This paper argues that large language model (LLM) hallucinations, or 'confabulations,' should be viewed as a potential resource rather than a negative flaw. The authors challenge the standard view that confabulations are inherently problematic and should be eliminated, presenting empirical evidence that measurable semantic characteristics of LLM confabulations mirror human tendencies to use increased narrativity for sense-making and communication. Specifically, they analyze popular hallucination benchmarks and find that hallucinated outputs display higher levels of narrativity and semantic coherence compared to veridical outputs. This suggests that the tendency for LLMs to confabulate may be associated with a positive capacity for coherent narrative-text generation. The paper also explores the cognitive and social benefits of confabulation, highlighting its potential utility in various domains such as domain-specific scenarios, text summarization, and adversarial examples. The authors define confabulation as a narrative impulse to generate more substantive and coherent outputs, mirroring human storytelling. They use narrative richness and coherence as key metrics to evaluate LLM confabulations, finding significant correlations between narrativity and coherence. The paper concludes by advocating for a more nuanced examination of confabulation, suggesting that it could enhance LLM capabilities and open new avenues for research and application.This paper argues that large language model (LLM) hallucinations, or 'confabulations,' should be viewed as a potential resource rather than a negative flaw. The authors challenge the standard view that confabulations are inherently problematic and should be eliminated, presenting empirical evidence that measurable semantic characteristics of LLM confabulations mirror human tendencies to use increased narrativity for sense-making and communication. Specifically, they analyze popular hallucination benchmarks and find that hallucinated outputs display higher levels of narrativity and semantic coherence compared to veridical outputs. This suggests that the tendency for LLMs to confabulate may be associated with a positive capacity for coherent narrative-text generation. The paper also explores the cognitive and social benefits of confabulation, highlighting its potential utility in various domains such as domain-specific scenarios, text summarization, and adversarial examples. The authors define confabulation as a narrative impulse to generate more substantive and coherent outputs, mirroring human storytelling. They use narrative richness and coherence as key metrics to evaluate LLM confabulations, finding significant correlations between narrativity and coherence. The paper concludes by advocating for a more nuanced examination of confabulation, suggesting that it could enhance LLM capabilities and open new avenues for research and application.