Article illustration 1

If you've ever winced at an AI-generated joke or yawned through an LLM-written story, you've witnessed a core limitation of modern artificial intelligence: its inability to master the delicate balance of surprise and inevitability. This concept, explored in depth by Dan Fabulich in his article Surprising, but Inevitable: Stories, Jokes, Puzzles, and Proofs, reveals why human-crafted narratives resonate while AI output often feels like 'slop.' For developers and tech leaders, understanding this gap isn't just academic—it's key to designing systems that can truly innovate.

The Essence of Surprise and Inevitability

At the heart of compelling human communication lies a paradox: the best jokes, stories, puzzles, and even mathematical proofs must be both unexpected and, upon reflection, utterly logical. As Fabulich argues:

"A good story has to be surprising. If you can predict what will happen next, then the story is boring. But the story events also have to follow from one another—they have to be inevitable in hindsight, or the story won't make sense."

This principle applies universally:
- Jokes: A punchline must catch us off guard, yet make us think, "I should have seen that coming." If it's too predictable, it's dull; if it's nonsensical even after explanation, it's a failure.
- Puzzles and Riddles: Solutions should be elusive but fair—requiring a leap of insight that feels earned, not arbitrary. Unfair puzzles frustrate users by demanding mind-reading.
- Journalism: Beyond reporting shocks (like 'man bites dog'), great stories contextualize events to show why they were inevitable, transforming news into meaning.
- Mathematical Proofs: Elegant proofs often arrive at surprising conclusions through steps that seem obvious in retrospect, challenging prior assumptions.

This duality engages our brains by rewarding pattern recognition while subverting expectations—a cognitive sweet spot that fosters trust, delight, and insight.

Why LLMs Inherently Struggle

Large language models, trained on next-word prediction, are optimized to minimize surprise. Their objective is statistical likelihood, not creative revelation. When generating a joke or story, LLMs default to the most probable output based on training data, producing content that's generic and predictable. As Fabulich notes:

"LLMs are trained to predict what the 'next word' would be in a sentence. Their objective requires the LLM to keep surprise to an absolute minimum... This is why we call LLM-generated content 'AI slop.' Slop is just more and more of the same thing we’ve already seen."

This flaw manifests in tangible ways for developers:
- Boring Output: LLM-generated narratives lack tension or originality because they avoid risks that could lead to novelty.
- Contextual Blind Spots: Models can't tailor 'surprise' to individual users' knowledge. What's obvious to a seasoned puzzle-solver might baffle a novice, making outputs feel unfair or irrelevant.
- Creative Stagnation: In mathematical proofs, LLMs reproduce established methods but rarely achieve breakthroughs, as true innovation requires leaps beyond existing data patterns.

Implications for AI Development

The ramifications extend across tech domains. In cybersecurity, AI that can't anticipate novel attack vectors (surprise) or logically trace their origins (inevitability) is less effective. For generative AI tools, users crave unpredictability in storytelling apps or game design—but only if it feels coherent. Fabulich suggests puzzles can be made more accessible through gradual hints, hinting at a solution: incorporating user feedback loops to calibrate surprise. Developers might explore:
- Hybrid Models: Combining LLMs with symbolic AI to enforce logical consistency while allowing controlled randomness.
- Reinforcement Learning: Training systems on engagement metrics that reward 'aha' moments, not just accuracy.
- Personalization Engines: Adapting outputs based on a user's expertise to make surprises feel earned, not jarring.

Ultimately, bridging this gap isn't just about better entertainment—it's about building AI that humans can trust and collaborate with. As we push toward artificial general intelligence, systems that master the art of the inevitable surprise could transform how we solve problems, tell stories, and understand our world. Until then, the most compelling narratives will remain distinctly human.