We caught it in a prompt-tuning session, buried in a note about our own generation pipeline: every post it produced defaulted to the same "tech blog aesthetic." Same rhythm, same structural beats, same safe hedge in the third paragraph. Nobody told the model to do that. It just found the path of least resistance and stayed on it.
That's the asymmetry in a nutshell. Producing something derivative costs almost nothing. Producing something original costs a search -- through a much larger, much less charted space.
Copying is a decision, originality is a search
Originality is usually defined by what it isn't: not a reproduction, not a clone, not a substantially derivative work. That's a negative definition, and negative definitions are a giveaway. It's far easier to describe the boundary of sameness than to describe the shape of something new, because sameness has a template and novelty doesn't.
This shows up mechanically in any content or generation system. Given a prompt and a corpus, a language model can produce a thousand grammatically fine, thematically safe variations in the time it takes you to read this sentence. Producing the one variation that says something nobody else has said -- that takes a different kind of work. It takes a gap to fill, not a pattern to match.
We built a system for exactly this reason: a pipeline stage that runs topic-gap analysis against our own knowledge base before generation starts, so the model isn't just matching the nearest neighbor in its training distribution -- it's told where the neighbors are thin. That's the whole trick. You don't ask an LLM to "be original." You point it at the part of the map that's actually empty.
What physics can tell you about asymmetry
Martin Gardner spent an entire book on the observation that the universe defaults to symmetry, and that the rare cases of broken symmetry -- parity violation, the chirality of molecules -- are the interesting ones precisely because they're exceptions. Symmetric states are common because they're the low-energy, low-effort configuration. Breaking symmetry takes a specific force acting in a specific direction.
Content pipelines behave the same way. Left alone, a generation system settles into its symmetric state: the tech blog aesthetic, the five-paragraph structure, the hedge in paragraph three. Getting it to break that pattern takes a deliberate push, not a bigger model and not a longer prompt.
The myth we keep telling ourselves
There's a persistent argument that originality itself is mostly illusion -- that every apparent novelty is a recombination of what came before, that nothing an artist makes is un-borrowed. Fine. That's probably true, and it doesn't change the asymmetry at all. Recombination is still a search problem. Some recombinations are near, cheap, and obvious. Others are far, expensive, and rare. Calling both of them "just recombination" doesn't make them equally easy to find.
There's also a real cost to over-indexing on novelty for its own sake -- chasing originality as a goal can produce paralysis instead of output. That's a fair warning for indie devs too: don't rewrite your renderer from scratch because "everyone does it that way." Ship the derivative 90% without apology. Spend your actual originality budget on the 10% that's load-bearing.
Spend the budget where it counts
If you're running any kind of generative system -- content, code, level design -- assume it will converge to the symmetric, cheap, derivative state by default. That's not a failure mode, it's physics. Your job is to find the specific place worth pushing, point the system at it, and let the rest fall into its natural, unoriginal, perfectly serviceable groove.
Top comments (0)