Lost on the Token-Room Floor

Is AI capable of “original thought”?

Everyone’s a critic — we arbitrate the “truly original” from the “entirely derivative.” We feel it in our bones (and sometimes, we write overwrought prose about it).

But the most successful IP often sits at the middle of these poles. Hit entertainment franchises (ex: Star Wars) and successful business models (ex: Uber for ____) are often defined by just the right amount of originality — twisting some elements within familiar frameworks so that people swallow the perfect cocktail of the novel and the known. (Read “Hitmakers” by Derek Thompson)

We know AI models are an infinitely combinatorial mirror of their training data. As users, we supply the lenses and directional gazes with our prompts, and AI now connects to the outside world — the internet, our shared drive documents, other agents — to inform its perspective. It can test ideas and code, refine its logic, and iterate. It can follow a scientific method. And as a result, AI has devised never-before-seen strategies to win at GO, map complex proteins or break into “secure” systems.

But if you ask most people if AI is “creative”? Most will say no — don’t expect anything “groundbreaking” from it. I think that’s just our way of saying that we view creativity as a distinctly human endeavor. Creativity is art, and we are the sole imperfect fountains of it. The tools may evolve, but creativity remains the raw conception of ideas, and the translation of ideas into forms. We value both the thought, and its translation process.

AI is so capable at being a translation layer, which is why we largely don’t value what it produces. It’s meaningful because a human organically worked at it.

Can AI be the seed, the essence, the DNA of the thought, and not just its conveyor belt? Can AI recognize in itself, through meta-cognition, its own sparks of brilliance?

The human brain has a unique ability to distinguish between our pedestrian thoughts and our brilliant ones. We know when to raise a well-reasoned objection to our boss, or call someone up to discuss the idea that just jostled us out of bed at 6am. We’ve needed this self-critical discernment to function in society, and it gets honed and validated through the “training data” of social acceptance or rejection — do others feel as we do about our ideas? Then strengthen that trust-your-gut synaptic connection.

AI receives feedback as well, but can it take that feedback to transform itself into a “truly original thinker”? “Originality” is an inherently subjective assessment based on the knowledge you bring to it, where the more expert you are, the better you should be at spotting it (where the inverse is often true of generating original ideas, with beginner’s mind).

Compared to us carbon forms, AI has infinite knowledge, but naturally pulls toward the average and optimizes for consensus. We don’t fully trust its judgment because true originality requires bravery — breaking the rules in some way, not just summarizing them.

Until AI develops a real mechanism for taste… until it can feel the friction between a safe idea and a groundbreaking one… it will still be the conveyor belt. The translator.

If AI is going to be a true source of creativity, it first needs to be able to discern the original from the derivative. Maybe it just needs our help to get there.

If only we could map the beautiful mysteries of human consciousness… now that is a job for AI.