In short

The 'AI boring law' describes how, as LLMs scale, they become more homogeneous, in outputs and in how they fail. Their architecture applies downward pressure from all digitised knowledge to reduce uncertainty, the opposite of creativity, and RLHF amplifies this. Cornell researchers find errors grow more correlated with scale, and LLMs lack Keats's 'negative capability'. Rather than treating homogenisation as inevitable, the essay argues for embedding LLMs in structured creative systems that add friction and diversity.

The better LLMs get, the more they sound alike, and the more they make the same mistakes. This essay names that tendency the “AI boring law” and explains why it’s baked into how the technology works.

A language model’s whole job is to reduce uncertainty about the next token, drawing on the average of everything it has read; RLHF then sands off the remaining edges. That’s close to the opposite of what Keats called “negative capability”, the capacity to sit with uncertainty and contradiction that genuine creativity depends on. Cornell research backs this up at the system level: as models scale, their errors become more correlated, meaning they fail in herds.

The constructive turn is the interesting part. The answer to AI homogenisation isn’t to brainstorm harder with ChatGPT; it’s to embed models inside structured systems, like MIT’s Supermind Ideator or DeepMind’s AlphaEvolve, that deliberately reintroduce friction and diversity.

Key takeaways

  • The 'AI boring law': more powerful LLMs become more homogeneous in outputs and failure modes.
  • LLMs reduce uncertainty by design and RLHF amplifies convergence, they lack Keats's 'negative capability'.
  • Cornell research shows model errors become more correlated as models scale, so they fail homogeneously.
  • AI may raise individual creativity while lowering collective creativity by homogenising the pool of ideas.
  • The fix is structured systems (MIT Supermind Ideator, AlphaEvolve, AutoTRIZ) that add friction and diverse perspectives.

Read the full piece

This is a summary. Read the complete essay, with all the sources and argument, on Substack.

Read on Substack →

Frequently asked questions

What is the 'AI boring law'?
The idea that as LLMs grow more powerful they become more homogeneous, producing convergent outputs and correlated errors, which makes them poor tools for genuine creativity.
Why are LLMs bad at creativity?
Their architecture applies downward pressure from all digitised knowledge to reduce uncertainty, and RLHF amplifies this convergence, the opposite of the uncertainty-embracing 'negative capability' creativity requires.
How can AI be used for creativity without homogenising ideas?
By embedding LLMs in structured systems, like MIT's Supermind Ideator, AlphaEvolve, or AutoTRIZ, that deliberately add friction and diverse perspectives rather than amplifying the obvious first idea.

People & ideas in this piece

John KeatsEthan MollickBernard StieglerCornell UniversityMIT Supermind IdeatorAlphaEvolveAutoTRIZRLHFNegative capabilityGenerative monoculture

Topics: How AI Actually Works , AI, Time & Society