DraftMarks: Inline Transparency for AI Writing

• DraftMarks is an augmented reading system that embeds process evidence using visual cues to reveal both AI and human contributions inline.
• It utilizes familiar editing metaphors, such as masking tape and eraser crumbs, to represent text insertion, revisions, and deletion for clear provenance.
• The system employs a model-view-controller architecture to capture keystroke events and adapt transparency for different stakeholders like teachers, reviewers, and general readers.

Searching arXiv for DraftMarks and closely related watermarking / transparency work to ground the article. DraftMarks is an augmented reading system for AI-assisted writing that makes the writing process visible inside the final document itself. It is designed for human-AI co-writing settings in which readers can no longer assume that a finished text directly reflects only human thought and effort. Rather than requiring readers to leave the document and inspect disclosures, chat logs, prompt histories, token attribution maps, or summary dashboards, DraftMarks embeds process evidence inline through visual encodings borrowed from the material world of writing and editing. Its contribution is a document-embedded transparency layer that combines interaction logging, provenance-aware document modeling, stakeholder-specific rendering, and empirical evaluation of how readers interpret AI-assisted text (Siddiqui et al., 27 Sep 2025).

1. Conceptual basis and transparency problem

DraftMarks is motivated by a change in the relationship between writer and reader under generative AI. The paper argues that readers need to understand not only what a text says, but also how it was produced, how much revision occurred, and how much of the intellectual labor came from the human versus the model. This need is framed differently for different readerships. Teachers want evidence of student thinking and productive struggle, because writing is part of learning rather than only a finished product. Reviewers care about intellectual contribution and may want to know whether AI shaped the ideas or merely polished the prose. General readers care about authenticity, trustworthiness, and whether they are reading a genuine human voice (Siddiqui et al., 27 Sep 2025).

A central claim of the system is that existing transparency approaches are too detached from reading. Disclosures, chat logs, prompt histories, token attribution maps, and summary dashboards can provide provenance information, but they force the reader to step away from the text to make local judgments. DraftMarks instead treats transparency as something that should remain legible during ordinary reading. This suggests a shift from provenance as external metadata to provenance as a property of the reading surface itself.

The system therefore does not merely expose that AI was used. It attempts to reveal where AI entered the document, where it was revised, what was deleted, and how prompt-driven iteration unfolded. A plausible implication is that DraftMarks belongs less to the tradition of watermarking and more to the tradition of process visualization, although it intersects with provenance research by making authorship and revision history inspectable at passage level.

2. Skeuomorphic process traces and visual semantics

The core design idea is “skeuomorphic process traces”: familiar physical-editing metaphors are used to encode hidden collaboration history inside the final artifact. The metaphor is that a final draft still carries evidence of drafting, much as a paper manuscript might show erasures, tape, notes, or overwritten lines. The paper emphasizes that these are not decorative icons; they are intended to encode process history in a way that readers can interpret through everyday document-literacy (Siddiqui et al., 27 Sep 2025).

Masking tape marks passages initially generated by AI. A single strip means AI inserted a contiguous chunk of text, while stacked masking tape indicates iterative generation at the same location. Scrunched masking tape represents human deletions within AI-generated text, and torn masking tape represents human insertions into AI-generated text. Segmented masking tape preserves phrase-level distinctions between what came from the prompt and what was newly generated by the model.

Smudge marks represent AI transformations of existing text for tone or meaning rather than new content insertion, and segmented smudges preserve phrase-level distinctions between original prompt text and AI-generated text. Eraser crumbs indicate the prompt trail adjacent to AI marks, functioning as a trace of the “virtually erased” prompt that produced the current passage. A solid crumb is a simpler prompt marker, whereas a density-varied crumb encodes prompt complexity through shades of gray; darker crumbs suggest more complex prompts.

Residual glue shows that AI text once appeared there but was later removed entirely by the writer. A single glue mark shows the most recent discarded generation, and a sequenced glue mark preserves multiple discarded generations. Font changes reinforce origin, using handwritten or script style for human-authored text and sans-serif for AI-generated text. Stencil marks indicate structured use of AI feedback: a single stencil mark marks one feedback point, a layered stencil mark indicates multiple feedback generations over the same passage, dotted strokes imply that the writer did not fully integrate the feedback, and lined strokes imply integration. Ghost text shows prompts or ideas that were generated but not included in the final draft, and it can reveal just the instruction or the full prompt with context.

These marks can overlap. A passage may be both masking tape and smudged, meaning AI inserted text that was later tone-adjusted. This layered visual grammar is one of the system’s main distinctions: it does not encode only origin, but also revision state, prompt lineage, and acceptance or rejection of AI contributions.

3. Data model, provenance capture, and versioning logic

DraftMarks is implemented as a model-view-controller system. The model stores the writing process through version-controlled editor states. The system extends a rich text editor’s node-based document structure so that each text node can store provenance fields: author for human or AI, prompt for the instruction or context used for generation, and generated text. It also stores “orphan” AI text nodes, meaning AI-generated outputs that were queried but never inserted into the writing canvas (Siddiqui et al., 27 Sep 2025).

The capture mechanism is event-driven rather than time-based. A new version is created only when one of three events occurs: an AI-authored text node is inserted, an AI-authored text node is fully removed, or 10 or more characters are deleted from an AI-authored text node. Human revisions that do not meet those criteria remain within the current version. This matters because the system is not attempting to record every keystroke equally; it is attempting to capture collaboration-relevant transitions that reveal where AI entered or left the draft.

On the capture side, the composer module listens to editor state changes and records keystroke-level data. Text that can be mapped to local keystrokes is treated as human, while text that cannot is treated as AI-generated. Copy-pasted text from outside the local app is also treated as AI-generated. The paper presents this as a heuristic that can approximate ground-truth writing history across split workflows such as Google Docs plus ChatGPT, integrated tools such as ScriptShift or ABScribe, and ambient assistants such as Grammarly. User permission is required, reflecting the privacy concerns explicitly discussed in the paper.

The rendering pathway converts these versioned editor states into a Process Schema, a nested structure optimized for a given reader type. Three controller components are responsible for this: the Intent Mapper, which sets stakeholder-specific priorities; the Trace Aggregator, which decides how much history to keep and at what granularity; and the Trace Annotator, which decides which visual mark to apply. This architecture makes the system not just a visualization layer but a transformation pipeline from raw interaction history to role-specific interpretive output.

4. Stakeholder-dependent interpretation and readership logic

The controller logic is explicitly stakeholder-dependent. Teachers receive more detailed traces, reviewers receive sparse summaries, and general readers receive a middle ground emphasizing authenticity and effort. These distinctions were derived from a formative study rather than imposed a priori, and the paper presents them as computational logic for different readership (Siddiqui et al., 27 Sep 2025).

For teachers, the system foregrounds process visibility for formative assessment. DraftMarks is intended to show struggle, revision, scaffolded learning, and whether the writer engaged with the task rather than delegated it wholesale. For reviewers, the preferred representation is lighter: reviewers often do not want a heavy process view, but may want quick signals about intellectual contribution, idea generation, or whether AI was used mainly for polishing. For general readers, the key value is authenticity and trust; they want transparency sufficient to judge whether a piece feels genuinely authored, heavily delegated, or iteratively refined.

This stakeholder stratification addresses a common misconception that provenance disclosure is a single universal interface problem. DraftMarks argues instead that process visibility should vary with the interpretive goals of the reader. A plausible implication is that provenance systems for human-AI collaboration may need audience-scoped rendering policies rather than a single fixed representation.

The system also treats authors as a relevant audience. The paper suggests that DraftMarks can function as a metacognitive mirror, helping writers reflect on whether they are using AI as a scaffold or as a substitute for their own thinking. This is presented not as a claim about automatic behavior change, but as an implication of making process visible.

5. Formative study and effectiveness evaluation

The formative study involved 21 participants: 7 teachers, 7 academic reviewers, and 7 general readers. Participants were shown a baseline DraftMarks prototype as a design probe and were asked to think aloud while assessing AI-assisted writing relevant to their role. They were then shown the actual writing process video so the researchers could identify what additional information participants wished they had seen. The qualitative analysis used reflexive thematic analysis, beginning with six transcripts to build a codebook and then refining it across all transcripts with an independent rater. The final codebook was reduced to 16 top-level codes after discussion. The resulting distinctions informed the controller mappings (Siddiqui et al., 27 Sep 2025).

The paper then reports a between-subjects study with 70 participants recruited on Prolific, split evenly between DraftMarks and a baseline. Participants were professionals who regularly work with writing, including teachers, journalists, and copywriters. The task was to read a 600-word essay about social media reform for children’s use, written with ChatGPT, and answer comprehension questions about both content and collaboration process. The baseline condition showed the essay with AI-highlighted text plus a chat log in split view; the treatment condition used DraftMarks.

The evaluation suggests that DraftMarks improved comprehension of the collaboration process. Aggregate comprehension improved from 2.86 to 4.29 out of 7, with significant gains on questions about where changes occurred and how AI feedback was integrated. Usability was also reported as strong: the system scored 80.5 on the System Usability Scale. Cognitive load findings were described as moderate intrinsic load, low extraneous load, and high germane load, suggesting that the representation was informative without being overwhelming. On transparency self-report measures, DraftMarks scored slightly higher than the baseline on confidence, sufficiency of information, and perceived human effort. Qualitatively, many participants reported that the marks helped them distinguish what AI contributed, what the writer changed, and what the writer ultimately rejected.

These results should not be conflated with a claim that DraftMarks solves attribution or authenticity in the strong forensic sense. The paper evaluates understanding of process, not immutable certification. What it does show is that embedded traces can improve readers’ ability to answer process-oriented questions compared with a baseline centered on highlighting plus chat logs.

6. Limitations, tensions, and relation to adjacent marking research

The paper emphasizes a tension between transparency and authorial agency. Making process visible may help readers, but it can also expose private or potentially embarrassing aspects of drafting, such as minimal revision or heavy AI dependence. This raises the question of whether writers should be able to redact, annotate, or audience-scope traces. The system also does not capture AI training-data provenance or citation-level accountability. It can show that AI generated or transformed text, but not where the output came from or whether it echoes copyrighted material. Additional limitations include the approximate nature of its authorship heuristics, especially in mixed workflows involving copy-paste, ambient assistants, or complex editing behaviors, and the possibility that skeuomorphic cues may be less legible for readers with limited exposure to paper-based revision practices (Siddiqui et al., 27 Sep 2025).

Future directions named in the paper include more configurable audience-specific modes, more culturally neutral alternatives to skeuomorphism such as typographic or color-based cues, writer-facing reflective tools, and a “forward-looking linkography” idea for tracing emerging ideas and gaps during co-writing. The authors also argue that future process-visualization systems should incorporate citation transparency in addition to collaboration transparency.

Within the broader landscape of marking systems, DraftMarks occupies a distinct position. UniMark is described as a multimodal content governance toolkit that unifies Hidden Watermarking and Visible Marking across text, image, audio, and video, with a dual-operation strategy and standardized benchmarking (Li et al., 13 Dec 2025). PostMark, by contrast, is a modular post-hoc black-box watermark for LLM-generated text that rewrites already-generated output to incorporate semantically selected watermark words, emphasizing detectability and robustness to paraphrasing rather than process legibility (Chang et al., 2024). PromptMark applies a black-box, prompt-guided watermarking framework to source code by inducing statistically detectable naming patterns under API-only constraints (Fahad et al., 18 Jun 2026). These systems address provenance, copyright protection, compliance, or attribution, but they do not attempt to surface the human-AI writing process inline inside the final artifact.

This suggests a useful distinction. DraftMarks is not primarily a hidden watermarking framework, a visible compliance marker, or a decoding-time provenance channel. It is an augmented reading system for interpreting collaboration history. Its central object is not a hidden signal embedded for later detection, but a stakeholder-specific reconstruction of writing process rendered through inline visual metaphors. In that sense, it extends provenance discourse from certification of origin toward interpretation of effort, revision, and collaborative agency.