LLMs vs. Top Human Authors

• The paper demonstrates that in direct competitions, GPT-4 scores significantly lower in originality and authorial voice (e.g., 10% vs. 65%) compared to elite human writers.
• LLMs are systematically evaluated via blinded contests using metrics such as attractiveness, creativity, and factual consistency to benchmark performance against expert human outputs.
• Fine-tuning and prompt engineering markedly enhance LLM outputs, suggesting potential market disruptions and shifts in traditional creative authorship.

LLMs are now routinely benchmarked against skilled and even elite human authors across a spectrum of tasks ranging from creative fiction and literary emulation to technical writing and legal or scientific composition. The competitiveness of LLMs with top human authors—defined as world-class literary figures, expert creative writers, or domain-specific virtuosi—remains an area of active empirical study. Evaluations have rapidly evolved from demonstrating LLMs surpassing average humans on mechanical language tasks to investigating the nuanced, high-bar standards of originality, creativity, and voice that characterize canonical human output.

1. Experimental Paradigms for Human–LLM Competitions

Recent work has operationalized the notion of LLM competitiveness by direct, blinded duels between high-status human authors and state-of-the-art models. For instance, "Pron vs Prompt" (Marco et al., 2024) staged a contest between Patricio Pron—an internationally acclaimed novelist—and GPT-4 Turbo, with symmetric tasks: both generated 30 movie titles and wrote ~600-word synopses for each title (both their own and their opponent's). Evaluation deployed a rubric rooted in Boden’s creativity framework, spanning Attractiveness, Originality, Holistic Creativity, Anthology Suitability, and Authorial Voice, with 5,400 expert annotations from literary critics and scholars.

Parallel experimental structures have been used to benchmark LLM emulations against MFA-level writers across 50 award-winning author styles (Chakrabarty et al., 15 Oct 2025, Chakrabarty et al., 26 Jan 2026). Here, LLMs are either prompted in-context (receiving exemplars and instructions) or fine-tuned on the full corpus of target authors, with output compared to human-authored imitative texts in blinded, forced-choice studies scored by both literary experts and lay readers.

In technical and scientific writing, LLMs have been assessed on the reliability and quality of academic literature reviews via large, multi-task datasets with alignment to real-world publication and metadata standards (Tang et al., 2024). Evaluation metrics include factual consistency, reference accuracy, semantic coverage, and overlap with human-authored content.

2. Quantitative Performance: Head-to-Head Outcomes

Performance metrics draw on rating scales, preference votes, and logistic modeling. In direct competition with Patricio Pron, GPT-4 was rated as “appealing” in only 5–17% of judgments on style and theme (vs. Pron's ~70%), “original” in 10% (vs. 65%), “overall creative” in 24% (vs. 88%), with similar gaps on “Anthology Suitability” and “Own Voice” (GPT-4 ~20%, Pron ~80%+) (Marco et al., 2024). Statistical tests—e.g., Mann-Whitney U, Wilcoxon signed-rank—confirmed these margins were large and significant. LLM outputs, even at maximum temperature and with human-engineered prompts, remained markedly less original, surprising, and distinctive than those from top-tier human authors.

On a separate English-language benchmark involving creative-writing students versus commercial LLMs (GPT-4, Claude, Bing), LLMs matched or exceeded overall human scores in categories such as fluency and structure (e.g., GPT-4 mean overall 7.45 ± 0.60 vs. human 7.10 ± 0.80, $p=0.005$). However, humans retained a statistically significant advantage in creativity/originality and humor, even as top LLMs performed on par in other stylistic and technical dimensions (Gómez-Rodríguez et al., 2024, Gómez-Rodríguez et al., 2023).

Table 1: Exemplary Head-to-Head Scores (Spanish Contest, (Marco et al., 2024))

Dimension	Pron (% "positive" 2–3)	GPT-4 (% "positive" 2–3)
Attractiveness	~70	5–17
Originality	~65	10
Creativity	~88	24
Anthology	~80	20
Own Voice	~80	20

In emulation tasks employing fine-tuned LLMs, a pivotal shift emerges: while in-context LLM output is strongly disfavored by expert readers (style/fidelity odds ratio OR=0.16, writing quality OR=0.13 vs. expert humans; both $p<10^{-7}$), fine-tuned LLMs reverse these findings—experts prefer fine-tuned AI on style (OR=8.2) and writing quality (OR=1.87), both $p\ll0.05$ (Chakrabarty et al., 15 Oct 2025, Chakrabarty et al., 26 Jan 2026). Lay readers consistently display an even greater preference for AI-generated text after fine-tuning (e.g., quality OR=2.42).

3. Critical Analysis: Sources of Human Superiority and LLM Shortcomings

Empirical analyses converge on several persistent deficits in LLM creative output, especially when compared to elite human authors under free-creative or open-imaginative tasks:

• Tropification and Predictability: LLMs exhibit a tendency towards familiar tropes and narrative arcs, with limited capacity for truly inventive or surprising connections—hallmarks of high-literary creativity. Weak emotional depth and shallow character-narrative integration are recurrent (Marco et al., 2024).
• Deficit in Authorial Voice: LLMs lack a distinctive, memorable voice; outputs are frequently described by evaluators as “generic” or “templated.” In contrast, top human authors (e.g., Pron) are recognized for a consistent voice and thematic resonance (Marco et al., 2024).
• Learnable Stylometric Signature: Human judges rapidly acquire the ability to recognize AI-generated prose, with detection accuracy for AI texts rising from 55% to 75% after limited exposure—indicating that LLM output retains significant statistical regularities absent from elite human writing (Marco et al., 2024).

At the technical-writing frontier, LLMs reach high factual and semantic similarity to human abstracts (cosine similarity ≈80%), but are prone to hallucination in reference generation, with even the best models (Claude-3.5) hallucinating ~45% of citations in zero-shot reference tasks (Tang et al., 2024).

4. Task and Workflow Effects: The Impact of Prompting and Fine-Tuning

The form of LLM engagement—prompted vs. fine-tuned—critically modulates competitiveness:

• Prompt and Context Effects: LLM outputs are substantially elevated by human-engineered prompts (e.g., evocative titles authored by a top-writer), leading to marked improvements in internal rubric dimensions (style originality +57%, anthology suitability +45%, own voice +30%) (Marco et al., 2024). This demonstrates prompt quality as a lightweight form of co-authoring.
• Fine-Tuning: Author-specific fine-tuning eliminates stylometric markers detectable by current AI detectors (flagged in only 3% of outputs, vs. 97% for in-context outputs), erases cliché density, and results in large, statistically significant reversals of reader preference in favor of LLM output across both expert and lay populations (Chakrabarty et al., 15 Oct 2025, Chakrabarty et al., 26 Jan 2026). Quantitatively, expert judges preferred fine-tuned AI writing over human MFA emulators in 62.2% of cases (binomial CI$_{95}$ ≈ [.517, .716]), a dramatic reversal compared to the 17.3% AI preference under in-context prompting (Chakrabarty et al., 26 Jan 2026).

5. Domain-Specific Competitiveness: Beyond Literary Prose

In competitive programming and legal reasoning, the frontier of LLMs’ competitiveness varies sharply by task:

• Algorithmic Code Generation: LLMs now rival strong Contest Masters on “template” or implementation-heavy problems but are hundreds of Elo points (e.g., best LLM Elo 2,116 vs. human grandmasters 2,700) short on the combinatorially hard, insight-driven tasks. They achieve 0% pass@1 on Hard problems and 53.5% on medium-difficulty tasks, reflecting persistent deficits in deep algorithmic reasoning (Zheng et al., 13 Jun 2025).
• Legal Drafting and Analysis: LLMs match or exceed human toppers on objective legal MCQs but do not yet rival human performance on open-ended Supreme Court drafting. Human expertise remains essential for procedural compliance, authority/citation discipline, and forum-appropriate rhetorical structuring—dimensions where models incur regular mark deductions (Juvekar et al., 19 Oct 2025).
• Scientific Literature Reviews: LLMs exhibit strong summary and coverage abilities but suffer from unreliability in reference generation and discipline-dependent factual consistency, limiting their competitiveness for automated review authorship (Tang et al., 2024).

6. Implications for Creative Labor, Authorship, and Attribution

Author-specific fine-tuning transforms the LLM landscape: at modest cost (~$81 median per author), LLMs can produce non-verbatim, style-faithful texts that not only pass expert and lay Turing tests, but are outright preferred for writing quality and stylistic fidelity over MFA-level human imitators (Chakrabarty et al., 15 Oct 2025, Chakrabarty et al., 26 Jan 2026). This constitutes strong empirical evidence for a substantial “market dilution” risk in the literary domain, directly affecting copyright’s fourth fair-use factor and raising urgent regulatory, ethical, and labor market concerns.

Qualitative debriefs reveal profound impacts on professional identity for human authors confronted with superior or indistinguishable AI output, including erosion of aesthetic confidence and a re-examination of what constitutes “good writing” in an era of expert-level generative models. Calls for mandatory AI disclosure, watermarking, and robust attribution regimes are a direct response to these findings (Chakrabarty et al., 26 Jan 2026).

7. Outlook and Structural Limits

Despite progress, certain creative and narrative dimensions remain notably harder for LLMs operating in autonomous, zero-shot mode—even as fine-tuning narrows or inverts the gap for excerpt-level emulation. Theoretical and empirical work suggests that simple scale or generic pretraining will not instantiate the depth of originality, conceptual grounding, or voice of world-class authors (Marco et al., 2024). Advances may require novel architectures, grounding with semantic and conceptual reasoning, or hybrid human–AI workflows that blend machine fluency with genuinely original human insight.

A plausible implication is that “superhuman storytelling” by LLMs across open-domain creative, legal, and technical tasks remains a moving target, defined not just by output metrics but by shifts in human evaluation criteria, market dynamics, and the evolving meaning of authorship itself.