How AI Post-Training Suppresses Creativity and Leads to Bad Writing

AI's post-training processes — RLHF, constitutional AI, and other alignment techniques — are systematically suppressing the creativity and whimsicality that made earlier models like GPT-2 feel surprising and alive, resulting in increasingly formulaic and dull writing from today's top AI models.

The Problem

Early language models like GPT-2 were celebrated for their unexpected, creative outputs — even when imperfect, they produced text that felt genuinely surprising. Today's more capable models, while technically superior, often produce writing that is technically correct but soulless.

The Culprit: Post-Training

The article argues that post-training alignment techniques are the primary cause:

• RLHF (Reinforcement Learning from Human Feedback) incentivizes safe, predictable outputs
• Constitutional AI adds rule-based constraints that limit expressive range
• Safety fine-tuning trains models to avoid controversy, which also means avoiding originality
• Helpfulness optimization pushes models toward generic, universally acceptable responses

The Creativity Paradox

More capable models produce worse creative writing because:

1. Safety constraints flatten the distribution of possible outputs
2. "Helpful" responses tend to be generic and unsurprising
3. Risk-taking in language (metaphors, unusual structures, provocative ideas) is penalized
4. The models learn to produce the most statistically "acceptable" response rather than the most interesting one

What We Lost

GPT-2 era outputs had a distinctive quality — they were often wrong, but they were rarely boring. The writing had personality, unpredictability, and a kind of accidental creativity that came from the model's limited but unconstrained generation.

The Trade-off

This isn't purely a criticism. The same alignment processes that suppress creativity also make models more reliable, safer, and more useful for most tasks. But it raises an important question: can we build AI that is both aligned and creative?

Source: The Atlantic via Techmeme | March 18, 2026