Abstract
An assistance system can be useful while changing the place at which an author’s decision occurs. That change is easy to miss when the result is fluent, when the interface offers an undo command, or when the action is described as an improvement. This paper introduces substitution drift: the movement from a visible, rejectable proposal toward a system action that determines wording, structure, evidence, or framing before the writer meaningfully encounters it as a choice. The paper offers the Agency Interruption Audit, a ten-field instrument for tracing one assistance episode through user purpose, intervention, proposal visibility, choice moment, alternatives, reason signal, recovery path, attribution cue, unresolved user-research question, and non-conclusion. The instrument neither rates a product nor decides whether a writer retained agency. It makes a more modest contribution: it distinguishes an available recovery action from a meaningful decision opportunity.
1. The three transitions that never arrive
Mara is revising a profile whose argument turns on a difficult hinge. The first half of the piece describes a neighborhood archive. The second half asks why its volunteer keepers distrust an institution offering to “preserve” it. Mara has drafted the first paragraph of the turn, crossed it out twice, and left a note in the margin: Need three ways of moving from care to custody without making the archive sound naïve.
She selects the marked passage and asks her revision assistant a plain question: “Give me three possible transitions.”
The answer arrives already inside the document. A polished bridge now stands between the paragraphs. The next paragraph has been moved below it. The final sentence of Mara’s archive description has been shortened, and a side notice says: Flow improved; repeated theme consolidated. There is no list of three transitions. There is no small gallery of alternatives, no view in which the old order sits beside a proposed order, and no explanation of why custody rather than care became the governing term. The draft simply reads more smoothly.
Mara can undo the change. That fact matters. An irreversible alteration would be a different and more serious condition. But undo is not the only question raised by the episode. Did the assistant answer the request Mara made? Did it offer transitions, or did it choose one? Did it assist a local sentence problem, or did it resolve an editorial relation between paragraphs? When did Mara have an opportunity to see that the order had changed? What would she have had to notice, reconstruct, and restore before she could return to the comparison she originally requested?
None of these questions requires a claim that the inserted passage is bad. It may be better than Mara’s draft. It may name a relation Mara had not yet found. It may save her useful time. The problem begins elsewhere: an interface can turn a request for material to consider into an action whose result is already installed. The difference is not captured by asking whether the writer still has a button labeled undo.
This paper calls that movement substitution drift. The term does not mean that any automation substitutes for a person, or that a system has taken authorship of a text. It names a more precise interactional possibility. An assistance system begins with a visible, rejectable option and then, by degrees, supplies a selection, a revision, a reordered structure, an unmarked omission, or a normalized framing that the writer does not meaningfully meet as a choice. The drift may be small. It may happen in a comma correction, a source summary, a paragraph move, or a sentence completion. It becomes more consequential when the intervention changes the work the writer thought they were doing.
A request for “three transitions” has a shape. It asks for alternatives. It presumes that comparison remains with the writer, even if the system generates the material. A request for “fix the repeated punctuation” has a different shape. It may authorize a predictable, low-consequence operation in which repeated review would add little value. The distinction is not whether one request is human and the other automated. Both involve a person working with an assistance system. The distinction is the relation between the requested help, the intervention actually made, and the point at which a person can still recognize and decide about that difference.
2. A help request contains a shape of help
The phrase assist the writer is too broad to teach with. It can include a prompt that catches a repeated word, a list of alternative transitions, a new paragraph placed into a draft, a source summary, an explanation of why a sentence seems ambiguous, a rearranged outline, or a revised argument. These acts are related but not equivalent. They distribute attention, initiative, and editorial consequence differently.
A writer’s request provides one starting point for that distinction. It does not supply a perfect contract. Writers ask imprecisely. They change their minds. They may say “make this flow” when what they really need is a stronger conceptual relation. An assistant cannot avoid inference merely because language is imperfect. Yet an imperfect request is not permission to ignore the requested form of help. A request for a list, a diagnosis, a question, a comparison, a mark-up, or a drafted passage carries a different expectation about what the writer will see and decide.
Human–AI interaction guidance gives names to the pieces that often vanish inside the word help: timing a service in relation to a current task, allowing efficient dismissal and correction, narrowing a service when the system is uncertain, making a reason available, and providing controls over system behavior.[1] Those guidelines do not prove that a particular interaction treats a writer well. They are not a scorecard for agency. They are useful because they move the conversation from a flattering noun—assistant—to observable features of an episode.
Mixed-initiative work provides a related model. In a guide–decide relation, a person’s action helps make an assistance system’s next move relevant; the system’s suggestion is made assessable; and the person may accept, refine, or dismiss it.[2] That model does not require a system to wait for a person at every moment. It does require a distinction between helping to surface a possible move and quietly selecting the move on the person’s behalf.
Mara’s incident is not troubling because the machine wrote a sentence. A writer might deliberately ask for a sentence and welcome one. The episode is diagnostically interesting because the requested object was a set of alternatives and the returned object was a document-level decision. The assistant has changed both the relation between paragraphs and the relation between Mara and the options she asked to inspect.
This leads to a first working question:
What form of help did the writer ask to encounter before deciding what to do next?
The question is more exact than “Did the user consent?” It also avoids the opposite error of requiring ceremonial confirmation for trivial, familiar, or access-supporting operations. A user might reasonably ask for spelling normalization to occur with no interruption. The writer may even establish a standing preference for it. The audit asks whether the intervention has crossed a decision boundary that the writer retained—not whether all automation has crossed one.
3. Recovery is a path, not the whole choice
A conventional editing interface already shows why recovery and choice need to be separated. Google’s support record for suggesting edits describes changes that remain visually distinct and leave original text unchanged unless an owner approves them.[10] Microsoft’s tracked-changes record describes review through visible changes, next/previous navigation, and accept/reject actions; it also distinguishes hiding a change from actually resolving it.[10] These are primary documentation records for named systems, not evidence that either system preserves agency or should be adopted. Their narrow value is mechanical: editorial interfaces can make an alteration legible before it becomes ordinary text.
That mechanism has at least four parts.
First, proposal visibility asks what the writer can perceive before an intervention changes the working draft. A suggestion in a margin and a change already written into the document are different states. Visibility does not guarantee comprehension, but a writer cannot compare what is never made available.
Second, the choice moment asks when a writer can act. A choice before insertion differs from a choice after insertion; a choice after an unmarked paragraph move differs from a choice after a marked one. Time matters because an installed revision can become the new cognitive starting point. Once a bridge is in place and the paragraph order reads smoothly, restoring the original question—what were the other ways to do this?—requires mental reconstruction.
Third, available alternatives asks whether the interface retained something with which the intervention could be compared. The relevant comparison may be the untouched original, two generated options, a diagnosis without a replacement, or an ability to request a distinct formulation. It is not a requirement to display endless variants. More options can obscure rather than support judgment. The issue is whether “alternatives” names a real relation in the episode or merely a menu label.
Fourth, the reject-and-recover path asks how a writer could refuse, restore, edit, or revisit an intervention. It matters that Mara can undo. It also matters whether undo restores the original order, whether it preserves the assistant’s alternatives, whether it can be found in the moment, and whether a writer must understand a technical history to recover a prior state. Recovery is a genuine protection. It is not evidence that the choice was initially meaningful.
The distinction has a quiet implication. A system may provide robust undo and still create substitution drift. It may let a writer restore the old document while making it difficult to revisit the reasoning, options, and framing that the system selected. Conversely, a system might place a visible proposal into a document while offering a poor recovery path. Agency is not a switch that appears when one control is present. It is a relation among the writer’s purpose, the intervention, the available evidence, the moment of decision, and the cost of declining the result.
4. Friction has a cost, and invisibility has one too
It would be easy to answer Mara’s case with a crude rule: make every assistance action wait for approval. The rule is attractive because it seems to return authority to the writer. It would also be impractical and, in some contexts, needlessly burdensome.
Research on automation bias offers a warning against simple stories in either direction. Goddard, Roudsari, and Wyatt’s systematic review concerns clinical decision-support systems, not writing. Within that bounded setting, it identifies over-reliance as a problem shaped by factors including experience, trust, workload, task complexity, time pressure, and advice presentation; it also notes that decision-support systems may improve performance while creating new error modes.[3] The paper cannot tell us that writers will follow an AI revision, or how often. It does show why a visible recommendation cannot be assumed to be neutrally received.
Later decision-making studies sharpen the point about cost. Buçinca, Malaya, and Gajos found that cognitive-forcing interventions reduced over-reliance relative to simpler explainable-AI approaches in their experiment, while receiving less favorable subjective ratings and benefiting people differently depending on need for cognition.[4] Vasconcelos and colleagues report across five studies that task difficulty, explanation difficulty, and incentives affect reliance; their account treats engagement with an explanation as a strategic cost–benefit decision rather than an automatic cure for over-reliance.[5]
Those studies do not prescribe a pause button for writers. Their settings and tasks are different. Their lesson is methodological: a reader should not assume that a delay, an explanation, a confirmation, or a review panel has the same effect in every circumstance. Verification takes time. Explanations can be hard to interpret. Repeated prompts can fracture concentration. A design that protects one kind of consideration may create another kind of interruption.
A counterexample helps. A copyeditor who has deliberately configured a document to standardize a recurring punctuation error may reasonably prefer a quiet, reversible correction to a confirmation dialogue after each instance. A writer using a language-support tool may value a rapid grammar suggestion because it lowers a barrier to finishing a draft. Professional guidance from the American Psychological Association describes possible uses of AI for routine tasks, grammar, manuscript-finalization steps, and language-related assistance while also emphasizing verification of output.[9] That guidance is not an accessibility efficacy finding and does not settle an individual workflow. It nevertheless prevents the audit from treating attentiveness as an unlimited resource.
The aim is proportionality. A request that authorizes a stable, low-consequence correction may support more automation than a request that calls for interpretive alternatives, strategic reordering, source selection, or editorial framing. The difference must be discovered in the episode, not imposed by a universal hierarchy of “human” and “machine” tasks.
5. The Agency Interruption Audit
The Agency Interruption Audit is designed for a writing-tool product steward, editorial-workflow designer, publishing-program teacher, commissioning editor, or researcher planning an AI-assisted revision exercise. It is performed on one episode, not on a brand or an institution. The audit begins before anyone praises or condemns the output.
| Field | Reader records | Why it matters | It does not decide |
|---|---|---|---|
| 1. User purpose | The requested help in the writer’s own terms. | Separates requested assistance from a later system action. | Whether the request was perfectly specified. |
| 2. System intervention | What the system generated, inserted, selected, reordered, marked, or withheld. | Makes the event concrete. | Whether the outcome is good writing. |
| 3. Proposal visibility | What the writer could see before the action affected the working text. | Locates the first decision opportunity. | Whether visible text was understood. |
| 4. Choice moment | When the writer could decide relative to the intervention. | Distinguishes proposal, confirmation, and after-the-fact notice. | Whether any timing is universally correct. |
| 5. Available alternatives | What distinct options, original text, or comparison views remained available. | Shows whether “alternative” meant more than one answer. | The ideal number of options. |
| 6. Basis/reason signal | What explanation, source cue, goal relation, or uncertainty signal accompanied the action. | Lets the writer assess what the system claims to have done. | That a reason is true, sufficient, or persuasive. |
| 7. Reject-and-recover path | How the writer could refuse, reverse, restore, or revise the action. | Separates recoverability from initial visibility. | That undo repairs every lost decision moment. |
| 8. Attribution cue | How the interface identifies a suggestion, automated action, or human-text relation. | Makes the intervention legible within the episode. | Authorship or responsibility. |
| 9. Unresolved user-research question | What must be studied with relevant users: timing, comprehensibility, interruption cost, access, trust, or felt control. | Prevents interface inspection from becoming an agency verdict. | A user outcome. |
| 10. Non-conclusion | What the episode cannot certify. | Holds the audit inside its evidence boundary. | Product quality, ethics, or policy compliance. |
The fields are deliberately asymmetrical. Several identify what is present in an interface. Two make absence explicit. The user-research field prevents a reviewer from mistaking a screenshot, a documentation page, or a carefully reasoned walkthrough for evidence about how writers actually understand an intervention. The non-conclusion field blocks a second temptation: turning a descriptive trace into a verdict about authorship, responsibility, quality, or the moral standing of automation.
The instrument also gives a more careful account of explanation. A basis/reason signal may say, “This transition repeats a phrase,” “This paragraph was moved to strengthen chronological order,” or “The system is uncertain between two organizational patterns.” Such a signal can help a writer recognize an intervention’s claimed relation to the draft. It does not make the intervention correct. A confident but opaque explanation may be worse than an explicit uncertainty cue. A terse but useful reason may serve better than a long narrative that a writer will not have time to inspect.
6. Worked trace: the transition and the moved paragraph
The following trace does not grade Mara’s assistant. It records what can be said from the constructed incident and where investigation would be needed.
| Ledger field | Constructed entry | Editorial reading |
|---|---|---|
| User purpose | “Give me three possible transitions” between care and custody. | The purpose names comparison, not a single installed bridge. |
| System intervention | A bridge is inserted; the following paragraph is moved; one sentence is shortened. | The intervention exceeds sentence suggestion and includes structural framing. |
| Proposal visibility | The constructed case gives no evidence that the bridge or new order was visible before insertion. | A small later notice is not proof of pre-change visibility. |
| Choice moment | The writer first sees a finished result in the working draft. | The decision opportunity arrives after the document has been altered. |
| Available alternatives | No three-option set is visible; the original order may be recoverable only through undo. | The requested comparison is not demonstrated in the episode. |
| Basis/reason signal | “Flow improved; repeated theme consolidated.” | The notice identifies a broad claimed purpose but does not disclose why this transition or order was selected. |
| Reject-and-recover path | Undo is available; the exact restoration path and its durability are not observed. | Recovery is a real protection, but it does not answer the original request for alternatives. |
| Attribution cue | A side notice identifies an improvement action, not the internal relation among generated and human text. | The cue may make an intervention visible without locating its scope clearly enough for review. |
| Unresolved user-research question | Would relevant writers notice the paragraph move, understand the notice, value the speed, or experience the change as a loss of choice? | No interface inspection can answer this. A study must ask and observe. |
| Non-conclusion | The trace does not establish that Mara lacks agency, that the revision is poor, or that the system is unsuitable. | The proper conclusion is partial evidence of a choice mechanism; no conclusion about agency outcome. |
The trace makes a narrow diagnosis possible. The assistant may have supplied useful language while drifting away from the requested interaction form. The drift occurs at the interval between request and installed result. It is visible without calling the assistant deceptive, harmful, or authorial. The audit exposes a design question: if a system receives a request for alternatives, should it preserve alternatives, make an insertion conditional, or state that it has changed the task? That question remains open until relevant writers and contexts are studied.
Studies of AI-assisted writing give further reason not to rush past that remainder. Co-Writing with AI, on Human Terms reports a PRISMA review of 109 HCI papers and interviews with 15 writers; it distinguishes planning, translating, reviewing, and monitoring, and finds that desired levels of intervention vary across process, goals, values, and conceptions of ownership.[6] The study does not settle which assistance level is right for Mara. It makes a broad claim of uniform preference untenable.
The ABScribe study is also suggestive without being definitive. It treats exploration of alternative wording as integral to writing and examines an interface that keeps variations adjacent for comparison rather than overwriting text. Its 12-writer study reports lower task workload and improved perceptions of the revision process relative to the selected baseline.[7] That finding does not prove that multiple variations preserve agency, nor does it specify the number that a writer needs. It supports a smaller possibility: how alternatives are preserved and organized is an interaction design question, not an ornamental feature.
Ownership research adds another caution. The AI Ghostwriter Effect study reports, in two study samples, that greater participant influence on text was associated with higher perceived ownership in that study setting.[8] It cannot decide who is an author, and this paper does not use it for that purpose. It reinforces the need to distinguish an interface’s visible output from the writer’s experienced relation to it.
7. From audit to research brief
A completed audit should not end with a sticker that reads agency preserved or agency lost. Its final value is to tell a design or editorial team what must be investigated next.
| If the audit finds… | The next research question is… | A proportionate comparison could be… |
|---|---|---|
| A finished change appears before the requested alternatives. | Do writers understand that the task changed from option generation to revision? | A visible three-option view versus an inserted default bridge. |
| An action is marked but the reason signal is generic. | Can writers identify the scope and rationale of the intervention in time to assess it? | A general “improved flow” notice versus a scope-specific rationale with an uncertainty cue. |
| Undo exists but restoration is cognitively or procedurally costly. | What must writers reconstruct to return to their original decision space? | Simple restoration of original text/order versus a history view that preserves alternatives and reasons. |
| Repeated confirmations interrupt a low-consequence task. | When does review support a writer, and when does it exhaust attention or access? | Default-on reversible correction with a global control versus per-instance confirmation. |
| An interface offers several alternatives. | Which comparisons are genuinely useful for the writer’s purpose, and which become noise? | Adjacent variations, a short ranked set, and an on-demand alternative request. |
This is not a laboratory protocol. It is a way of writing better questions before claiming a design has solved a human problem. A team might recruit writers who work in the relevant genre, observe revision episodes with representative material, ask writers to describe the help they expected, and compare what they recognized before and after an intervention. It might measure completion time, but it should not confuse speed with control. It might collect preference statements, but it should not assume that a preference in a brief task explains the meaning of a revision in a long project.
The audit also makes a useful teaching exercise. A publishing-program instructor can show three versions of the same constructed interaction: first, three visible transitions next to untouched text; second, one inserted transition with a clear tracked-change record; third, a silent insertion with a later notification. Students can trace the three episodes without deciding which is morally correct. They can then ask which field changed, which consequence depends on context, and what evidence would be needed before making a claim about writers’ experience.
The point is not to discipline writers into perpetual suspicion. It is to preserve a vocabulary for seeing where a system has already acted. Fluent assistance can make its own intervention difficult to notice precisely because it decreases visible friction. An editorial practice needs ways to decide when that smoothness is a gift and when it quietly rearranges the writer’s work.
8. A decision can vanish before anyone rejects it
Mara’s problem is not that a machine offered language. Writers have always worked with suggestions—from colleagues, editors, dictionaries, templates, and their own earlier drafts. The problem is that she asked for a particular relationship to language: three possible transitions that she could set beside the passage and judge. The assistant returned a relationship already chosen.
An undo control may allow Mara to travel backward. It does not automatically restore the moment before the assistant chose the category of help, selected a bridge, or moved a paragraph. A visible explanation may help. A marked change may help. Several alternatives may help. A quiet automatic correction may be exactly right elsewhere. None of these features can, alone, establish that a writer remained in command of a decision.
The Agency Interruption Audit therefore refuses both comfort and alarm. It does not presume that assistance erases a writer, and it does not treat a later recovery control as proof that no decision was displaced. It asks a reader to hold the episode still long enough to see the request, the intervention, the moment of choice, the alternatives, the recovery path, and the uncertainty that remains.
Did the system offer a choice, or did it make one and leave only an afterimage?
A paper, a tool review, or an editorial workflow should not answer that question on behalf of writers. It should create the conditions in which the relevant writers can answer it with evidence.
References
[1] Microsoft Research. “Guidelines for Human-AI Interaction Design.” 2019.
[2] PNAS. “Agency plus automation: Designing artificial intelligence into interactive systems.” 2019.
[3] Kate Goddard, Abdul Roudsari, and Jeremy C. Wyatt. “Automation bias: a systematic review of frequency, effect mediators, and mitigators.” Journal of the American Medical Informatics Association 19, no. 1 (2012): 121–127.
[4] Zana Buçinca, Maja B. Malaya, and Krzysztof Z. Gajos. “To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making.” Proceedings of the ACM on Human-Computer Interaction (2021).
[5] Helena Vasconcelos et al. “Explanations Can Reduce Overreliance on AI Systems During Decision-Making.” Proceedings of the ACM on Human-Computer Interaction (2023).
[6] “Co-Writing with AI, on Human Terms: Aligning Research with User Demands Across the Writing Process.” ACM Digital Library, 2025.
[7] “ABScribe: Rapid Exploration & Organization of Multiple Writing Variations in Human-AI Co-Writing Tasks Using Large Language Models.” ACM Digital Library, 2024.
[8] “The AI Ghostwriter Effect: When Users Do Not Perceive Ownership of AI-Generated Text but Self-Declare as Authors.” ACM Digital Library, 2024.
[9] American Psychological Association. “The promise and perils of using AI for research and writing.” 2024.
[10] Google Docs. “Suggest edits in Google Docs.”; Microsoft Support. “Accept or reject tracked changes in Word.”. Accessed 20 August 2026.