💬 NLP

LLMs Exhibit Significantly Lower Uncertainty in Creative Writing Than Professional Writers

This paper argues that large language models exhibit significantly lower uncertainty than human writers in creative tasks due to alignment strategies prioritizing factuality, a constraint that suppresses literary richness and necessitates new uncertainty-aware paradigms to achieve human-level creativity.

Original authors: Peiqi Sui

Published 2026-02-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Peiqi Sui

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are asking two different chefs to cook a meal based on a single ingredient: a potato.

Chef A is a professional human writer. Chef B is a Large Language Model (LLM), the AI behind tools like the one you are using right now.

The paper you shared argues that while Chef B is incredibly fast and can make a perfectly edible potato dish, it lacks a specific secret ingredient that makes Chef A's food truly artistic: Uncertainty.

Here is the breakdown of the paper's findings using simple analogies.

1. The Core Problem: The "Safe" Chef vs. The "Risky" Artist

The paper starts with a quote from the poet John Keats, who said that great art requires the ability to sit with "uncertainties, mysteries, and doubts" without rushing to find a quick, factual answer.

The AI Chef (The "Safe" Chef): AI models are trained to be helpful, factual, and safe. If you ask them a question, they want to give you the most likely, correct answer. They are like a GPS that never takes a detour because it wants to get you to your destination in the shortest time. In creative writing, this means the AI always picks the most predictable, cliché path. It avoids "mistakes" or weird twists because its training tells it that "uncertainty" usually equals "hallucination" (lying).
The Human Chef (The "Risky" Artist): Human writers, especially in fiction, thrive on the unknown. They introduce ambiguity, weird metaphors, and plot twists that don't immediately make sense. They are willing to take a detour down a dark, foggy alley because that's where the interesting story happens.

The Finding: The paper measured this using math (information theory) and found that human writing is 2 to 4 times more "surprising" (uncertain) than AI writing. The AI is too comfortable; the human is delightfully uncomfortable.

2. The "Uncertainty Gap"

The researchers created a "Uncertainty Gap" meter. They fed the same story prompt to both humans and AI and asked: "How surprised are you by what comes next?"

The Result: The AI was rarely surprised. It knew exactly what word would come next because it was playing it safe.
The Human: The human writer often chose words that the AI didn't expect. This "surprise" is what makes a story feel alive.

The Twist: The paper found that the more you "train" the AI to be helpful (using something called "alignment" or "instruction tuning"), the worse it gets at being creative. It's like training a jazz musician to only play sheet music perfectly. The more you force them to follow the rules, the less they can improvise. The "Thinking" models (AI that reasons step-by-step) were actually more rigid and less creative than the basic models.

3. Why "Mistakes" Are Actually Good

In the world of AI safety, "hallucinations" (making things up) are bad. But in literature, a little bit of "making things up" is essential.

The Analogy: Imagine a painting. If the AI paints a perfect, photorealistic apple, it's technically correct. But if a human painter paints an apple that is slightly blue, or floating, or melting, it creates a feeling. It makes the viewer stop and think, "What does this mean?"
The Paper's Point: The AI is trained to scrub away these "errors" to make the text smooth and clear. But by smoothing out the rough edges, the AI removes the mystery. The paper argues that literary richness comes from "gaps" in the story—places where the meaning isn't 100% clear, forcing the reader to use their imagination to fill them in. The AI fills in all the gaps, leaving the reader with nothing to do but consume.

4. The "Sweet Spot" of Creativity

The researchers looked at whether being "uncertain" actually makes writing better. They found a Goldilocks Zone (an inverted U-curve):

Too Low Uncertainty: The story is boring, repetitive, and predictable (like a robot reading a manual).
Too High Uncertainty: The story is nonsense, incoherent, and confusing.
Just Right (The Sweet Spot): The story is surprising but still makes sense. This is where high-quality human writing lives.

The Problem: The AI is stuck on the far left side of the curve. It is so afraid of being wrong that it never reaches the "sweet spot" of creative surprise.

5. The Big Picture: Why This Matters

The paper concludes that we are heading toward a "Textpocalypse"—a world flooded with smooth, frictionless, but soulless content.

The Risk: If we rely on AI to write our stories, news, and art, we might lose the ability to handle ambiguity. We might get used to answers that are always clear and certain, losing the human capacity to sit with a mystery.
The Solution: We need to teach AI to be comfortable with "not knowing." We need new ways to train these models that distinguish between a dangerous lie (a medical diagnosis that is wrong) and a creative ambiguity (a metaphor that is open to interpretation).

Summary in One Sentence

Great stories need a little bit of chaos and mystery to be interesting, but our current AI models are trained to be so safe and certain that they accidentally kill the magic of storytelling.

1. Problem Statement

The paper addresses a critical limitation in Large Language Models (LLMs): their inability to produce high-quality creative writing, often resulting in outputs that are trite, repetitive, cliché-ridden, and lacking in dramatic tension.

The Core Conflict: Current LLM alignment strategies (RLHF, SFT) are optimized to reduce uncertainty, hallucinations, and ambiguity to ensure factuality and instruction-following in functional domains (e.g., QA, coding). However, literary theory posits that uncertainty (ambiguity, indeterminacy, and "negative capability") is a fundamental requirement for literary richness and creative expression.
The Hypothesis: The authors argue that the systematic aversion to uncertainty in aligned LLMs creates a "human-model uncertainty gap," causing models to collapse into safe, homogenized attractor states that lack the constructive ambiguity necessary for compelling storytelling.

2. Methodology

The study employs a controlled, information-theoretic framework to quantify the distributional divergence between human-authored fiction and model-generated continuations.

A. Experimental Setup

Task: Story completion. Given a context prefix ( $T_1$ ), the study compares the uncertainty of the human ground-truth continuation ( $T_2$ ) against a model-generated continuation ( $\hat{T}_2$ ).
Models: A diverse suite of 28 open-weights models across families (OLMo, LLaMA, Qwen, Mistral, Gemma, Phi), including base, instruction-tuned, and "reasoning" (thinking) variants.
Datasets:
- The New Yorker (2010–2019): High-quality, professionally edited literary fiction.
- Tell-Me-A-Story: Professional workshop-generated stories.
- Ghostbuster: Used for domain comparison (Creative Writing vs. News vs. Essays).
Evaluation Mechanism: The same LLM ( $P_\theta$ ) acts as both the generator and the evaluator. It calculates the log-probability of human tokens given the context, ensuring the uncertainty metric reflects the model's perspective on human creativity versus its own generation.

B. Uncertainty Metrics

The authors utilize several token- and sequence-level metrics derived from log-probabilities:

Mean Token Entropy (Surprisal): Length-normalized Negative Log-Likelihood (NLL). Measures the average "unexpectedness" of tokens.
Perplexity (PPL): The exponentiated mean token surprisal.
Pointwise Mutual Information (PMI): Measures how much the context ( $T_1$ ) predicts the continuation ( $T_2$ ) versus the unconditional probability. High PMI indicates over-reliance on context; low PMI indicates divergence/creativity.
Conditional PMI (CPMI): A thresholded variant focusing on high-entropy decision points to isolate narrative ambiguity.

C. Quality Correlation

To determine if uncertainty correlates with quality, the authors used a Writing Quality Reward Model (fine-tuned on expert preferences) to score continuations. They performed linear and quadratic (inverted-U) regression analyses to test for a "sweet spot" of uncertainty (Wundt curve).

3. Key Results

A. The Human-Model Uncertainty Gap

Consistent Divergence: Across all models, datasets, and metrics, human writing exhibits significantly higher uncertainty than model outputs.
- NLL Ratio: Human stories are 2.03x to 3.9x more "surprising" to the model than the model's own generations.
- Perplexity Ratio: The gap widens to 2.76x to 8.82x.
- PMI: Models show significantly higher context-dependence (more negative $\Delta$ PMI), indicating they over-rely on predictable narrative paths, whereas humans deviate from expected trajectories.

B. Impact of Alignment and Model Type

Alignment Tax: Instruction-tuned and "reasoning" (thinking) models exhibit a larger uncertainty gap compared to their base counterparts.
- Example: Gemma-3-4B's NLL ratio doubles from base to instruct.
- Reasoning Models: Chain-of-thought reasoning does not recover creative uncertainty; in some cases, it exacerbates the gap.
Model Families: While base models generally preserve more diversity, specific families (like Gemma) show particularly rigid generation distributions, likely due to knowledge distillation techniques.

C. Domain Specificity

Amplification in Creative Writing: The uncertainty gap is present across domains (News, Essays) but is significantly amplified in creative writing.
- The $\Delta$ PMI gap for creative writing is 25–30% wider than for essays or news.
- This suggests models struggle specifically with the narrative-level tension and structural ambiguity required for fiction, even if they can handle local lexical uncertainty similarly across domains.

D. Correlation with Writing Quality

Positive Correlation: Higher uncertainty (NLL/PPL) correlates positively with writing quality scores for both human and model text.
The "Sweet Spot": Quadratic regression reveals an inverted-U relationship (Wundt curve).
- Optimal quality occurs at high entropy levels (approx. 1.48–1.98 standard deviations above the mean).
- Crucial Finding: Aligned LLMs operate to the left of this peak (low entropy), trapped in a safe, low-uncertainty regime that precludes high-quality narrative. Human writers naturally operate closer to the peak.

4. Key Contributions

Formalization of the Uncertainty Gap: The paper provides the first systematic, information-theoretic quantification of the gap between human and model uncertainty in creative writing, moving beyond qualitative assessments.
Theoretical Synthesis: It bridges Literary Theory (concepts of indeterminacy, ambiguity, and co-creativity) with NLP, arguing that "hallucination" in a creative context is structurally analogous to the "infinite play" of meaning required for literature.
Empirical Evidence on Alignment: It demonstrates that current post-training alignment strategies (RLHF, SFT) actively degrade creative potential by suppressing the very uncertainty mechanisms that drive literary richness.
Quality-Uncertainty Link: It establishes a robust, non-linear correlation between uncertainty and writing quality, identifying a "sweet spot" that current models fail to reach.

5. Significance and Implications

Redefining "Hallucination": The paper challenges the binary view of hallucination. In creative writing, semantic drift and ambiguity are not errors but constructive features necessary for engagement. Current alignment penalizes these features, leading to "stuplimity" (a mix of shock and boredom).
Future Alignment Paradigms: The authors argue for uncertainty-aware alignment. Future models must be trained to distinguish between destructive hallucinations (factual errors in medical/legal contexts) and constructive ambiguity (metaphor, plot twists, open-endedness).
Systemic Risk: The homogenization of creative output poses a risk to cultural diversity. If LLMs dominate creative generation, they may flatten the "long tail" of human expression into a narrow set of safe, predictable patterns.
Interdisciplinary Approach: The paper advocates for "literary-theory-informed" machine learning, suggesting that solving creative AI problems requires integrating domain-specific humanities frameworks rather than relying solely on computer science metrics.

In conclusion, the paper posits that achieving human-level creativity in AI requires not just better data or larger models, but a fundamental shift in how models are aligned to tolerate and generate pragmatic uncertainty.