Why AI's Creative Writing Feels Flat, and What's Behind It

JUST IN: We've all noticed it. Ask a large language model to write a story and it delivers something that feels a bit, well, flat. But what's going on under the hood?

The Study's Breakdown

A recent study tackled this by analyzing story continuations from the OLMo 32B language model at four different checkpoints: Base, SFT, DPO, and RLVR. These variations share the same architecture, scale, tokenizer, and pretraining, so what differentiates them? It's the post-training effect. That's where the flattening kicks in.

The researchers used three types of story sources: StoryStar (a public platform), TMAS (prompt-guided), and The New Yorker (the big leagues of literary fiction). They compared the AI's story continuations with human-written texts along three dimensions: thematic motion, affective prevalence, and linguistic diversity.

Narrative Flattening: The AI Effect

Here's the skinny. Post-training seems to compress what's dynamic. Think of it like smoothing out all the bumps in a road. Transitions in themes become samey. High-intensity emotions? They fizzle out into neutrality. Style? It gets monotonous.

But here's the kicker, professional literary fiction suffers the most from this flattening. Why? Human writers in this domain take risks, play with style, and embrace strong emotions. AI, post-training, does the opposite. Public-platform and prompt-guided stories? They fare a bit better. Their human versions are already closer to the model's default rhythm.

Why Does This Matter?

So, why should you care? Because this pattern has implications for how we use AI in creative industries. If we're relying on AI to generate creative content, we might want to rethink the training tweaks. Who wants a world where all AI stories sound the same?

Sources confirm: Post-training alignment converges everything into a one-size-fits-all model. And just like that, the leaderboard shifts in how we perceive AI's role in creativity. It's a reminder that while AI can mimic, true creativity still needs the human touch.

So, the question is: Are we ready to accept a homogenized AI narrative, or will we demand more diversity and depth from our robotic scribes?