable traces of prior modes of doing (Simondon, 2017/1958). Marx, encountering this residue historically in the industrial machine, names it dead labour: the living activity of the worker sedimented into fixed capital, functioning without the worker’s continued presence (Marx, 1973). Read together, Simondon’s technical leftover and Marx’s objectified labour expose how each wave of technicity builds by hollowing out embodied life—Simondon tracks this in recursive layers of technical codes; Marx frames it as the vampiric logic of capital. Heidegger turns both toward ontology: once abstraction wins, beings endure as Bestand—timed for availability (Heidegger, 1977/1954). Generative AI radicalises all three at once.

Scope and consequence. OL is narrower in scope—naming the remainder specifically as duration at the H↔R seam—yet its consequence is broader: it reframes AI not as rupture but as history’s fulfillment, the phase in which poiesis (formative bringing-forth) is effectively exhausted downstream and only technē-as-residue endures. In short: across technical, economic, ontological, and informational registers, each thinker isolates a structural remainder. My claim names it temporally: every H→R (history → retentional substrate) transfer leaves non-recoverable duration—time that did the work yet cannot be re-extracted from R without doing it again.

Third Inflection: The Statistical Stack

Generative AI does not merely store craft in machinery; it entangles it in statistical weights, dispersing poiesis into latent tensors. Statistical embedding secures pattern recall yet dissolves the situated gesture that once grounded meaning. The gesture survives only as encoded correlation in R; its duration survives as OL, not latency. This fracture is not failed fidelity; it is the irreducible remainder of temporality—what cannot be reduced to a vector, even if fully sampled. Craft persists as OL: a non-invertible temporal remainder—present in outcomes, absent from the substrate.

Hui’s theory of digital individuation strengthens the relocation thesis: form no longer emerges from a hand–material dyad but from a reticulated milieu of models, protocols, and datasets. Read this way, authorship’s disappearance is not an industry accident but the systemic effect of recursive infrastructures that pre-compose appearance via tertiary retentions. Where we part is the status of poiesis and remainder: Hui’s cosmotechnics keeps open the possibility of multiple ontologies and a poiesis of relations at the milieu level; I diagnose a de facto monism in which recursion exhausts disclosure and leaves only technē as endurance. Seams and glitches thus index not versional drift but non-reticulable duration—the preindividual residue a statistical ontology cannot phase. Cosmotechnical plurality remains rhetorically invoked; within my H/R/OL frame it adds no further theoretical load. The diagnosis hinges on the mechanics of retentional encoding and duration loss, not on adopting plural ontologies (Hui, 2016b, 2019, 2021).

Software as Fiction—Without Duration

Having located the remainder across traditions, I now show how statistical embedding operationalizes that remainder in practice.

Reading Trogemann & Damm, generative systems are fiction engines: they instantiate mathematical and software fictions that delimit what can appear, opening large possibility spaces whose runtime artefacts can surprise (Trogemann & Damm, 2019). I accept novelty yet insist: novelty ≠ poiesis. The model’s surprise is parametric—bounded by its ontology and timing—whereas craft’s poiesis is risked time. Accordingly, glitches that Trogemann & Damm treat as interpretive affordances are, in my terms, OL: seams that index the non-invertible duration the substrate cannot carry back. Against a merely versional account of digital time, the lag I name is non-reticulable duration—time that cannot be promoted to a new state without loss. It is not a versional diff; it is OL—a remainder of duration that cannot be promoted without loss.

Glitch theory clarifies how seams signify. Rosa Menkman argues that systems become legible through their interruptions—the breakdown is epistemic, not accidental (Menkman, 2011/2010). In my terms, any AI misrender (e.g., the misread kintsugi seam) functions as an index of OL: time-at-risk becoming visible as misfit. The “artefact” is disclosure, not defect; it converts noise into insight by surfacing the substrate’s compression assumptions. This directly supports the thesis that the value of craft appears at the moment generative models fail: where reproduction falters, duration speaks (Menkman, 2011/2010). As Menkman argues, glitches operate as negative definitions: by failing, they redraw the norm. The norm exposed here is the model’s inability to carry duration; the seam is OL made legible. Thus the break is epistemic: it forces a shift from image to infrastructure, making the politics of the code—and the temporal theft it performs—explicit.

At this juncture, failure becomes theory. Even advanced models misrender culturally idiosyncratic techniques—kintsugi seams read as visual noise. These glitches dramatise how tacit cultural timings—once entrusted to muscle memory—re-enter production only as pattern noise. Every misrendered seam is thus a temporal hiccup: evidence that embodied pacing cannot be compressed without distortion—as glitch aesthetics maintain, interruption is diagnostic rather than accidental (Menkman, 2011/2010). That models can generate novelty is beside the point: novelty without embodied pacing remains parametric. The vector holds pattern; it cannot carry duration. Trogemann & Damm would call this an emergent artefact of a possibility space; I call it time becoming legible as misfit. This marks the boundary of pseudo-poiesis: the model’s “surprise” is pre-calibrated. Chronos can permute; it cannot hazard Kairos. Hence re-risking time is not an aesthetic afterthought but an ethical injunction—to stage an interval of openness the stack cannot pre-time. The readings converge empirically while diverging ontologically: theirs marks fictional potential; mine marks temporal remainder.

Where Sennett reads craftsmanship as an ethical discipline of attention (Sennett, 2008), generative computation converts that discipline into a style token, retaining the gesture while hollowing the moral core. Thompson’s account of mechanisation reclassifying artisanal knowledge as redundancy (Thompson, 1963) finds its algorithmic correlate. Among these handles, Marx’s general intellect supplies a prognosis of skill becoming code—precisely the scenario AI enacts (Marx, 1973). Because these thinkers theorised technicity at moments of infrastructural shift, their frameworks remain diagnostic for AI: they offer handles for a reality where material machines are replaced by probabilistic models yet enact the same extractions. Each stage increases substrate competence while increasing OL relative to H—better recall, less recoverable duration. The ethical thinning of attention has an economic correlate: skill becomes code.

Marx’s general intellect anticipates a moment when living skill crystallises into dead objectivity, relieving capital of embodied competence (Marx, 1973, pp. 692–693). Generative AI radicalises that diagram: artisanal nuance is vectorised, archived, redeployed by large models whose optimisation loops complete dead labour’s autonomisation. Lazzarato’s immaterial labour updates Marx by stressing how aesthetic/affective contents become productive forces; vectorisation is the computational realisation of that turn (Lazzarato, 1996). Vectorisation consummates a three-stage arc: Thompson’s manual redundancy (labour removed from the workshop), Marx’s mechanical dead labour (skill fixed in gears), Lazzarato’s immaterial surplus value (affect encoded in signs). AI’s latent vectors compress all three into a single statistical ontology. Yet each mobilises a distinct infrastructure of extraction: steam loom captures bodily repetition; advertising engines harvest attention; data centres ingest correlation itself. Collapsing them into a single ontology risks masking infrastructural shifts even as it reveals a common logic of temporal appropriation. Contra Hui’s cosmotechnical plurality, I treat present AI as a de facto universal—an enforced monism of statistical ontology. Hui articulates a horizon of cosmotechnical plurality (2016b, 2019). My claim is diagnostic, not metaphysical: plurality can be affirmed rhetorically, but it is orthogonal to the OL thesis; the argument turns on the non-invertibility of H ⁣→R, independent of cosmotechnical commitments.

Viewed through Heidegger, each stage marks a further withdrawal of poiesis; what began as craft-driven revealing terminates as standing-reserve embeddings (Heidegger, 1977/1954). Thus, classical labour theories do not expire; they metastasise inside neural weights. Craft’s labour is transmuted into a fungible index without temporal commitment. The ontological surplus I track is precisely what survives that absorption—a temporally charged residue of living labour that remains present yet functionally redundant. Craft thus stands as empirical evidence that, even when subsumed by the general intellect, the time of the human persists inside the machine as a non-assimilable lag. This temporal custodianship, rather than symbolic resistance, is what links Marx’s prognosis to twenty-first-century design practice. Each optimisation step deletes experienced time even as it perfects technical recall.

Metrics fail to register what Heidegger calls Verweilen—the quiet staying of being (Heidegger, 1977/1954). Craft’s surplus supplies that lingering in machinic contexts that otherwise accelerate toward absence. Not as relevance, nor as resistance: it persists as ontological surplus—not required by the system, not excluded from it, but irrelevant to it. Its agency is negative: to testify, by sheer duration, that total optimisation cannot silence time. Craft, no longer needed for making, continues to mark the human as being-in-time, whose relation to matter, gesture, and worlding exceeds what is capturable by alignment objectives.

The question, then, is what kind of being belongs to something fully legible yet functionally void. I therefore argue for craft not as an aesthetic category or cultural practice, but as a temporal ontology within the age of computational sovereignty. I trace its logic not through linear evolution but through recursive return—a rhizomatic unfolding in which craft appears, disappears, and reappears as a punctum, rupturing machinic coherence without ever opposing it. Craft is not lost; it is simply not needed. And precisely in this unnecessity it begins to matter again.

To comprehend what persists, we must descend from political economy into ontology, beginning with technē. The task is to reread technē not as productive mastery (pre-industrial interpretations) but as endurance without disclosure—a post-AI reinterpretation in which craft’s value resides in refusing conversion to instantaneous availability.

Cyborg Interlude: Transitional Sympoiesis

Human–AI “co-making” functions as a bridge arrangement whose timing, loss functions, and data hygiene are preset by the apparatus. It stages the hand inside machine cadence. Its horizon is not mutual poiesis but parametric sufficiency: collaboration as a wrapper that normalises the drift rather than reverses it.

Craft as Ontological Residue: Technē Without Poiesis

Where Hui would count infrastructural reticulation itself as a new poiesis of relations, I reserve poiesis for time-at-risk. By that criterion, what remains is technē without poiesis: endurance after disclosure has been outsourced to the milieu. To grasp craft in its post-AI settlement, we must return to technē divested of its classical metaphysical load. In Heidegger’s account, technē is not merely technique, but a mode of bringing-forth—a revealing that is neither instrumental nor representational (Heidegger, 1977/1954). That poietic origin presumes a world where the human is the locus of disclosure. Such a world no longer obtains. In the age of generative AI, form, coherence, and even taste are computationally synthesised without recourse to human intention (Wen, 2024a).

The locus of world-disclosure has migrated into AI computing.

Workshop → factory → AI computing: what the factory did to embodied rhythm, AI computing does to ontological presence—compressing disclosure into microseconds of alignment and inference. On a Wardley map, craft has run its historical course—Genesis → Custom-built/Bespoke → Product → Commodity/Utility—and nothing escapes that curve (Wardley, Chapter 3: Evolution). But this turn is not merely another production revolution or an enabler shift. It is a competence revolution. Carbon intelligence (form negotiated in time-at-risk) now meets silicon intelligence (appearance pre-composed in manifolds). In Chomskyan terms: earlier preservations of craft were performance choices (tempo, seam, refusal) atop a shared competence of making; today, competence itself has moved upstream into infrastructure. Post-selection can still style outputs, but it cannot restore poiesis once disclosure is housed in the substrate. Hence OL: the non-invertible remainder of duration at the H↔R interface—what performance can display but not regenerate.

Generative AI can be creative; that does not restore craft. Creativity here is a search over parametric space; craft is risked duration in material time. As carbon hands cede to silicon stacks, variance becomes a slider, and contracts into operational latency (while OL persists as duration remainder)..

“Noise,” here, names the experiential trace of Bestand: present, usable, yet never fully heard (Heidegger, 1977/1954). In this post-AI horizon, craft becomes not a site of disclosure, but a residue—what persists after revealing has been absorbed into indifferent infrastructure. The kintsugi anomaly signals OL: an appearance-level trace of duration non-invertible from R. Heidegger’s Bestand names the mode of total availability (Heidegger, 1977/1954). Craft remains as OL made legible—what Agamben calls the remnant: a form left out of calculation yet still present (Agamben, 1999). Craft is the gesture after relevance. It does not reveal; it endures. It does not signify; it persists as an ontological remainder within a world that has no symbolic need for it. It is technē without mastery, without authorship, without consequence—yet not without weight. This reconfiguration aligns with posthuman ontologies that decentre the human not only from knowledge production but from ontological primacy (Braidotti, 2013; Hayles, 1999). Craft is not the site of human triumph over the machinic, but the index of a milieu in which human gestures no longer organise the field of meaning. And still, they endure.

Information-theory “noise” denotes low-amplitude residue after compression; craft resembles such noise, yet differs in that its residue is durational, not statistical. Craft occupies a similar structural role: the signal that lingers without contributing to alignment or coherence. Its persistence is not functional noise, but metaphysical remainder—a reminder that temporal variance cannot be entirely smoothed out. Thus craft can manifest as noise, but OL ≠ noise: noise is a phenomenology, OL is a structural non-invertibility—use noise as an index of OL. If Bestand flattens meaning into availability, Serres shows how its remainder hums as parasitic noise—present, enabling, unassimilable (Serres, 2007/1980). In this sense, Serres’ parasite and Heidegger’s Bestand converge: both name that which remains structurally external yet internally enabling. Generative AI foregrounds this paradox by enrolling craft’s gestures as training noise while simultaneously emptying them of poietic force. Where Pye’s “workmanship of risk” valorised variance and tactile feedback (Pye, 1968), probabilistic smoothing eliminates risk in favour of reproducible novelty. Yet precisely because risk is unnecessary, its echo persists as ontological noise—an inaudible crackle inside the perfectly compressed signal. Craft, then, is not resistance; it is resonance without need. Sloterdijk’s anthropology of resonance frames human–world relations as vibratory rather than oppositional; craft’s value lies in this vibro-affective register, not in negation (Sloterdijk, 2013).

Dispersed Agency: Post-Curation and the Fragmenting of the Human

Post-curation does not eliminate the human; it redistributes and fragments human agency across computational systems, infrastructures, and data flows. The human is no longer the initiating subject of design but a node within a constellation of generative and selective processes. This shift echoes Hayles’ posthuman subject (Hayles, 1999) and Braidotti’s affirmative ethics (Braidotti, 2013). Contemporary making persists, but as heterochronies interfacing computation rather than reversing its cadence. Haraway’s cyborg, once subversive, is absorbed in late posthumanism as capital’s affirmative biopolitics (Haraway, 1991). The craftsperson becomes a ghost-function—one who validates outputs without initiating them. Agency becomes diffused curation; embodiment lingers only as OL between H and R—registered as latency in operations, but not reducible to latency. Under Trogemann & Damm, this marks the shift from problem-solving to world-instantiating algorithms: code as a compact description of a possibility space whose exploration is curated rather than authored (Trogemann & Damm, 2019). My addition is temporal: such exploration occurs after poiesis has been outsourced; the human’s leverage is to tend the lag—curate, annotate, mis-time—rather than originate form. That latency is not an inefficiency; it is the final habitat of craft’s phenomenology. In that narrow habitat, what counts is not the object produced but the temporal atmosphere the lag sustains.

What remains of craft is not its form but its remnant ontology. Craft persists as a habit of care, friction, and orientation in a world that no longer asks for such things. It is folded into the loop not as disruption, but as faint lag—a presence not strong enough to alter system behaviour, yet irreducible to it. Marchand’s ethnography of Yemeni masons shows how apprenticeship encodes social memory in gesture—memory that post-curation flattens into data (Marchand, 2001). Adamson’s critique of modern craft marginalisation finds a post-AI correlate: the craftsperson re-enters the loop only as a stylistic curator, validating outputs whose provenance is machinic, not manual (Adamson, 2013).

Time Without Flow: Custodianship as Non-Alignment

If Bergson distinguished mechanical time from durée (Bergson, 1911/1907), AI radicalises the former: microseconds suspend qualitative flow altogether, while craft embodies durée that cannot be discretised. AI time is a calculus of microseconds—training epochs, inference calls, latency budgets. This machinic nanosecond is not simply faster; it is calibrated to erase anything that cannot be sliced into identical quanta. Craft time is heterochronic: the slow hydration of clay, the seasonal cure of hardwood, the generational relay of embodied know-how. Pye names this indeterminacy the workmanship of risk (Pye, 1968); Ingold casts it as dwelling along lines of material becoming (Ingold, 2013). Neither tempo is slower or faster; they are non-commensurable.

AI temporality seeks optimality through compression—obliterating continuity in favour of calculable efficiency. In contrast, craft carries temporal heterogeneity that AI systems neither accommodate nor value. Pye’s workmanship of risk names precisely the durational indeterminacy that optimisation seeks to eradicate (Pye, 1968). Ingold reframes making as dwelling, not a task to be compressed (Ingold, 2013). Craft is not slower than AI; it is timed differently. This asynchrony is what political economy misprices and what philosophy must name. Its duration is intergenerational, ecological, affective—with no dashboard, metric, or KPI. Its temporality is not retrograde but non-aligned.

Virilio warned that speed leads not to immediacy but to disappearance (Virilio, 1991/1980). Inference loops compress time so tightly that presence vanishes. Craft, by contrast, is neither fully visible nor invisible—it is durational opacity: a time-form too slow for machinic scanning, too dense for data erasure. Bergson’s durée underlines this opacity: qualitative time resists spatial-numerical translation (Bergson, 1911/1907). Virilio’s dromology accelerates; it cannot dissolve durée (Virilio, 1991/1980). Consider the 14-hour vitrification cycle of a porcelain glaze versus the 0.4-second inference call of a diffusion model. In Trogemann’s terms, the glaze cycle is kairotic not because it is slow but because its right time cannot be reduced to clock time; inference budgets enforce chronos (Trogemann, 2025). Custodial craft maintains a reservation window for OL—holding times apart so duration is not erased in replay.

If Bergson’s durée resists spatialization, and Virilio warns that pure speed erases presence, generative AI accomplishes both: it vectorizes making and accelerates it past perception, effectively disappearing the formative moment. Re-risking time is thus the attempt to reclaim Kairos within Chronos.

This incommensurability reprises the factory-craft split: industrial cadence once clashed with vernacular pacing; post-AI cadence renders craft an unalignable interval—too heterogeneous to synchronise. Each duration indexes a distinct cosmology of time. The point is not slowness but non-alignment: a small, irreducible friction that time-shifts the field.

To be a custodian of craft is not to preserve tradition in a museum sense, but to hold open incompatible durations. This is temporal stewardship that does not resist machinic time but simply does not answer to it. Craft does not slow AI down; it marks that not all durations collapse into instantaneity. The custodian reenacts Heidegger’s Gelassenheit—a letting-be that refuses the tyranny of speed. Gelassenheit operationalises Verweilen, turning lingering into deliberate practice (Heidegger, 1977/1954). Custodianship becomes a refusal of synchronisation. It names temporal practices that endure without utility or resolution—marked not by use but by refusal to be timed.

Tensions Without Resolution: Craft as Philosophical Impasse

Where Braidotti reads dispersed agency as an affirmative opening, Adorno hears in instrumentality the non-identical shard that refuses absorption. Craft, here, inhabits that split: useless to the system, yet not nothing—an enduring difference rather than a solution.

Because craft oscillates between moral remnant and inert code, it operates as a philosophical relay—short-circuiting Heideggerian dwelling with Deleuzian becoming, Adornian negativity with Braidottian affirmation. Adamson’s genealogy of craft’s marginal status in modernity anticipates its new ubiquity-as-irrelevance in the post-AI settlement (Adamson, 2013). To situate craft in this posthuman ecology is to face unresolved tensions: Heideggerian dwelling vs. Deleuzian becoming; Adornian negativity vs. Braidottian affirmation; Morton’s hyperobjects vs. Latour’s actor-networks (Morton, 2013; Latour, 2005). Craft becomes the site where these contradictions are neither reconciled nor resolved—only inhabited.

Adorno’s critique of functionalism as repression of subjective expression under instrumental rationality resonates here (Adorno, 1973). For him, modernist design collapses into abstraction when it abandons inwardness for utility. Generative AI replicates this collapse procedurally, converting gestures of care into empty style tokens. In Negative Dialectics, the non-identical persists beneath functional rationality (Adorno, 1973). Craft embodies that non-identical shard. It is not a solution to these tensions. Its endurance resembles a rhizome: popping up wherever systems attempt closure, refusing root yet refusing erasure. It allows contradiction to endure materially. The hand that moves without reason, the form that persists without function, the process that continues without endpoint—these are not errors, but metaphysical affirmations. Derrida’s critique of presence clarifies why: gestures become iterable marks once intention decouples, yet technics still carry temporal weight (Derrida, 1976/1967; Stiegler, 1998). Craft tends OL—institutionalising hesitation so that duration’s non-invertibility remains audible inside coherence. It endures where meaning stalls, systems disarticulate, temporality suspends its flow. Craft is not coherence withheld—it is coherence rendered irrelevant.

Epistemically Captured, Ontologically Surplus: Non-Significant Signal

What is captured is visibility; what is lost is time. In Baudrillard’s terms, the craft image persists as simulacrum; my addition is that the murder of the real is a temporal homicide: remove the duration, keep the look. And if Feenberg reads design as encoded choices, then one encoded choice of this stack is to delete pacing and uncertainty—precisely the variables OL preserves as remainder.

Marx foresaw in the “Fragment on Machines” that living skill would ossify into inert code once subsumed by automated systems (Marx, 1973). In the machinic logic of generative AI, craft is not misunderstood or overlooked; it is fully absorbed, parsed, intelligible. Its gestures, histories, materialities, and symbolic associations are mapped into high-dimensional vectors, reassembled as aesthetic options, indexed by prompt. Craft has been epistemically captured—its forms parsed, indexed, recombined. Yet capture is not recognition. Foucault warns that visibility is control (Foucault, 1977). Baudrillard calls it the murder of the real (Baudrillard, 1994). Feenberg insists that optimisation circuits smuggle normative choices about whose labour, culture, and temporality may be flattened (Feenberg, 1999). Baudrillard diagnoses loss of referential ground; Feenberg politicises that loss by tracing it to design choices. Combined, craft’s surplus appears not as tragic residue but as an index of selection and exclusion.

To the AI system, craft is just another parameterisable surface—used because it is available, not because it matters. Centuries of symbolic labour reduce to training data. What was once cosmology, temporality, ritual becomes a reference token in a generative chain. Derrida’s critique of presence finds an echo: in generative AI, craft no longer carries intentional trace—signifiers without source, endlessly recombined (Derrida, 1976/1967). Where Derrida foregrounds the slippage of signifiers, Stiegler reminds us that technics, even emptied of semantic load, retain temporal weight (Stiegler, 1998). Craft becomes this memory trace—gestures stripped of meaning yet temporally sedimented.

And yet the ontological surplus persists—not because AI cannot process it, but because its ontological charge is irrelevant to AI. Craft, having lost all exceptional status, becomes metaphysically residual: not unseen, but entirely seen and still unaccounted for. This surplus is not functional; it is not recoverable as value-add or efficiency metric. It is the remainder: the excess of worlding that survives reduction. Craft endures as OL rendered operationally surplus: fully legible in R, yet its duration non-recoverable. Stiegler argues that technics are mnemotechnical supports (Stiegler, 1998); even stripped of meaning, they store retentional traces. Craft’s echo is that stored retention: not semantic, but temporal memory. In Hui’s terms, latent vectors are tertiary retentions that pre-compose appearance; the remaining surplus is not missing data but non-retentional duration—time that resists addressability (Hui, 2016a, 2019, 2021).

Custodial Craft Practice: Guarding the Ontological Lag

Firstly, custodial craft crystallises the arc from industrial mechanisation to statistical vectorisation: what began as the loss of authorship and the translation of poiesis into standing-reserve embeddings re-emerges as a micro-politics of almost-nothing. Craftspersons materialise only after the inference pass, curating outputs they did not generate; by pausing inside that machinic after-image they preserve OL (the duration remainder at H↔R).

Custodial craft / ζ (speed-minus-one). Rather than “restore” poiesis, introduce a small damping (ζ > 0) at decision, render, or review stages—slow just enough that duration remains audible without changing the attractor. Here ζ names a positive damping ratio applied to R-side processes (selection, scheduling, exposure): a deliberate micro-friction that holds the H↔R seam so OL stays legible instead of being smoothed away. It does not generate a new H; it guards the remainder left by H→R encoding. The politics is minimal but concrete: timing becomes a site of intervention—a measurable drag rather than a reversal.

Secondly, the lag is not sentimentality for slow making but a structural counter-force—an infinitesimal drag on acceleration that exposes optimisation’s blind spot. The kintsugi misrender, previously read as noise, becomes an asset: a self-registering glitch (index of OL) that the custodian deliberately leaves unpolished, letting embodied tempo resonate inside statistical smoothness. In Latourian terms, the craftsperson functions as a relay-node that inserts friction into the network—speed minus one, measurable only in the aggregate yet crucial for re-grounding agency in duration (Latour, 2005).

Thirdly, custodial practice reframes critique as maintenance. Against nostalgic heroism or artisanal fetish, the custodian performs Feenberg’s demand to keep open the normative margins that alignment objectives erase—specifically, the margins where OL would otherwise be smoothed away (Feenberg, 1999). Repair replaces innovation; annotation replaces correction; withholding erasure becomes an affirmative tactic. By tending temporal fault-lines instead of producing novelties, the craftsperson demonstrates that value can lie in conserving irreducibility rather than extracting it. Unlike conservation (restore to original) or hacker repair (functional subversion), custodial craft sustains failure without end-use: the glitch is neither corrected nor revalorised; it is kept open as an ontological wrinkle.

Finally, this practice retroactively knots the paper’s threads—Bestand’s standing-reserve, Serres’ parasite-noise, Bergsonian durée, and the rhizomatic tensions—into a paradox: craft persists precisely because it has no functional reason to do so. It marks time without adding “content,” lending weight where metrics register none. In that gesture, we witness the living form of ontological surplus diagnosed throughout: a presence fully legible to the system yet forever exceeding its use.

Coda: Shadow Protocols (Tanizaki)

Returning to Tanizaki: what began as an optical metaphor reads, in this account, as a temporal protocol of reservation. Analogy, not alibi. Tanizaki’s claim—“we find beauty not in the thing itself but in the patterns of shadows”—is a temporal thesis disguised as an optical one (Tanizaki, 1977/1933). “Shadow” names reservation: a kept interval of low information in which matter and meaning are still negotiating. In the AI stack, that interval is operationally collapsed: auto-correction pipelines eliminate hesitation in favour of legibility. When the interval goes, time-at-risk goes with it.

Patina vs. patination. Generative AI and fabrication automation can convincingly simulate patina and even operationalise patination as protocol. But protocolised patination is not patination-as-duration. Patina is a look; patination is a process—time-at-risk. Once rendered as parameters and control loops, its uncertainty is bounded and replayable. Style (or even “process”) without non-reticulable duration is not poiesis; it is parametric novelty.

Shadow as protocol, not palette. Here “shadow” does not mean a colour grade; it names a rule of deferred disclosure. Pipelines that treat darkness, grain, or seam as error produce parametric chiaroscuro—a picture of shadow with the time removed. Neutrality is a policy: what passes as “default good quality” is optimisation to chosen losses and priors. The appearance of neutrality indexes the pipeline’s values, not their absence.

Hui: reticulation and pre-composed appearance. Today’s “brightness” is infrastructural. Tertiary retention (weights, datasets, prompts, protocols) pre-compose appearance upstream; what reaches the eye is a stabilised phase of a reticulated milieu (Hui, 2016a, 2019). Tanizaki’s kept dimness maps to what this milieu cannot retain: reservation that is not addressable as data. The seam, glitch, or uncorrected underexposure indexes not version drift but non-reticulable time—what the reticulation cannot phase.

Serres and the hum. Where Heidegger names total availability (Bestand), Serres names its leftover as noise—a residue that both enables and exceeds order (Heidegger, 1977/1954; Serres, 2007/1980). Tanizaki’s shadow functions as an OL-window: a kept reservation of duration inside exposure pipelines. AI reads it as a defect; this paper reads it as ontology.

Trogemann: chronos vs. kairos. The tea room’s dimness is a kept interval, not a mood; it withholds instant readability so that meaning condenses. AI’s exposure logic is chronos: minimise underexposure, maximise signal. My custodial tactic—speed-minus-one—keeps that interval open inside chronos without pretending to restore pre-AI poiesis (Trogemann, 2025).

What Tanizaki anticipated—and could not foresee. He anticipates that excess illumination flattens depth. He could not foresee a pipeline that treats every shadow as a loss term to be minimised—electric glare becomes algorithmic glare: not a lamp but a policy (no latency, no seam, no grain). The consequence is the thesis in miniature: variance becomes parametric; duration becomes OL—non-invertible by the pipeline. The task of custodial craft is to keep a remnant of shadow as rule—reservation that the stack cannot absorb—so that duration remains audible inside a pipeline of perfect recall.

Method note—simulation vs. duration. “Simulation” here includes any reproducible protocol for achieving the look or steps of patination. The distinction is ontological, not technical: simulation closes the space ex ante; patination-as-duration remains indexical and non-reticulable—time that cannot be parameterised without loss.

There is no ending to craft. No climax. It does not win, or resist, or return. It stays. It persists. It circulates as irrelevant in systems that consume everything but care for nothing. That weight—neither metric nor message—outlasts every cycle of optimisation. Such staying is not resignation. To endure as irrelevant is to oppose the system’s grammar without declaring war—refusal by duration, not defiance. Minimal agency resides here: not in intervention, but in presence. Technē without poiesis.

What else could we call meaning, if not that? Returning full circle, that quiet pulse is the same lagging presence first observed in craft’s initial misreadings. What began as a gap in data pipelines returns as an ontological injunction: remain. In that latency, technē survives—not as mastery, but as the quiet pulse of irreducible time. Thus, Heidegger’s technē, once a venue of revealing, is revealed here as latency: craft remains, but only as temporally thick code (Heidegger, 1977/1954). Craft-making is always in flux; under AI, the vector of that flux points toward the compression of durée (Bergson, 1911/1907). If magic names anything here, it is simply this: a kept interval—kairos held open inside chronos—where endurance lets meaning condense without mandate. Technē without poiesis.

---

References

Adamson, G. (2013). The invention of craft. Bloomsbury Academic.
Adorno, T. W. (1973). Negative dialectics (E. B. Ashton, Trans.). Seabury Press. (Original work published 1966)
Agamben, G. (1999). Remnants of Auschwitz: The witness and the archive (D. Heller-Roazen, Trans.). Zone Books.
Baudrillard, J. (1994). Simulacra and simulation (S. F. Glaser, Trans.). University of Michigan Press. (Original work published 1981)
Bergson, H. (1911). Creative evolution (A. Mitchell, Trans.). Henry Holt. (Original work published 1907)
Braidotti, R. (2013). The posthuman. Polity Press.

Chomsky, N. (1965). Aspects of the theory of syntax. MIT Press.

Derrida, J. (1976). Of grammatology (G. C. Spivak, Trans.). Johns Hopkins University Press. (Original work published 1967)
Feenberg, A. (1999). Questioning technology. Routledge.
Foucault, M. (1977). Discipline and punish (A. Sheridan, Trans.). Pantheon. (Original work published 1975)
Haraway, D. J. (1991). A cyborg manifesto. In Simians, cyborgs and women (pp. 149–181). Routledge.
Hayles, N. K. (1999). How we became posthuman: Virtual bodies in cybernetics, literature, and informatics. University of Chicago Press.
Heidegger, M. (1977). The question concerning technology (W. Lovitt, Trans.). Harper & Row. (Original work published 1954)

Hofstadter, D. R. (2007). I am a strange loop. Basic Books.
Hui, Y. (2016a). On the existence of digital objects. University of Minnesota Press.
Hui, Y. (2016b). The question concerning technology in China: An essay in cosmotechnics. Urbanomic.
Hui, Y. (2019). Recursivity and contingency. Rowman & Littlefield.
Hui, Y. (2021). Art and cosmotechnics. University of Minnesota Press.
Ingold, T. (2013). Making: Anthropology, archaeology, art and architecture. Routledge.
Latour, B. (2005). Reassembling the social: An introduction to actor-network-theory. Oxford University Press.
Lazzarato, M. (1996). Immaterial labor. In P. Virno & M. Hardt (Eds.), Radical thought in Italy: A potential politics (pp. 133–147). University of Minnesota Press.
Marchand, T. (2001). Minaret building and apprenticeship in Yemen. Routledge.
Marx, K. (1973). Grundrisse: Foundations of the critique of political economy (M. Nicolaus, Trans.). Penguin. (Original work written 1857–1858)
Menkman, R. (2011). The glitch moment(um). Institute of Network Cultures. (Original manifesto published 2010)
Morton, T. (2013). Hyperobjects: Philosophy and ecology after the end of the world. University of Minnesota Press.
Pye, D. (1968). The nature and art of workmanship. Cambridge University Press.
Ribault, P. (Ed.). (2022). Design, Gestaltung, Formatività: Philosophies of making. Birkhäuser.
Sennett, R. (2008). The craftsman. Yale University Press.
Serres, M. (2007). The parasite (L. R. Schehr, Trans.). University of Minnesota Press. (Original work published 1980)
Simondon, G. (2017). On the mode of existence of technical objects (C. Malaspina & J. Rogove, Trans.). University of Minnesota Press. (Original work published 1958)
Sloterdijk, P. (2013). You must change your life (W. Hoban, Trans.). Polity Press.
Stiegler, B. (1998). Technics and time, 1: The fault of Epimetheus (R. Beardsworth & G. Collins, Trans.). Stanford University Press.
Tanizaki, J. (1977). In praise of shadows (T. J. Harper & E. G. Seidensticker, Trans.). Leete’s Island Books. (Original work published 1933)
Thompson, E. P. (1963). The making of the English working class. Penguin.
Trogemann, G., & Damm, U. (2019). Algorithms rethought: On the fictional potential of software. In W. Speitkamp & C. Weinreich (Eds.), Idee Inhalt Form: Beiträge zur Gestaltung der Gegenwart. Bauhaus-Universität Weimar.
Trogemann, G. (2025). Software and magic: Or an attempt to re-enchant the world. In The unknown in design, art, and technology (pp. 225–254). transcript Verlag. https://doi.org/10.14361/9783839476819-009
Virilio, P. (1991). The aesthetics of disappearance (P. Beitchman, Trans.). Semiotext(e). (Original work published 1980)

Wardley, S. (n.d.). Wardley Maps: Topographical intelligence in business — Chapter 3: Evolution. WardleyMaps.com. https://www.wardleymaps.com/learn/book-passages
Wen, J. (2024a). Taste of nations: Geopolitical logics in generative design. Medium. https://medium.com/@jimiwen/taste-of-nations-e7f0fae56115
Wen, J. (2024b). The collapse and contraction of post-curation. Medium. https://medium.com/@jimiwen/the-collapse-and-contraction-of-post-curation-eed0fc9a2424