Authored: Rethinking Work Attribution

• Authored is the attribution of creative works, defining production, control, originality, and evidentiary trace across both human and machine contributions.
• Methodologies range from stylometric analysis and spatial patch aggregation to process-based logging, demonstrating diverse approaches to identifying authorship.
• Legal and collaborative frameworks now emphasize thresholded human contribution, challenging traditional notions of originality in mixed human-AI outputs.

“Authored” denotes attribution of a work, or of the relevant aspects of a work, to the entity or entities that produced it. In current research, the term extends well beyond signed human texts. It now spans copyright doctrine, academic governance, stylometric attribution, process-based verification, provenance analysis for black-box LLM outputs, and mixed human-machine production in art, code, and scientific writing. Across these literatures, authorship is treated as a context-dependent relation among production, control, originality, and evidentiary trace (Räz, 2024, Mei, 2024, Pereira, 6 Apr 2026).

1. Conceptual scope of authorship

A general formulation appears in the philosophical and political analysis of LLM watermarks: the author of a text is “the entity who produced relevant aspects of the text,” where production is not mere reproduction and relevance is determined by the kind of text and the context (Räz, 2024). This definition is deliberately weak. It does not require intentionality or personhood, and therefore can accommodate both human and machine candidates for authorship. Its significance lies in separating the attribution question from assumptions that only human-like agency can matter.

Technical work operationalizes “authored” in narrower ways. In binary document classification, the task is to decide whether a document is single-authored or multi-authored, with labels $y \in \{0,1\}$ and predictions $\hat y = f(x)$ over a document representation $x$ (Zamir et al., 2023). In multi-class authorship identification, the task is to map a document feature vector to one of $50$ possible authors in a standard text categorization setting (Iyer et al., 2019). In collaborative painting, authorship is spatial rather than document-level: patch predictions are aggregated into whole-painting labels, and mixed regions are treated as genuinely ambiguous rather than simply misclassified (Chen et al., 19 Feb 2026). In black-box LLM provenance, the problem is source-model identification under arbitrary, user-chosen prompts, where prompt semantics dominate surface form and model-specific traces are weak (Liu et al., 9 Jun 2026).

This breadth matters because it shows that authorship is not a single invariant property. It may refer to legal entitlement, intellectual responsibility, stylistic signature, behavioral provenance, or source attribution under observational constraints. A plausible implication is that disputes about “who authored” a work often arise from moving between these distinct senses without making the shift explicit.

2. Human authorship, originality, and legal thresholds

Under the U.S. Copyright Office’s March 16, 2023 “Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence,” a “work of authorship” must be an original work of human authorship, fixed in a tangible medium, and independently created with at least a modicum of creativity. The Office states that when “an AI technology determines the expressive elements of its output,” the generated material “is not protected by copyright and must be disclaimed in a registration application.” It therefore applies a two-step test: determine whether traditional elements were “conceived and executed by a machine,” and if so, refuse registration of those elements while registering only any human-authored selection or arrangement (Mei, 2024).

Mei challenges this position by arguing that the Guidance misdescribes the interaction between user and model. In text-to-image systems, users iteratively craft and refine prompts, evaluate each returned image, select, crop, recombine, and post-process outputs, and repeat this cycle of “adjustment, refinement, selection, and arrangement” until the image matches their vision. On this view, generative AI is neither purely random nor fully autonomous, and controlled chance does not vitiate human authorship any more than wind or gravity does in the analogies to Jackson Pollock or Tim Knowles (Mei, 2024). Mei therefore proposes a simplified registration regime centered on whether AI was used in the creative process and whether that use was substantial and human-directed.

European analysis reaches a related result through a different doctrinal path. Pereira treats authorship as a qualitative threshold rather than a binary status. European copyright protects only works that are the product of a natural person’s “own intellectual creation,” requiring free and creative decisions about form, structure, and expression, and creative freedom that reflects the author’s personality. This is rendered as a threshold predicate

$$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$

and, in a second framing, as attribution when meaningful human engagement $X_h$ exceeds a critical $\theta$ (Pereira, 6 Apr 2026). The paper’s five exclusionary indicators—failure of explainability, lack of intellectual control, substantial identity with unmodified AI output, substitutability by identical prompts, and prompt-only contribution—mark the point at which AI displaces creative autonomy.

Taken together, these approaches treat authorship less as the absence of automation than as the presence of human intellectual control above a legally relevant threshold. This suggests a convergence around thresholded human contribution, even though the U.S. discussion centers on registration and disclaimer while the European discussion centers on “own intellectual creation.”

3. Co-authorship and the redistribution of human labor

In scientific writing, AI assistance has been described as a shift from computers as deterministic tools to LLMs as “virtual collaborators.” In a case study of drafting a computational physics manuscript, the human author adopts a Human-in-the-Loop workflow and an “Inside-Out Writing Strategy”: context loading, conversational genesis of core concepts, and drafting followed by expert verification. The AI drafts structure, syntax, LaTeX, and figure prompts, while the human Principal Investigator catches physics inaccuracies, enforces modern taxonomy, manages academic diplomacy, and anticipates peer review (Zhou, 5 Apr 2026). The central claim is not that authorship disappears, but that the human contribution migrates from boilerplate production to high-level intellectual steering.

In software engineering, large-scale public evidence of mixed authorship appears in AgentPack, a corpus of $1{,}337{,}012$ commits/PRs co-authored by Claude Code, Codex, and Cursor Agent across public GitHub projects up to mid-August 2025. The curation pipeline identifies agent-tagged activity, keeps only commits that landed on the primary branch, removes edits under node_modules/, and joins GitHub metadata with cleaned diffs. Median patch spans are $\sim 70$ lines, median number of files touched per commit is $2$–$\hat y = f(x)$0, and median messages are $\hat y = f(x)$1–$\hat y = f(x)$2 characters, roughly $\hat y = f(x)$3–$\hat y = f(x)$4 longer than prior human-only datasets (Zi et al., 26 Sep 2025). Because maintainers merge these changes, the corpus functions as an implicitly human-filtered record of accepted human-agent collaboration.

Authorship also has a motivational dimension. In a preregistered experiment on goal-setting, self-authored goals and LLM-authored goals derived from a personal reflection were compared. LLM-generated goals scored higher on SMART criteria with $\hat y = f(x)$5, but participants in the LLM condition reported lower psychological ownership ($\hat y = f(x)$6), commitment ($\hat y = f(x)$7), and perceived importance ($\hat y = f(x)$8). At two-week follow-up, $\hat y = f(x)$9 of self-authored participants had acted on two or more of their goals, compared to $x$0 in the LLM condition (Chi et al., 12 May 2026). Mediation analyses identified psychological ownership as the mechanism, while objective goal quality did not mediate downstream motivational and behavioral outcomes. The result is a direct challenge to the assumption that better-formed AI output necessarily strengthens the human project it is intended to support.

A plausible implication is that co-authorship is not reducible to output assembly. In identity-relevant, behavior-dependent tasks, authorship may be part of the mechanism by which a work acquires commitment, accountability, or value.

4. Authorship attribution as an inference problem

Authorship attribution research treats “authored” as an inferable signal in linguistic style, code structure, visual texture, or latent model representations. The tasks differ, but the underlying pattern is consistent: authorship is framed as a classification or evidence-accumulation problem over features that remain stable enough to survive topical variation.

Task	Method	Reported result
50-way text authorship identification	Lexical, syntactic, and 23 stylometric features with LibLINEAR SVM	91.30% 10-fold CV; 81.6% holdout
Single vs. multi-authored documents	Merit-based late fusion with Powell optimization	$x$1
Human-authored vs. GPT-4-generated Python code	XGBoost over 140 stylometric features	$x$2, AUC-ROC $x$3
Human-robot collaborative paintings	Patch-based VGG-style CNN plus majority vote	88.8% patch accuracy; 86.7% painting-level accuracy
Dynamic black-box LLM provenance	READER with proxy activations and Bayesian evidence accumulation	31.0%–42.4% top-1 at $x$4; 70.0%–84.0% at $x$5

These results are reported in (Iyer et al., 2019, Zamir et al., 2023, Idialu et al., 2024, Chen et al., 19 Feb 2026, Liu et al., 9 Jun 2026).

The methodological diversity is substantial. In classic text attribution, performance improves when unigram features are enriched with word bigrams, POS bigrams, word/POS pairs, and stylometric meta-features such as punctuation usage, lexical diversity, and sentence-length statistics (Iyer et al., 2019). In multi-authored document analysis, late fusion of BERT-base, DistilBERT, ALBERT, RoBERTa-base, and XLM-RoBERTa improves performance across three tasks: single vs. multi-authored classification, single author-switch detection, and multiple author-switch detection. On the PAN-21 dataset, the best raw-data fusion scores are $x$6 for Task 1 and $x$7 for Task 2, while clean-data fusion reaches $x$8 for full author attribution in Task 3 (Zamir et al., 2024).

The code domain exhibits the same structure under different features. One line of work distinguishes GPT-4-generated from human-authored CodeChef solutions with a grouped 10-fold split on problem ID, using XGBoost over layout, lexical density, AST, and complexity features; the non-gameable variant, excluding emptyLinesDensity and whiteSpaceRatio, still achieves $x$9 and AUC-ROC $50$0 (Idialu et al., 2024). Another line of work separates ChatGPT-generated from human-authored Java and C++ code using lexical, structural-layout, and semantic features, reporting for the best J48 model an accuracy, precision, recall, and $50$1 of $50$2 in Java and $50$3 in C++ when all features are used (Ke et al., 2023).

Visual attribution introduces explicitly local uncertainty measures. In human-robot painting, patches from manually annotated hybrid regions show a mean conditional entropy of $50$4 bits versus $50$5 bits in pure works, a $50$6 relative increase with $50$7, supporting the interpretation that elevated uncertainty reflects mixed authorship rather than mere failure (Chen et al., 19 Feb 2026). In black-box LLM provenance, READER maps outputs into the hidden state space of a frozen proxy LLM, averages sampled token states to form per-response representations, and accumulates log-posterior evidence across prompts: $50$8 This replaces fragile mean-pooling over prompt-dependent embeddings with calibrated multi-query attribution (Liu et al., 9 Jun 2026).

A recurrent empirical finding is that signals often survive in materials standard NLP pipelines discard. Retaining punctuation, contractions, stop-words, and very short words improves authorship-sensitive performance on single-vs-multi-authored classification and style-change detection, indicating that authorial trace is frequently encoded in low-level form rather than overt topical content (Zamir et al., 2023, Zamir et al., 2024).

5. Process evidence and the construction of human-authored corpora

Not all authorship systems infer from the finished artifact. The “Writer’s Integrity” framework argues that verification should target the writing process rather than the final text. Its architecture comprises a front-end text editor, real-time logger, change detection module, log cleaning and condensation, analysis engine, and certificate generator/verifier. The system captures keystroke, delete/backspace, and paste events, computes typing speed, edit frequency, paste ratio, and average changes per word, and issues a GUID-based certificate linked to the stored metrics and cleaned log (Aburass et al., 2024). Pilot results report 100% successful certificate generation for all $50$9 human writing sessions and log cleaning reductions of $$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$0–$$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$1 per session. The framework’s central premise is that genuine human authorship leaves an iterative, non-linear behavioral footprint that cannot be reconstructed by post hoc paraphrase.

Human-authored datasets can also be engineered at corpus scale. PARHAF is a large open-source corpus of French clinical documents describing entirely fictitious patient cases. It was created by $$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$2 medical residents across $$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$3 specialties, covers $$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$4 patient cases and $$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$5 reports, and combines clinician-authored content with epidemiological guidance from the French National Health Data System. The protocol defines a clinical case tuple, builds a sampling database of $$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$6 distinct cases covering $$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$7 ICD-10 codes, and uses square-root-plus-cap sampling

$$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$8

Each draft is peer-reviewed by a second resident, $$orig(W)= \begin{cases} 1, & \text{if } H(W)\ge \tau\ 0, & \text{otherwise} \end{cases}$$9 of submissions were rejected and rewritten, and generative AI was discouraged to preserve authentic clinical style (Tannier et al., 20 Mar 2026).

These two strands treat authorship as an infrastructural property. In one case, the goal is certification of a live writing trace; in the other, it is the systematic production of a shareable corpus whose provenance, realism, and human origin are specified at design time. This suggests that authorship can be embedded upstream, in the protocol that governs creation, rather than inferred solely downstream from a completed artifact.

6. Controversies, misconceptions, and regulatory directions

A persistent misconception is that prompting is analogous to merely instructing a commissioned artist. The U.S. Copyright Office adopts this analogy in denying users “ultimate creative control,” whereas Mei argues that it overlooks iterative prompt-refinement, curation, and post-processing by which users actively shape outputs (Mei, 2024). A second misconception is that final-output inspection alone is an adequate basis for authorship determination. Process-based logging, transcript publication, and multi-query provenance systems all challenge that assumption by relocating evidence to the interaction history or to aggregated latent traces (Aburass et al., 2024, Zhou, 5 Apr 2026, Liu et al., 9 Jun 2026).

Watermarking introduces a different controversy. Private watermark schemes give model providers sole possession of the key that both embeds and verifies a hidden signal. On the account developed in the political and ethical analysis of LLM watermarks, this grants private firms sweeping practical power to determine authorship, including the power to generate false positives or reserve unwatermarked modes for favored clients (Räz, 2024). The same paper argues that watermark detectability depends on entropy, creating the possibility of group-dependent rates of machine-text detection when topic, register, language size, or translation constraints induce systematically lower-entropy outputs.

Empirical work on spear phishing shows how difficult authorship judgments already are at the perceptual level. In a TRAPD-based experiment on personalized SMS phishing, participants’ overall accuracy in labeling messages as Human vs AI was $X_h$0, barely above random, and logistic regression using Emoji, LinkMod, and CharacterCount found no reliable cues (Francia et al., 2024). The study therefore complicates the expectation that end users can reliably infer provenance from surface style.

Emerging proposals reflect these tensions. Mei advocates a flexible, principles-based regulatory framework with transparency about AI use but no blanket exclusion, plus periodic review as capabilities evolve (Mei, 2024). Pereira proposes institutionally operationalized threshold assessment keyed to explainability, intellectual control, and the nature of AI output (Pereira, 6 Apr 2026). Zhou argues that full, unedited AI interaction transcripts should be published as standard supplementary material and treated as raw experimental data (Zhou, 5 Apr 2026). The common thread is not a single settled doctrine of authorship, but a movement toward evidentiary regimes that can register hybrid creation without collapsing human responsibility into either total ownership or total displacement.