That's cool -- I quite dig the idea of using it to generate conspiratorial content!

But I'm pretty sure Marcus is right; this sort of statistical pattern-recognition breaks down pretty fast, because it doesn't seem capable of any sort of abstraction.

You can really see that with numbers and math ... these models can do simple arithmetic, but quickly break down once the numbers get even slightly large -- because they're just absorbing/mimicking patterns they've seen written down of mathematics being done. They can only "do" math that they've seen before. But this isn't how humans (or even animals that appear to understand quantities, including bees) seem to process numbers; we're not just doing pattern recognition of math we've seen before -- we understand abstractly that 3 is a number, a quantity ...