The concept of 'fair play' is an abstract yet widely accepted principle, both in sports and in crime literature. In the sporting context, it refers to the idea that competitors must ensure fair conditions for both to win, even if it means giving up an immediate advantage, such as when an opponent passes the ball so an injured teammate can receive medical attention.
In crime fiction, 'fair play' implies that the reader actively participates in the narrative, attempting to solve the mystery and identify the culprit before the conclusion. The author must provide all necessary elements for the reader to solve the enigma independently, requiring only sharp reasoning to connect the dots, without revealing the solution obviously.
The opposite of this principle is 'deus ex machina,' a situation where the mystery is resolved by elements introduced suddenly, without having been mentioned or suggested earlier in the plot. An example would be a detective announcing the discovery of a hidden camera recording in the final chapter, when such a camera was never mentioned in the book.
A 2026 study conducted by the Hebrew University of Jerusalem sought to quantify this literary concept, investigating whether stories generated by Artificial Intelligence (AI) can achieve a satisfactory level of 'fair play.' The researchers argue that a good narrative must offer a solution accessible to two types of readers: the 'naive' one, who interprets clues superficially, and the 'brilliant detective,' who makes the most accurate inferences.
To conduct the measurement, AI developers were asked to create 54 mystery stories. Eleven human-written works were added to this group, including six stories by the detective Sherlock Holmes, of Arthur Conan Doyle, and five stories by the detective Hercule Poirot, of Agatha Christie.
The narratives were evaluated in two distinct ways. Firstly, the language models themselves simulated the two types of readers. Secondly, twenty real people read the stories section by section, trying to determine the culprit among four suspects. At the end, these participants rated the criteria of surprise, coherence, 'fair play,' and reading pleasure on a scale of 1 to 5.
In the tests conducted by humans, the only criterion where AI-created stories matched the level of original works was surprise. In the other three aspects—coherence, 'fair play,' and reading pleasure—the evaluators pointed to the superiority of stories written by Agatha Christie and Conan Doyle.
This result indicates that AIs are capable of generating unpredictable endings, meaning twists, but generally fail to establish the 'fair play' necessary to sustain that twist. The study's conclusion points out that the models can produce surprise or coherence in isolation, but they struggle to combine them, which is essential for building 'fair play.'
In summary, AIs can create narratives with shocking conclusions, such as when the culprit is the least suspected person. However, they often fail to perform the reverse process: ensuring that all clues presented throughout the story make sense and allow an attentive reader to reach the correct conclusion without cheating.