Authorship Drift in LLM Writing

• Authorship drift in LLM-assisted writing is the detectable shift in stylistic and attributional markers between human and machine-generated text.
• Bayesian classifiers and changepoint detection methods quantify drift, showing significant drops in attribution accuracy after iterative LLM paraphrasing.
• This phenomenon raises ethical, cognitive, and design challenges, prompting strategies like provenance tracking and revised authorship criteria.

Authorship drift in LLM-assisted writing denotes the progressive divergence or redistribution of authorial agency, style, or perceived contribution between the human operator and the LLM during text production. This phenomenon encompasses both measurable stylistic transitions within individual documents and broader conceptual, psychological, and ethical shifts concerning authorship, ownership, and accountability across academic and creative writing. The following sections synthesize the principal empirical, computational, and philosophical insights from recent research to provide a technical overview and analysis of this complex, rapidly evolving subject.

1. Operational Definitions and Measurement

Authorship drift is formally understood as a systematic and detectable shift in the attribution or stylistic footprint of LLM-assisted text relative to the original human author or baseline human-generated writing. At the document level, drift is operationalized as a change point or gradient in the density of words statistically associated with LLM outputs, as determined by a Bayesian classifier trained on uni-gram frequency differences between canonical human corpora and LLM-regenerated texts. The drift is often localized using changepoint algorithms such as PELT (Pruned Exact Linear Time), yielding a segmentation of the text that reveals clustered regions of LLM-like or human-like prose (DeHaan et al., 22 May 2025, Tripto et al., 2023, Wang et al., 2023).

Two primary types of drift are tracked:

• Stylistic drift: The erosion or replacement of the original author’s idiosyncratic style, e.g., through iterative LLM paraphrasing or copy-pasting completions.
• Attributional drift: The migration of perceived or measurable authorship from the human to the LLM, especially when model-generated text dominates or human revision/ownership fades (Tripto et al., 2023, Wasi et al., 2024).

In stylometric attribution tasks, key metrics to quantify drift include:

• Decrease in classifier accuracy, e.g., macro-F1 score of BERT-based models on paraphrased or LLM-adapted texts (sharp drops after just one paraphrase round: e.g., from ≈0.77 to ≈0.33) (Tripto et al., 2023).
• Declines in style similarity (cosine or Mahalanobis distance between feature representations of original and drifted texts).
• Section or revision-level embedding drift (e.g., per-draft l2-norm changes in a universal authorship representation space) (Wang et al., 2023).

2. Psychosocial and Cognitive Dynamics

Research demonstrates that authorship drift is not solely computational but is fundamentally intertwined with users’ evolving psychological states during LLM collaboration. Authorship drift often arises as a function of:

• Declining self-efficacy: Users whose confidence in their ability to write independently decreases over prompts are more likely to cede creative control to the LLM, resulting in higher literal overlap between LLM outputs and final submitted texts (Park et al., 5 Feb 2026).
• Increasing trust in the LLM: As users experience (or perceive) LLM output as helpful, trust typically grows with each turn, which can accelerate drift unless accompanied by sustained self-efficacy.
• Prompting strategy: Users in a 'decrease' self-efficacy trajectory disproportionately issue editing prompts ("edit X for clarity"), leading to adoption of LLM language; those maintaining self-efficacy are more likely to request review and give more substantial feedback, retaining agency and style (Park et al., 5 Feb 2026).

Multi-stage survey data reveal that "sense of ownership" and "authorship" are modulated both by the type of writing (creative vs. routine) and by prompts reminding the user of their contributing role in prompt engineering or editing. Notably, explicit reminders of active intervention (editing) increase the proportion of users who claim ownership of LLM-assisted text from 17% to over 60% (Wasi et al., 2024).

3. Methodological and Computational Frameworks

Empirical studies deploy a variety of methodologies to track and model authorship drift:

• Bayesian Unigram Log-Odds Classifier + PELT Segmentation: Assigns each word a smoothed log-odds of being LLM-generated; detects changepoints where cumulative LLM attribution shifts abruptly. A high penalty parameter (β) required to suppress changepoints indicates partial LLM generation; low β suggests uniform editing or generation (DeHaan et al., 22 May 2025).
• Style Embedding Trajectories: Computes drift scores across revisions/drafts using authorship-sensitive embedding models. Cumulative and per-revision distances can flag abrupt or gradual drift (Wang et al., 2023).
• Iterated Paraphrasing Experiments: Quantifies the effect of repeated LLM paraphrasing on classifier accuracy and style similarity, demonstrating rapid initial drift and then plateauing (Tripto et al., 2023).
• Word Frequency Surveillance: Tracks term frequency over time (e.g., "delve" peaking with LLM adoption, then falling as flagged; "significant" rising steadily), capturing macro-level co-evolution of human and LLM usage (Geng et al., 13 Feb 2025).

Table: Key Computational Approaches for Authorship Drift Detection

Approach	Measurement Granularity	Key Metric/Signal
Log-odds + PELT segmentation	Intra-doc/section	# of changepoints / M_min (seg. penalty)
Paraphrasing iteration cascade	Corpus-level	Drop in macro-F1, S_style, S_cont
Universal Authorship Embeddings	Revision/document	Per-drift Δ_t, cumulative drift D_T
Lexical-frequency surveillance	Population/periodic	f_w(t): normalized freq. per abstract

4. Authorship Drift in Human–LLM and LLM–LLM Collaboration

The presence and character of authorship drift depend on the collaboration scenario:

• Human–LLM Writing: Most academic preprints exhibit minimal authorship drift (measured by lack of changepoints, high inter-section log-odds correlation), implying that when LLMs are used, they are typically applied uniformly for editing or document-wide revision, not for localized or partial generation. This practice reduces risk of hallucinations and preserves document integrity (DeHaan et al., 22 May 2025).
• Distributed LLM Multi-Author Texts: In purely LLM–LLM collaborative writing (e.g., CollabStory), detection of authorship drift (e.g., via shift in vocabulary richness or explicit authorship verification) becomes more tractable as the number of collaborating LLMs increases, but per-segment attribution accuracy degrades as author interleaving rises. Segment-to-segment drift measures (embedding distance) increase with number of collaborating LLMs (Venkatraman et al., 2024).
• Iterative Human + LLM Paraphrasing: Each LLM paraphrase round produces an order-of-magnitude reduction in authorial style recognizability, with macro-F1 in attribution tasks dropping from ~0.77 to ~0.33–0.55 after only a single iteration (Tripto et al., 2023).

5. Philosophical and Policy Dimensions

Authorship drift raises acute philosophical and policy questions:

• The Ship of Theseus Paradox: Successive paraphrasing or machine revision eventually yields a text for which the original human stylistic signature is unrecognizable; the point at which authorship properly transfers to the LLM is nontrivial and may depend on both content preservation and style-signature erosion (Tripto et al., 2023).
• ICMJE and Senior Author Analogy: Under established biomedical authorship criteria, LLM-assisted textual generation supervised and critically reviewed by a human is functionally analogous to senior authorship (conception, review, approval, accountability), suggesting such collaborations merit recognition under current norms—or else the criteria themselves must be revised (Hurshman et al., 5 Sep 2025).
• Ownership vs. Authorship Attribution: Cognitive research reveals a divergence between felt ownership and formal authorship, especially in creative domains, with users often willing to submit LLM-generated content under their name despite reluctance to claim full ownership (Wasi et al., 2024).

6. Interface Design, Provenance, and Mitigation Strategies

To address and manage authorship drift, research proposes:

• Provenance Tracking: Recording and surfacing metadata distinguishing human-typed from LLM-generated spans, color-coding, and logging provenance at the chunk or sentence level (Yukita et al., 22 May 2025).
• Editable Prompting and Reflective Feedback: Emphasis on UIs that require manual translation of LLM suggestions (e.g., rendering AI output as "views" or annotations rather than direct text insertions), integrated prompt editing, and explicit contextual cues for user intervention (Kim et al., 2024).
• PaperCard Framework: Structured declarations at submission time (including per-stage AI involvement, prompt logs, license, and risk disclosures) to make transitions in authorship transparent and auditable (Cho et al., 2023).
• Self-Efficacy Scaffolding: Detecting declines in user self-efficacy and prompting review-oriented (rather than direct-editing) interactions to reinforce active human agency in authorship (Park et al., 5 Feb 2026).

7. Limitations, Open Challenges, and Future Directions

Substantial challenges remain:

• Detection Robustness: Macro-level frequency monitoring can partially adapt to coevolution, but stylometric or embedding models rapidly degrade as LLMs and humans blend techniques, especially given the rise of prompt engineering and adversarial avoidance of flagged markers (Geng et al., 13 Feb 2025).
• Generalizability and Drift across Domains: Evidence from academic preprints may not extend to peer-reviewed journals, code repositories, or creative platforms with different social or stylistic conventions (DeHaan et al., 22 May 2025).
• Ownership Attribution in Highly Personalized and Informal Writing: LLMs struggle to replicate the subtle features of personal and informal style, yielding severe drift in blogs and forums even with advanced few-shot or snippet prompting; mitigating this will require hybrid training regimes and explicit style-preservation controls (Wang et al., 18 Sep 2025).
• Socio-technical Integration: As LLMs become embedded "co-authors" in workplace tools, clarification of credit, accountability, and human oversight at scale will require institutional, cultural, and interface-level interventions (Yukita et al., 22 May 2025, Cho et al., 2023).

In sum, the phenomenon of authorship drift is multifaceted, entangling computational diagnostics, cognitive psychology, interface theory, and academic norms. The trajectory of both LLM technology and professional writing practice will likely hinge on the successful detection, mediation, and transparent disclosure of drift, preserving human agency and accountability amid powerful but easily misattributed generative models.