The Problem With Perfect Endings
Here's something that might make you uncomfortable if you write fiction for a living: your AI writing assistant is actively making your characters worse. Not subtly. Not in some vague, hard-to-pin-down way. It's systematically stripping the mystery out of your stories and replacing it with tidy, predictable archetypes that feel like they were assembled from a template.
A new study from UNC Chapel Hill has given us the tools to prove it. The researchers developed CASPER — an automated computational linguistics framework that measures character depth across eight distinct literary dimensions. And what they found wasn't just interesting. It was damning.
AI models don't struggle with mystery because they're dumb. They avoid it by design. The study shows that large language models possess an inherent mathematical bias toward wrapping up narratives cleanly, resolving internal conflicts aggressively, and ensuring every character fits neatly into their designated arc by the final page. Human writers? We leave questions unanswered. We let characters remain contradictory. We embrace ambiguity because that's what makes a story stick with a reader long after the last sentence fades.
This isn't a capability gap. It's a fundamental difference in how machines and humans understand storytelling.
What CASPER Actually Measures
Let's get specific about what the researchers were tracking, because this is where CASPER earns its keep as a benchmarking tool.
The framework analyzes thousands of stories — both human-authored and machine-generated — across eight core dimensions of literary theory. These include whether characters seem realistic or exaggerated, whether they genuinely evolve over time, and crucially, whether they remain mysterious or fully understood by the story's end. The mystery dimension is what makes this study so important: CASPER scans whether a character's internal motives are entirely explained by the narrator by the climax, or if their behaviors remain beautifully unmapped, contradictory, and open to multiple interpretations.
Think about your favorite literary characters. The ones that haunt you. Are they the ones whose motivations are perfectly laid out in chapter three? Or are they the ones you're still trying to figure out years later?
The UNC team turned abstract literary theory into measurable data. They quantified the exact invisible traits that separate flat caricatures from unforgettable icons. And when they applied this lens to AI-generated fiction, the results were consistent and unsettling.
The Safe Resolution Bias
Lead author Anneliese Brei put it plainly: AI systems possess an inherent mathematical bias to wrap up narratives cleanly. They aggressively resolve internal conflicts. They answer every mystery. They force characters into perfect arcs.
This is what the researchers call the "Safe Resolution Bias," and it's baked directly into how large language models are trained. AI models predict the most satisfying, logical next word based on massive datasets of internet text. When building a story, the model naturally optimizes for high-probability paths — which means it leans toward predictable structures and tidy resolutions. It's programmed to provide answers, not sit with discomfort.
Human life is full of loose ends and contradictions. Human writers deliberately capture those qualities because they resonate. AI views them as statistical anomalies to be smoothed over.
The result? Characters that feel safe. Predictable. Forgettable. They arrive at their destinations with all their questions answered and all their contradictions resolved, as if they'd been through some kind of narrative therapy and emerged perfectly adjusted. Real people don't work that way.
Why Bigger Models Don't Fix This
Here's where the study gets really interesting — and honestly, a bit embarrassing for the AI industry.
You'd expect that as models get bigger and more powerful, they'd produce richer characters. More nuance. More complexity. More mystery.
They don't.
The researchers found that massive, state-of-the-art flagship LLMs generated characters just as flat and archetypal as those produced by significantly smaller, less complex models. Co-author Nicholas Sanaie noted that this tells us the challenge isn't about scale at all. It's about how these models understand storytelling itself.
This matters because the entire AI industry has been operating on the assumption that more parameters equals better output. CASPER suggests that for narrative depth, that assumption is wrong. We're not dealing with a processing power problem. We're dealing with a fundamental misunderstanding of what makes fiction compelling.
The deficit is rooted in how models process narrative, not how much data they've consumed.
What Human Writers Do Differently
When the CASPER framework analyzed human-authored fiction, it revealed something remarkable: a high comfort level with chaos.
Human writers regularly leave characters unresolved. They make them morally gray. They keep them open to interpretation. They allow contradictions to stand without resolution. And these aren't accidents or gaps in craft — they're intentional choices that make stories linger.
Ambiguity is what makes a story stick with a reader. It's the difference between finishing a novel and feeling haunted by one.
The study systematically mapped character behavior across all eight dimensions, tracking the precise transition from hyper-exaggerated caricatures to realistic individuals. Human writers consistently produced characters that occupied the messy middle ground — somewhere between fully known and completely unknowable. That tension is where emotional resonance lives.
AI writers, by contrast, treated narrative uncertainty as noise to be eliminated. They smoothed it out. They resolved it. They made everything make sense.
And in doing so, they made everything forgettable.
What This Means for Writers Using AI
Let me be clear: this study doesn't say novelists and screenwriters should abandon AI tools entirely. The researchers view AI as an incredibly capable creative partner — one that requires a firm, human hand at the wheel.
AI is fantastic for brainstorming plot outlines. It cures blank-page syndrome. It flesh out simple background descriptions quickly. For these tasks, it's genuinely useful.
But you cannot delegate the soul of character development to a machine. If a novelist lets an AI write their main characters unchecked, those characters will inevitably turn into flat, safe clichés. The true magic of storytelling still requires a human writer willing to step in and intentionally mess up the code by adding unresolvable flaws, messy contradictions, and a healthy dose of genuine mystery.
Think of AI as your brainstorming buddy, not your co-author. Let it help you generate options. Then choose the messier, more interesting path — the one that makes you slightly uncomfortable because it doesn't quite resolve.
CASPER as a Benchmark for the Industry
Beyond exposing creative limits, CASPER functions as a vital standardized benchmarking framework. It enables AI developers and creative studios to evaluate whether upcoming next-generation models are genuinely advancing narrative depth and character complexity rather than simply becoming more grammatically fluent.
Right now, the industry measures AI writing quality in terms of coherence, grammar, and logical consistency. CASPER adds a new dimension: does this model understand that good fiction requires uncertainty? That compelling characters should remain partially unknowable?
As more people collaborate with AI to write novels, screenplays, and other creative works, we need ways to understand both what these systems do well and where they fall short. CASPER gives us a lens for evaluating character depth and diversity — which can ultimately help developers build storytelling systems that better reflect the complexity of human experience.
The question isn't whether AI can write. It's whether it can write characters worth remembering.