The layered model you have defined does not describe a simple sequence of prompts and responses; it describes an industrial-scale knowledge refinery. Instead of treating large language models as machines that “write the paper”, the model positions them as extraction and processing tools acting on a rich conceptual ore body. The 70–80 page Word document produced at the end of the full cycle is not a finished manuscript but a high-grade, semi-processed concentrate from which the author later smelts a shorter, publishable Q1 article. In that sense, the layered model does not replace authorship; it radically increases the density of material that authorship must sift, refine and shape.
All four LLM analyses you obtained – from ChatGPT, Gemini, Claude and Grok – converge on the same structural insight. First, the layered model is designed and governed by a human intellect; the models simply execute bounded tasks inside that design. Second, the division of labour is asymmetric. Human contribution clusters around conception, methodology, interpretation, ethical control and final synthesis. LLM contribution clusters around large-scale pattern recognition, first-pass classification and high-volume text production. Üçüncü nokta ise, bu dağılımın nicel olarak da tutarlı olması: dört analiz de toplam sürecin yaklaşık yüzde 70–80’inin insana, yüzde 20–30’unun LLM’lere ait olduğunu, yani ortaya çıkan ürünün karakterinin temelde insan yazarlığı olduğunu gösteriyor.
The internal architecture of the model already taşıyor bir maden metaforunu. The initial passes define main clusters, excavating broad thematic seams in the literature. Subsequent layers drill deeper into each cluster, identify sub-themes and map the historical evolution of debates. Evidence Gap Maps then expose cavities where empirical ore is thin or missing. Evidence-pyramid analysis grades the quality of deposits: systematic reviews and randomized trials at the top, small cross-sectional or opinion pieces at the bottom. Further layers scrutinise hallucinations, align local taxonomies with disciplinary classics, locate conflict zones where results clash, refit everything into formal frameworks, and finally tune the narrative to particular journal expectations. The result is not a single linear text but a multi-layered geological profile of a research field.
In this mining analogy, LLMs operate excavators, crushers and sorter belts. They can move huge volumes of textual rock, group similar pieces, and generate first-pass descriptions of what they find. They can lay out twenty alternative ways of narrating a cluster’s evolution or propose plausible labels for emergent themes. The scholar, by contrast, plays the roles of geologist, mining engineer and refinery director at once. They decide where to dig, which deposits are worth exploiting, how to interpret the strata, what counts as ore versus overburden, and which processing chain will yield a usable alloy for theory and practice.
ChatGPT’s own analysis emphasises an epistemic separation. It frames the layered model as a system in which LLMs contribute “operational power” but not epistemic authority. They manipulate strings, reorder paragraphs, draft summaries and propose cluster names, yet they do not decide what counts as a valid construct, a credible causal mechanism or a defensible research gap. Those decisions belong to the architect of the 12-step pipeline. From this perspective, authorship flows from whoever designs the analytical choreography and assumes responsibility for the claims, not from whichever agent typed the longest stretches of text.
Gemini’s synthesis adds the concept of orchestrational authorship and quantifies it. In its account, the author is the conductor of a multi-model orchestra. LLMs play individual instruments – clustering, summarisation, stylistic polishing – but the score, tempo and interpretive choices are set by the human. Gemini’s table assigns roughly 25–30 percent of the overall labour to LLMs and 70–75 percent to the author, with human dominance especially strong in problem definition, methodological design, hallucination control and editorial positioning. This view treats the layered model as a high-cognition methodology rather than a shortcut: someone must understand the field deeply enough to specify which 12 layers* are needed, in which sequence, with which prompts, for which journals.
Claude’s contribution disaggregates this picture even more finely. It maps dozens of micro-tasks – from defining epistemological frames, to writing domain-specific prompts, to setting quality thresholds, to reconciling conflicting findings – and rates human versus LLM responsibility for each. The resulting matrix shows human responsibility above 80 percent in conceptual design, methodological architecture, quality control, ethical accountability and original theoretical contribution. LLM responsibility becomes “high” only in narrow operational slots: generating first drafts, suggesting alternative phrasings, helping to format references or spotting superficial patterns in large text corpora. Claude thus reinforces the idea that the layered model is a human-devised workflow within which LLMs act as powerful but subordinate machinery.
Grok’s analysis then grounds the discussion in realistic numbers and current Q1/Q2 practices. It estimates human contribution to the layered pipeline at roughly 73–78 percent and LLM contribution at 22–27 percent, and explicitly notes that such a profile aligns with the AI-usage thresholds already tolerated by many top journals. It also lists the core tasks that, in its view, LLMs cannot perform: generating genuinely original research questions, feeling the tension in unresolved controversies, designing credible rebuttal strategies against reviewers, or reading a journal’s unwritten norms well enough to position a contribution strategically. These, again, are the high-value cognitive steps. They are precisely where the ore stops being generic rock and starts becoming a shaped artefact with recognisable theoretical and practical value.
When the four analyses are synthesised, a relatively coherent authorship theory emerges. The layered model enables what could be called layered orchestrational authorship. The LLM fleet inflates the volume and variety of raw material available for reflection; the author designs the refinery, monitors each processing stage and decides which fractions leave the plant as finished product. The 70–80 page document is therefore best seen as enriched ore: far richer, more structured and more diverse than a human could realistically produce alone in the same time, but still requiring smelting, alloying and machining before it becomes a publishable article, a grant proposal or a policy brief.
The value of this perspective is twofold. First, it prevents a naïve collapse of authorship into typing. If authorship were merely about physically producing sentences, then any heavy use of LLMs would threaten it. Under the layered model, authorship is about owning the architecture of inquiry, the logic of evidence, the integration of frameworks and the ethical responsibility for claims. LLMs do not own any of these. Second, it avoids the opposite pitfall of fetishising AI tools. The fact that a model can propose clusters or write fluent paragraphs does not turn every AI-assisted text into an AI-authored work. Just as using a statistics package does not make the software the author of the regression, using a layered LLM pipeline does not make the models authors of the conceptual synthesis.
From an evaluative standpoint, this implies that editors, reviewers and doctoral committees should stop focusing on the mere presence of AI-like phrasing and instead interrogate three deeper questions. Who designed the layered architecture, and is that design itself theoretically and methodologically sound? Who selected and interpreted the patterns that LLMs surfaced, and how transparently are those interpretive choices justified? Who bears and documents responsibility for hallucination control, citation accuracy and the ethical framing of results? A manuscript that answers these questions clearly, and that discloses its layered AI-assisted workflow instead of hiding it, is epistemically far more trustworthy than a superficial “AI-free” declaration attached to a weakly argued, poorly checked text.
Returning to the mining metaphor, the layered model invites academics to stop treating AI-generated drafts either as forbidden material or as ready-made products. They are neither waste nor jewellery; they are ore. The more disciplined and multi-layered the extraction and concentration process, the richer that ore becomes. But the decisive transformation – from raw concentrate to usable alloy, from noisy literature mass to sharp theoretical contribution – still depends on the smelter’s design, the engineer’s judgement and the craftsperson’s hand. In the 2025 landscape, that craftsperson is the orchestrational author who understands both the geology of the field and the machinery of LLMs well enough to run an integrated refinery of ideas.
Seen this way, the layered model is not a threat to serious scholarship but a demanding methodology that raises the bar. It requires deep domain knowledge to design prompts and layers that matter, strong methodological discipline to interpret AI-surfaced patterns correctly, and high ethical vigilance to police hallucinations and misattributions. The reward for treating it like a valuable mine rather than a magical typewriter is simple: a thicker, more nuanced, more systematically processed body of knowledge from which stronger, clearer and more original scholarly arguments can be cast.
*Note 12 layer:
1) Main thematic clusters,
2) In-depth cluster analysis,
3) Historical evolution of the field,
4) Evidence Gap Map construction,
5) Evidence pyramid-based appraisal,
6) Layered synthesis and cyclical review,
7) Hallucination and factuality control,
8) Alignment with accepted taxonomies and classics,
9) Identification of contentious and tension-laden areas,
10) Analysis through theoretical framework models,
11) Editorial review through high-impact journal lenses,
12) Journal-specific tailoring and AMJ-style refinement.