Human-Authorship Indeterminacy

• Human-authorship indeterminacy is a phenomenon where the origin of a text becomes indistinguishable between human creativity and AI generative processes due to overlapping styles and recursive collaboration.
• Empirical studies reveal a significant decline in attribution accuracy as advanced LLMs mimic human writing, challenging existing stylometric and detection methods.
• Legal, ethical, and methodological frameworks are evolving to address the blurred lines of attribution, emphasizing transparency, revised copyright policies, and process-based certification.

Human-authorship indeterminacy denotes the epistemic, methodological, and legal inability to confidently assign the origin of a text, or its creative components, to a human as opposed to an artificial agent, in contexts where the boundaries of authorship are blurred by automation, collaboration, and recursion. This concept arises both from the convergence of machine-generated and human-generated text distributions and from the recursive, stochastic, and often collaborative nature of modern writing workflows involving LLMs and agentic AI systems. Human-authorship indeterminacy is now central to debates in forensic linguistics, legal frameworks, attribution science, empirical aesthetics, and creative labor as AI systems attain parity or even dominance in domains previously reserved for human creativity.

1. Conceptual Foundations and Theoretical Frameworks

The notion of human-authorship indeterminacy emerges from a departure from classical, unitary authorship models. Historically, authorship presupposed an identifiable, intentional, and typically human agent, as codified in aesthetic philosophy and intellectual property law (e.g., the "romantic genius" paradigm and the legal requirement for “conception by a human mind”) (Sun et al., 25 Jan 2025, Mukherjee et al., 5 Apr 2025). Sun & Gualeni’s puppet–actor spectrum reframes AI authorship agency as occupying a continuum between externally-determined puppets and semi-autonomous actors: LLMs both follow human prompts and improvise, rendering definitive assignment of credit ambiguous when their generative contributions become co-constitutive (Sun et al., 25 Jan 2025).

Philosophically, authorship is increasingly seen not as a binary (human versus machine) but as a function of “degrees of distance” and context-dependent relevance: the entity (human or machine) responsible for the “relevant aspects” of a text (i.e., those aspects germane in a given context) is deemed the author (Räz, 2024). Authorship thus becomes distributed, recursive, and sometimes fundamentally unmappable, especially in agentic AI workflows where creative elements are irreversibly fused across human and machine feedback loops (Mukherjee et al., 5 Apr 2025, Zhou, 5 Apr 2026).

2. Empirical Studies of Indistinguishability and Attribution

Recent behavioral and computational studies show that the convergence between LLM outputs and human writing erodes the reliability of both expert and lay human attribution. Fine-tuned LLMs can so closely emulate authorial style that MFA-trained writers prefer AI-generated text to human originals, and their attribution accuracy drops below chance after exposure to such fine-tuned mimicry (Chakrabarty et al., 26 Jan 2026). In reception experiments outside English, such as Czech poetry with GPT-4.5, participants average 45.8% accuracy—below random guessing—and regard AI text as equally or more aesthetically valuable, except when they suspect AI authorship, which triggers devaluation regardless of actual origin (Marklová et al., 26 Nov 2025).

At the forensic level, stylometric classifiers and neural networks exploiting hundreds of surface or narrative features have achieved high accuracy for earlier LLMs and unconstrained genres, but their performance degrades as AI systems optimize for stylistic diversity and alignment with human text distributions. For example, even compact stylometric sets plus XGBoost models still achieve ≳89% accuracy on literary/political mimicry tasks, but their discriminative power depends on statistical regularities (e.g., perplexity, affective variance) not yet fully captured by current LLMs (Alsadhan, 24 Mar 2026). In long-form fiction, discourse-level narrative features (structure, agency, temporality) are fundamentally more resilient to surface editing and remain robust signals for attribution, yet cluster analyses indicate that LLM outputs continue to occupy a narrower region of narrative space than human productions (Russell et al., 3 Apr 2026).

3. Formal Impossibility Results and Limits of Detection

Information-theoretic arguments establish that, as LLMs converge to human writing distributions via scale and fine-tuning, the minimum error achievable by any detector approaches random guessing (Bayes-optimal error $E_{\min} \to 1/2$ as $\|P_H - P_{AI}\|_{TV} \to 0$) (Ganie, 15 Sep 2025). Attempts to enhance separability—by injecting cryptographic watermarks, altering token sampling, or exploiting stylometric artifacts—face the detection–authenticity trade-off: improving detection by boosting signal necessarily perturbs semantic/narrative fidelity ($\Delta S\cdot\Delta C \gtrsim \mathcal{K}$, with $\mathcal{K}>0$ set by inherent linguistic diversity). As a result, perfect detection is mathematically impossible in the limit of converged distributions.

Keystroke-timing and behavioral authentication schemes provide only "motor provenance"—they confirm that a human physically operated the instrument, but remain formally non-identifiable with regard to compositional authorship (Condrey, 24 Jan 2026). Copy-type and timing-forgery attacks, as well as scripting of editing logs, bypass these defenses; hence, process-based certification can only assert process integrity if tied directly to observable semantic content (Aburass et al., 2024).

4. Methodological Taxonomy and Practical Frameworks

A recent survey proposes a four-problem taxonomy for attribution under LLM co-authorship (Huang et al., 2024):

Problem Type	Description	Indeterminacy Source
Human Attribution	Assign text to one of $m$ human authors	Style/genre overlap, limited text length
LLM Detection	Binary label: Human (H) vs. Machine (M)	Convergence of LLM/human distributions
LLM Attribution	Multi-class: assign to human or LLM family	Fading between-model and human-model distinctions
Human-LLM Co-authored Attribution	Hybrid; possibly segment-wise attribution	Interleaving and iterative editing

Stylometric, feature-based, neural, transformer-based, and process-tracing approaches (including challenge-response and revision-history analysis) are surveyed, each with domain-specific strengths and weaknesses. Watermarking introduces both clarification (when public and verifiable) and muddying (when proprietary or entropy-dependent), potentially shifting authority and responsibility from creators to corporate keyholders (Räz, 2024). Process-level frameworks, as with the Writer’s Integrity system, attempt to bind behavioral metrics (typing speed, revision statistics, paste ratios) to certification––but are ultimately susceptible to simulation and motor presence–composition separation (Aburass et al., 2024, Condrey, 24 Jan 2026).

Narrative-feature pipelines such as StoryScope highlight that long-form, discourse-level structures (e.g., thematic explicitness, temporal complexity, protagonist agency) retain discriminatory value for attribution, and narrative "rarity" can act as a proxy for human originality (Russell et al., 3 Apr 2026).

5. Legal, Ethical, and Policy Consequences

Authorship indeterminacy produces significant challenges for copyright, patent, and liability regimes (Mukherjee et al., 5 Apr 2025, Sun et al., 25 Jan 2025). Classical IP law often presupposes a human principal as the originator of creative or inventive content. However, agentic AI with stochastic, dynamic, and fluid autonomy creates inseparable, recursively interleaved contributions. The practical impossibility of tracing distinct expressive elements to human minds forces a shift away from intent- or origin-based standards to outcome-based, "functional equivalence" regimes—vesting rights and responsibilities in those who initiate and curate projects, regardless of the provenance of particular components.

Ethically and epistemically, indeterminacy necessitates greater transparency. Radical transcript publication (e.g., full prompt–response archives as scientific "apparatus" logs) is advocated as the sole means for maintaining accountability in collaborative human–AI research (Zhou, 5 Apr 2026). Disclosure and explainable attribution become paramount as stylometric and annotation-based approaches lose efficacy. Policy recommendations include mandatory AI interaction supplementation, adapted role-taxonomies (CRediT extension), and layered attribution fields.

Watermarking centralizes evidentiary power and introduces new inequities (e.g., entropy-dependent error rates, ghostwriting privileges for privileged clients, and exposure to vendor lock-in) (Räz, 2024). The phenomenon of the “AI Ghostwriter Effect”—psychological and external attribution disjunction—exacerbates exploitation risks and complicates the adaptation of existing frameworks to collaborative and AI-mediated writing (Draxler et al., 2023).

6. Philosophical and Methodological Open Questions

The “Ship of Theseus” paradox applies directly: as LLMs paraphrase, co-write, or recursively edit texts, at what threshold does authorship shift from the human source to the machine, or become ambiguous enough to warrant shared or hybrid attribution? Quantitative experiments show that even a single paraphrase iteration can reduce classifier attribution accuracy from ≈77% to ≈33%, and that further edits quickly render origin inference uninformative (Tripto et al., 2023).

Proposed criteria for attribution after recursive co-editing or paraphrasing vary by context: content-centric use cases (e.g., scientific reporting) may retain human attribution if content remains unchanged, whereas style-centric or originality-focused domains (e.g., literature, patentability) may treat deeply-altered or LLM-recursive outputs as either collaborative or fundamentally ambiguous (Tripto et al., 2023, Sun et al., 25 Jan 2025, Chakrabarty et al., 26 Jan 2026). No uniform quantitative threshold has achieved consensus, though some frameworks recommend explicit transparency once more than 50% of content is AI-originated.

Future work is directed at formalizing “degrees of distance” metrics (Sun et al., 25 Jan 2025), developing robust, cross-domain and cross-linguistic datasets and benchmarks (Huang et al., 2024), and integrating psycholinguistic and actor–network insights with practical forensic, ethical, and legal guidance.

7. Implications for Scientific Publishing and Creative Labor

Human-authorship indeterminacy directly impacts academic integrity, scientific reproducibility, and the economics of creative labor. Empirical studies suggest that for scientific writing, hybrid human-AI composition with full transcript traceability is necessary to preserve the integrity of the record (Zhou, 5 Apr 2026). In fiction, narrative-space dispersion, rarity scores, and narrative-feature profiling may supplant speaker’s intent or stylometry as key markers for literary originality and copyright claims (Russell et al., 3 Apr 2026). For professional authors, indeterminacy in audience and peer attribution triggers identity re-evaluation and undermines traditional markers of expertise, uniqueness, and aesthetic authority (Chakrabarty et al., 26 Jan 2026).

In summary, human-authorship indeterminacy arises from the mathematical, methodological, recursive, and psycho-social convergence of human and AI creative processes. It erodes the epistemic ground for product-based attribution, rendering classical frameworks obsolete in favor of outcome-oriented, procedural transparency, and adaptive legal and scientific responses. The problem is not a contingent limit of technology, but a fundamental property of convergent autopoietic systems in language and creativity (Ganie, 15 Sep 2025, Mukherjee et al., 5 Apr 2025).