LLMs Exhibit Significantly Lower Uncertainty in Creative Writing Than Professional Writers

LLMs consistently generate story continuations with 2–4× lower entropy and substantially higher context-dependence than human-authored ground truth—a gap that widens under post-training alignment and persists across model families and scales.

The Gist

LLMs consistently generate story continuations with 2–4× lower entropy and substantially higher context-dependence than human-authored ground truth—a gap that widens under post-training alignment and persists across model families and scales.

Theoretical Framework of Uncertainty in Creativity

Literary theory posits that uncertainty is a necessary condition for creative expression, arising from concepts like indeterminacy, ambiguity, and the openness of a work to multiple interpretations. This suggests that literary richness stems from the simultaneous presence of conflicting meanings and gaps in meaning-making. The paper formalizes this by proposing an information-theoretic framework to quantify the human–model uncertainty gap between human-authored fiction and model-generated continuations. The core tension identified is that current alignment strategies, aiming to reduce ambiguity for instruction-following, steer models away from uncertain outputs to ensure factuality and reduce hallucination. This systematic aversion to uncertainty is described as a mechanism that effectively forecloses this co-creative process, optimizing away a key precondition of the literary experience.

Experimental Methodology

The research employs a controlled story completion experiment comparing the uncertainty profiles of human-written versus model-generated sequences under identical settings. The comparison is structured by treating context as a cumulative prefix, constructing context–continuation pairs where the model's log-probability is evaluated for each token position in both the human continuation and the model continuation. To ensure a fair comparison, models are used as both generators and evaluators of human text uncertainty, with generation lengths constrained to match the ground truth length. Uncertainty is quantified using several metrics derived from log-probabilities:

1. Mean Token Entropy (Surprisal): Approximated using length-normalized Negative Log-Likelihood (NLL).

2. Perplexity (PPL): The exponentiated mean token surprisal, serving as a normalized measure of predictive uncertainty.

3. Pointwise Mutual Information (PMI): To disentangle context-dependent uncertainty from intrinsic token frequency, PMI normalizes conditional likelihood by an unconditional baseline.

4. Conditional PMI (CPMI): A thresholded variant that upweights the contribution of unconditional probabilities for tokens that are already uncertain given context.

Key Findings on Uncertainty Gaps

The analysis reveals a pervasive human–model uncertainty gap: when given identical contexts, LLMs consistently generate story continuations Tˆ2 with substantially lower intrinsic uncertainty than the human-authored ground truth T2. This gap is quantified by NLL ratios ranging from 2.03 to 3.9 and PPL ratios amplifying this disparity to 2.76–8.82×, indicating that human-written stories are on average 2–4× more 'surprising' to the model than its own generations. Furthermore, the gap is amplified in creative writing compared to other domains; for instance, the ∆PMI gap is 25–30% wider for creative writing (−2.15) compared to essays (−1.71) and news (−1.73). The paper also notes that instruction-tuned and reasoning models exacerbate this trend compared to their base counterparts.

Correlation with Writing Quality

The study establishes a robust correlation between higher uncertainty metrics and automated quality scores, finding evidence of a possible “sweet spots” (inverted U-curve) where optimal writing quality exists at high entropy levels. For human ground truth, the mean Spearman correlation between token-level uncertainty (NLL, PPL) and quality scores is positive: for human-authored continuations, the mean Spearman correlation is ρ¯ = 0.072. Crucially, PMI shows a robust negative correlation with quality for both human and model text (PMI shows a robust negative correlation with quality for both human (ρ¯ = −0.106, 75% significantly negative) and model text (ρ¯ = −0.076, 69% significantly negative)). This suggests that divergence is a linear driver of quality: the more a continuation creates its own information distinct from the strong priors of the prompt, the higher it is rated. The quadratic regression analysis confirms this non-monotonic relationship for human text, with optimal uncertainty occurring at z¯∗ = 1.98 standard deviations above the mean.

Domain Generalization and Model Effects

The uncertainty gap is not unique to creative writing but is amplified in creative writing, especially on context-dependent uncertainty metrics. The analysis across three domains (Creative Writing, News, Essays) confirms that the ∆PMI gap is consistently larger for creative writing.

Here are specific improvements to AI systems based on the findings of this research:

1. Enhance creative generation capabilities by implementing a Uncertainty-Aware Alignment Module. This module should explicitly incorporate information-theoretic metrics (Token Entropy, Perplexity, PMI, and CPMI) during the reinforcement learning from human feedback (RLHF) or alignment phase.

2. Implement a dynamic uncertainty thresholding mechanism in the decoding process for generative models. Instead of forcing outputs into a low-entropy safe regime immediately, the system should be capable of sampling based on a calculated uncertainty metric that allows for controlled bursts of high-entropy, creative exploration when the context demands ambiguity (e.g., during narrative tension building).

3. Develop a quality assessment feedback loop that specifically weights outputs based on their creative uncertainty profile. The reward model should be fine-tuned to positively correlate with the observed sweet spot in the uncertainty-quality relationship (the inverted-U curve), thereby rewarding outputs that exhibit optimal, controlled ambiguity rather than those that are perfectly predictable or maximally random.

4. For LLMs used in narrative co-creation, develop a mechanism to distinguish between destructive hallucinations and constructive ambiguity. This involves training the model not just on factual correctness but also on literary affordances—learning when and how to introduce indeterminacy (unbestimmtheitsstellen) or ambiguity (Empson's gridiron) as a necessary aesthetic feature of the text.

5. Integrate domain-specific uncertainty calibration into alignment strategies. Since the paper shows that the human–model uncertainty gap is amplified in creative writing compared to functional domains (News/Essays), future systems should utilize domain classifiers to apply more aggressive, uncertainty-aware constraints specifically for open-ended, narrative tasks while relaxing them for structured reasoning or factual retrieval tasks.

6. Improve model interpretability by providing Uncertainty Attribution Maps. When a model generates a text segment, it should be able to highlight which tokens contributed most significantly to the local surprisal (Htoken) and how contextual relevance (PMI/CPMI) was calculated for that specific decision point, allowing researchers to diagnose why the model chose an ambiguous or predictable path.

These improvements will enable AI systems to move beyond merely generating safe or average text, allowing them to produce creative writing that exhibits the necessary features of literary richness: tension, ambiguity, and genuine surprise.