Anyone who writes a book, a paper, or a piece of reportage today with the help of a language model runs into an uncomfortable question sooner or later: who actually wrote this sentence? The question sounds reasonable. Unfortunately it points in the wrong direction. A paragraph now takes shape over several passes, half dictated, half suggested, rearranged three times, a joint production in which nobody can finally say who earns the applause and who gets the bill. Authorship of individual words is barely reconstructable. The question that can be answered is a different one: who stands behind this sentence?
This piece describes how a framework for distributing decision authority between humans and AI can be applied to writing itself. What came out of it is an app, linked here: the Writing Workshop.
Capability is not legitimacy
The Adaptive Complementarity Framework (ACF) was developed for organizations that have to decide which decisions an AI system may make on its own, which ones it should prepare, and which ones are better left with a person (Russmann 2026). Its starting point is an observation that is simple and easily overlooked: the fact that a system performs a task well says nothing yet about whether it should perform it. Capability and legitimacy are different quantities. Confuse the two, and you automate, with impressive consistency, exactly those decisions nobody can answer for afterward.
Writing is no different. A language model often phrases a transition more elegantly than an exhausted author at midnight. With the same elegance it also supplies a page reference that looks plausible and simply does not exist. Both performances feel strikingly alike, and they call for entirely different responsibilities.
A text is a sequence of decisions
The first step is to stop treating "the book" or "the article" as one indivisible unit. Look closely and a chapter falls apart into decisions of very different kinds:
- the core thesis and its place in the overall arc,
- a direct quotation with a page reference,
- an interpretation, for instance of Kant's distinction between determinative and reflective judgment,
- a transition between two sections,
- the LaTeX markup of a formula.
These decisions have different properties, and it is precisely those properties that the ACF captures through five dimensions.
Five dimensions, translated to writing
The first two dimensions go back to Kant. Determinative judgment subsumes the particular under a given universal; reflective judgment first has to find the universal that fits the particular (cf. Kant 2000: 179–181). From this follow uncertainty and novelty as signals that plain rule application no longer suffices.
Uncertainty (U) asks whether a statement is covered by available sources. Page references, publication years, and attributions carry structurally high uncertainty, because a model can produce them fluently and with conviction without their being true. Fluency here, unlike in this text, is no seal of quality.
Novelty (N) measures the distance from what is already established. A paragraph that reformulates an existing definition for a new audience has low novelty. A paragraph that introduces a new concept has high novelty.
Value conflict (W) concerns interpretation, positioning, and statements about real people. Here it is not the solution that is unclear but the goal itself that is contested. Such a passage needs a decision point where someone can give an account, an address, in other words, not an inbox without a sender.
Precision (ϖ) describes how reliable the AI is for this particular kind of task, not how capable it claims to be in general. The term comes from precision-weighted inference in the Free Energy Principle, where signals are weighted by their expected reliability (cf. Friston 2010: 127–138; Clark 2013: 181–204). For LaTeX markup, precision is high. For bibliographic detail, it is soberingly low.
Prediction error (ε) only shows up after the fact: as the share of AI drafts the author had to revise substantially or discard. This dimension does not steer the individual passage. It steers the calibration of whole categories of work.
Gates that nothing offsets
The most important building block of the framework is pleasantly inconspicuous. Before any weighing takes place, non-compensatory gates check whether a decision is eligible for the AI at all (Russmann 2026). High precision cannot offset a high value conflict. A brilliant draft does not make an unverified page reference more reliable, it only makes it prettier.
For writing, this yields a short list of categories that always stay with the author:
- quotations, page references, and bibliographic details,
- figures, data, and empirical claims,
- statements about real people,
- the core thesis and central interpretations,
- the decision to publish.
Each type of text adds its own categories. In reportage, every factual claim and every direct quotation from an interviewee belongs on the list. In a monograph, it is axioms, proofs, and decisions about notation.
Three modes and one asymmetric rule
Every work step is assigned one of three modes:
- AI-led: formatting, style checks, reconciling in-text citations with the bibliography.
- Hybrid: first drafts from the author's own material, translations, transitions. The AI delivers, the author revises actively.
- Human-led: theses, interpretations, verification of quotations, publication.
More interesting than the assignment itself is the rule by which it may change. Making a step stricter, that is, requiring more human involvement, is possible at any time. All it takes is a short justification, which doubles as the learning signal for calibration.
Loosening a step, by contrast, is not a spontaneous decision at the individual step, however tempting that may be at midnight. It is a change of policy, with a justification and a new version number. Steps behind a gate cannot be loosened at all.
This asymmetry has a philosophical reason. Heinz von Foerster formulated as an ethical imperative that one should act so as to increase the number of choices (cf. von Foerster 1993: 133–134). Applied to tools, that means the option of taking a decision back must never disappear behind a procedure. The option of handing it over may well do so.
The log of all decisions is more than bookkeeping. Luhmann showed that in complex systems legitimacy arises less from insight into every single operation than from reliable, traceable procedures (cf. Luhmann 1969: 26–29). The inside of a language model remains opaque, and that much you can count on. The procedure in which it is used need not be. Luhmann called this constellation controlled intransparency: the internal workings stay hidden while the externally relevant behavior is fixed and can be checked (cf. Luhmann 2017: 186–187).
The ACF condenses these requirements into three properties every implementation has to satisfy (Russmann 2026):
- Traceability: every decision can be reconstructed.
- Contestability: every assignment can be challenged.
- Governance closure: rules change only through an explicit, logged change.
The card index
In his guide to academic work, Umberto Eco described a system of index cards that keeps reading, quotations, and one's own thoughts apart (Eco 1977). The Writing Workshop takes up that idea and adds what collaboration with AI demands: a status.
There are bibliographic cards, reading cards, quotation cards, fact cards, topic cards, person cards, connection cards, and working cards. Two markers matter most:
- Bibliographic, quotation, and fact cards count as unverified until the author confirms them against the original.
- Whatever travels from a conversation with the AI into the card index stays marked as an AI suggestion until the author adopts it in substance.
The origin of a thought is therefore not lost, not even after months in which one would have sworn one would remember.
The work as a text adventure
The most unusual component is a text adventure. Every chapter, scene, or step of an argument becomes a room, and the order of the rooms is the reading order. Motifs, concepts, and facts are objects. A reader possesses an object only after entering the room in which it is introduced.
Move through the work with commands such as next, go 5, or inventory, and you get messages like this: Blocked: the concept "address of responsibility" is still unknown to the reader, it is introduced in chapter 2. A character who speaks in chapter 3 but is only introduced in chapter 5 stands out immediately. So does a motif that is introduced and never picked up again, or a fact in a piece of reportage that rests on no verified card.
The metaphor is more than play. Kant's distinction is at work here too: a concept can only be applied if it is available. What a reader does not yet have, a reader cannot subsume anything under.
This check works from rules and is therefore AI-led. In addition, an AI reader can play through the same world and report structural observations, such as chapters without a recognizable function or theses that do not build on one another. That is hybrid: judging those observations stays with the author.
The AI as a partner in discussion
In the discussion area, the AI appears in fixed roles: Socratic questioner, devil's advocate, rigorous reviewer, first reader, editor, or a checklist that flags every claim in need of evidence. All roles follow the same workshop rules:
- no invented sources, quotations, figures, or statements by real people,
- conjectures are marked as conjectures,
- no decision taken in the author's place.
Why a partner in discussion and not a co-author? The place to which questions about a text are addressed demands more than fluency. Whoever occupies it must be available to be asked. They must be able to revise a standard and make it their own. And they must bear a loss if the text does not hold. No increase in capability satisfies these conditions by itself, however impressive the model cards read.
The tool does not claim that only a human can occupy that place. It makes sure the place is occupied, and that it stays visible by whom.
The real risk: passivity drift
The greatest danger in writing with AI is not the obvious error. It is gradual habituation. Research on automation bias and complacency describes how people increasingly follow automated recommendations and scale back their own checking (cf. Parasuraman/Manzey 2010: 381–410; Skitka/Mosier/Burdick 1999: 991–1006). Parasuraman and Riley distinguished early between misuse and disuse of automation (cf. Parasuraman/Riley 1997: 230–253). In writing, misuse shows up as revision that imperceptibly turns into nodding along, the classic career path of a critical reader.
The Writing Workshop counters this with three design decisions:
- Nothing is preselected. When the AI proposes a structure derived from a draft, not a single suggestion comes checked. Adoption is an act, not an omission.
- Drift protection in the book process. For each chapter, at least one section is sketched by the author first, before any AI draft is read.
- A falling revision rate is a warning sign. A decline can mean the AI has improved. It can just as easily mean attention has slipped. The log makes that ambiguity visible instead of booking it as success.
A proposal
The Writing Workshop is a proposal, and that is how science works. A design theory is spelled out as a concrete artifact so that others can use it, test it, and develop it further. Design science research describes this path as an interplay of building and evaluating (cf. Hevner et al. 2004: 75–105). A working instance makes a design theory tangible and belongs among its components as an expository instantiation (cf. Gregor/Jones 2007: 322–335).
The mode assignments are a starting configuration that can be calibrated through the log. Every type of text, every gate, and every discussion role is defined as data and can be changed and extended. Anyone who uses the app for a project of their own generates exactly the experience against which the proposal has to be measured.
The app is linked here: Open the Writing Workshop
The question of who wrote a sentence will get harder to answer with every generation of models. The question of who stands behind it can be organized. One only has to want it.
References
Clark, A. (2013): "Whatever Next? Predictive Brains, Situated Agents, and the Future of Cognitive Science", Behavioral and Brain Sciences, 36(3), 181–204. https://doi.org/10.1017/S0140525X12000477
Eco, U. (1977): Come si fa una tesi di laurea. Milan: Bompiani.
Friston, K. (2010): "The Free-Energy Principle: A Unified Brain Theory?", Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787
Gregor, S./Jones, D. (2007): "The Anatomy of a Design Theory", Journal of the Association for Information Systems, 8(5), 312–335. https://doi.org/10.17705/1jais.00129
Hevner, A. R./March, S. T./Park, J./Ram, S. (2004): "Design Science in Information Systems Research", MIS Quarterly, 28(1), 75–105. https://doi.org/10.2307/25148625
Kant, I. (2000): Critique of the Power of Judgment. Ed. by P. Guyer, trans. by P. Guyer and E. Matthews. Original work published 1790. Cambridge: Cambridge University Press.
Luhmann, N. (1969): Legitimation durch Verfahren. Neuwied: Luchterhand.
Luhmann, N. (2017): Die Kontrolle von Intransparenz. Berlin: Suhrkamp.
Parasuraman, R./Manzey, D. H. (2010): "Complacency and Bias in Human Use of Automation: An Attentional Integration", Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055
Parasuraman, R./Riley, V. (1997): "Humans and Automation: Use, Misuse, Disuse, Abuse", Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
Russmann, M. (2026): The Address of Responsibility: A Design Theory for Allocating Decision Authority in Human-AI Systems. Version 1. Zenodo. https://doi.org/10.5281/zenodo.21043361
Skitka, L. J./Mosier, K. L./Burdick, M. D. (1999): "Does Automation Bias Decision-Making?", International Journal of Human-Computer Studies, 51(5), 991–1006. https://doi.org/10.1006/ijhc.1999.0252
von Foerster, H. (1993): KybernEthik. Berlin: Merve.