A study of how platform-based digital structures came to organise creative capability into largely separate domains, and of how generative AI — by lowering the cost of moving work between writing, visual design, music, video, code and research, alongside the shift from ranked search results toward composed answers — may be changing the position of the person at the centre. The station is at the question-forming stage.
define the boundary first, or the study drifts
Over roughly two decades, much of the commercial digital environment organised creation by medium. Writing was supported by one set of infrastructures, images by another, sound by a third. Many capabilities acquired their own account, grammar, audience and metric of worth — a process the literature describes as the platformisation of cultural production (Nieborg & Poell, 2018). Where that organisation held, professional identity tended to be described by medium: a person was a writer or a designer, rather than by subject or intent.
Generative systems lower the cost of crossing those lines. A paragraph can become a storyboard, a score sketch and a working prototype in one sitting, by one person. The research question is not whether such output is good. It is what happens to the human position — authorship, intention, identity, continuity — where the seams between capabilities no longer function as enclosures. A parallel shift is under way in knowledge access, where ranked lists of sources are in some settings supplemented or replaced by composed answers (Metzler et al., 2021; Shah & Bender, 2022).
An independent research station. It studies a structural transition; it does not describe, explain or promote Earth LinC. Its outputs at this stage are provisional questions and evidence positions — neither findings nor normative design proposals.
Not anti-platform. Platforms built reach, livelihoods and craft communities, and remain important infrastructure. The argument is structural rather than moral, and it describes a tendency, not every platform.
Every claim is filed either as supported — attributed to cited peer-reviewed work — or as provisional, meaning this station has not yet established what the literature contains. The reading behind this page was preliminary: its search scope, databases and dates are undocumented, so a source not located here is not evidence that none exists. The two are never blurred.
Separation by medium is a property of particular infrastructures,
not a necessary property of creative work.
Working proposition · Station paper 01, §2 · provisional, not a finding
where the pattern holds: one account per capability
Much platform architecture followed the logic of the distribution channel it required: commonly one medium, one format, one upload path, one dominant ranking signal (Gillespie, 2010; van Dijck, Poell & de Waal, 2018). This made discovery possible at large scale, and it tended to ask makers to arrive pre-sorted. A maker selected a lane, learned its dialect, accumulated an audience there, and left other capabilities outside the door; general-purpose and multi-format services are partial exceptions. The labour consequences of that arrangement are well documented (Duffy, 2017; Poell, Nieborg & Duffy, 2021).
The cost was rarely visible inside any single tool. The station locates it in the seams: the essay whose argument thins when it becomes a video, the composition separated from the narrative it belonged to, the research that does not reach the design decision it could have shaped. Finished work was portable between enclosures; the reasoning around it travelled less readily. This is a working description of the seams, offered for testing, not a measured result.
Figure 1a — Held separately (schematic). Six capabilities, six enclosures, each drawn with its own audience and scoring logic; only finished files cross the dashed boundary. The centre is left unmodelled: the person is implied but not represented.
Figure 1b — Routed through one person (schematic). The same capabilities drawn through one maker; translation between media is cheaper and the centre becomes load-bearing, since intention, judgement and context are supplied from it. The two are idealised modes that coexist in practice, not a sequence.
a second line: work and its sources
A second line of change concerns where the evaluation of sources takes place. Early web search returned an index — a list of documents whose evaluation remained largely with the reader, though the politics of that list were contested from the start (Introna & Nissenbaum, 2000). Ranking and personalisation then moved part of that first act of judgement upstream into the system (Pariser, 2011; Noble, 2018; Bucher, 2018). Retrieval had already begun to alter what people retain, not only where they look (Sparrow, Liu & Wegner, 2011).
Generative retrieval extends that movement. The proposal to make “domain experts out of dilettantes” by replacing ranked lists with synthesised answers (Metzler et al., 2021) describes a regime in which the artefact returned is a composition rather than a source list. Retrieval-augmented generation (Lewis et al., 2020) keeps sources in the loop technically; how far they remain visible at the surface varies by system and interface. Critiques of that surface argue that search and answering are different acts that should not be conflated (Shah & Bender, 2022; Bender et al., 2021).
Figure 2 — Three modes of knowledge access (schematic). Each step moves the first act of evaluation further from the reader. The dates mark when each mode became prominent, not when the previous one ended: index search, ranked feeds and composed answers are all in use today. The station treats this as the knowledge-side counterpart of capability separation — a proposition it intends to test, not an established result.
For a maker working across several capabilities, these two lines meet. Knowledge may arrive pre-composed at the point where the person is asked to supply more judgement rather than less. That conjunction — connected making under composed knowledge — is the condition this station proposes to study. Whether the two literatures address each other is itself an open question here: the preliminary reading behind this page found little that did, but it was not a documented systematic search.
The station's working reading is that the most consequential effect of generative systems here is connective: they lower the cost of carrying an idea between capabilities. On that reading, the useful unit of analysis shifts from the artefact within one lane to the intention as it travels across lanes. Early field evidence is consistent with part of this: measured gains concentrate where a person can now cover ground that previously required another specialism (Noy & Zhang, 2023; Brynjolfsson, Li & Raymond, 2023), while effects near the boundary of capability are uneven and easily misjudged (Dell'Acqua et al., 2023). Those studies address specific tasks and settings and do not test the broader reading.
Where these systems work well, execution within a single capability becomes cheaper and more plentiful, and selection becomes comparatively more demanding: knowing which of many plausible versions is the one meant, and being able to say why. On this account, taste, context and continuity move from background conditions toward the centre of the work — the position Human-Centered AI argues should be designed for rather than assumed (Shneiderman, 2020, 2022; Heer, 2019).
The station's hypothesis about difficulty is that it relocates rather than disappears: from acquiring several crafts to holding one thread coherent while several move at once. This is consistent with the long-observed irony of automation, in which the residual human task can become harder rather than easier (Bainbridge, 1983); it has not been tested for cross-capability creative work.
each one needs its own evidence — they do not answer each other
Seven provisional research questions organise the study at this stage. They were formulated before a documented systematic review and may be revised, merged or withdrawn once one is complete. They are not stages and do not resolve into one another; each would require a distinct method and a distinct body of evidence. Each card states the question and the literature it is read against.
Authorship has often been inferred from visible labour: you made the marks. Where capability is connected, the maker's contribution shifts toward framing, selection and refusal. The preliminary reading behind this page did not locate settled vocabulary or measures for a position that is neither hand-craft nor supervision; that may reflect the limits of the reading.
Read against: Suchman 1987 · Horvitz 1999 · Heer 2019 · Epstein et al. 2023Much evaluation measures output quality, novelty or user preference. The station did not locate a measure of fidelity to intent across a chain of translations — text to image to sound to build. Intention is commonly treated as an input rather than as something that can be lost in transit.
Read against: Cherry & Latulipe 2014 · Lubart 2005 · Lee, Liang & Yang 2022Platform identity has been supported by account structures: a handle, a follower count, a consistent format. A maker working across several capabilities may have a recognisable voice with no single home for it. How that coherence is perceived, by others and by the maker, was not described in the material read here.
Read against: boyd & Ellison 2007 · Marwick 2013 · Duffy 2017Drafts encode reasons that are not always recorded alongside the file. Files travel more readily than the reasoning behind them. The station treats context loss as a candidate failure mode of connected creation, and looks for practices — annotation, provenance trails, versioned reasoning — that survive it.
Read against: Bush 1945 · Engelbart 1962 · Whittaker 2011 · C2PA 2021—Much co-creation research studies the dyad: one human, one system. Ordinary practice is often messier — several people, several models, overlapping capability. Credit, trust and disagreement in that setting were thinly covered in the material read here.
Read against: Davis et al. 2016 · Rezwana & Maher 2023 · Bansal et al. 2021Craft knowledge accumulates in hands and habits (Sennett, 2008). It now also accumulates in prompts, presets, archives and model versions, all of which can be deprecated. Continuity across years, tools and model generations is a question the station proposes to study.
Read against: Jones 2007 · Sellen & Whittaker 2010 · Dreyfus & Dreyfus 1986Generative retrieval can remove the ranked list from the surface of the work. The station asks how makers re-establish an evidentiary trail they no longer see, and whether the capacity to evaluate sources changes when it is exercised less often.
Read against: Metzler et al. 2021 · Shah & Bender 2022 · Lewis et al. 2020 · Sparrow et al. 2011left column is anchored in literature; the right is not yet — a statement about this review, not a finding
The first duty of a research station is to avoid rediscovering what is already known. The left column lists fields with an existing peer-reviewed base, each anchored to named work in §VIII. The right column lists questions this station has not yet been able to place against the literature: the reading behind them was preliminary and its search scope and dates are undocumented, so each is a question to be checked by a documented systematic search, not a demonstrated absence. Each is stated with an illustrative study design; none of those designs has been carried out.
Figure 3 — The intersection this station works in (schematic). Each strand is separately well developed: Human-Centered AI, platform studies, and the emerging evidence on AI capability expansion. The object of this station sits where they meet — connected making, under platform-formed identity, with composed knowledge. The preliminary reading behind this page did not locate literature addressing that intersection directly. That is a statement about this reading, not a demonstration that none exists; a documented systematic search is the next step. The right-hand column below states the intersection as seven provisional questions.
The programme reads Human-Centered AI strictly: not AI made pleasant to use, but AI studied with the person's position, intention and continuity treated as dependent variables. The methods below are proposed; none of the studies has been completed. Work is intended to proceed slowly, in small studies, published in full including negative results.
Construct a documented systematic map of the seven fields in §VI — scope, databases and dates recorded — before advancing any claim of the station's own.
Long-form qualitative interviews with makers already working across capabilities, analysed in their own terms rather than survey scales.
Follow one intention through a complete project — draft to image to sound to build — and record where and how it changes shape.
Draw the structure before arguing about it. Drawing is used here as a test of whether a structure is understood well enough to state; the drawings are schematic.
Publish questions, including those that remain unplaced. A question stated with an illustrative design is a contribution, though not a finding.
No claim that platforms are ending. The study asks what the structure is becoming, not what should replace it.
The station holds one hypothesis, stated plainly: that when capabilities rejoin, the person at the centre becomes more legible rather than less. That is a hypothesis, not a conclusion, and it is falsifiable. The study remains in progress until evidence arrives either way.
Station paper 01 · question-forming stage · findings published as completedthe ground the station stands on
Short-form citations, grouped by research strand. Inclusion records the ground this station reads from; it does not imply endorsement of each position.
Citation practice: claims in §I–§IV are attributed to the works above. Where a claim carries no citation it belongs to the station's own argument and is filed as provisional in §VI until evidence is produced.