Academia has a new preprints problem

Academia has a new preprints problem. Thousands of researchers are essentially playing slot machines with large language models, pulling the lever over and over hoping to hit the jackpot of novel theoretical physics insights. As Figshare acts as a preprint platform, I'm watching this unfold in real-time. It's equal parts fascinating and maddening.

The strategy is simple. Fire up Claude or ChatGPT, pepper it with increasingly esoteric questions about quantum mechanics or string theory, collect the responses, wrap them in LaTeX, and boom, you've got yourself a preprint. It's the academic equivalent of the infinite monkey theorem, except instead of waiting for monkeys to accidentally type Shakespeare, we're waiting for transformer models to accidentally solve the mysteries of the universe. I get the appeal. LLMs are genuinely impressive at synthesizing existing knowledge, making unexpected connections, and occasionally producing those "huh, I never thought about it that way" moments. The barrier to entry is practically non-existent. There is no need for expensive lab equipment, years of mathematical training, or even a particularly deep understanding of the field. Just you, a chatbot, and the audacity to ask "what if gravity is actually just electromagnetic force wearing a disguise?" The problem isn't that this approach can't work. Broken clocks, infinite monkeys…pick your metaphor. 

The issue is the signal-to-noise ratio is absolutely abysmal. 

For every potentially interesting nugget buried in these preprints, there are hundreds of pages of what amounts to sophisticated-sounding nonsense. LLMs are, at their core, pattern-matching machines trained on existing human knowledge. They're exceptional at remixing and recombining ideas, but genuine novelty? What's particularly frustrating is watching legitimate repositories get flooded with these papers. They're not exactly wrong. Many are internally consistent, use appropriate terminology, and follow logical structures. But they're also not exactly right, or useful, or advancing human understanding in any meaningful way. 

The researchers doing this seem to fall into two camps. First, there are the true believers who genuinely think they're onto something revolutionary. They'll point to that one time an LLM suggested a novel approach to protein folding or helped solve a mathematical proof as evidence that their method is valid. Fair enough, but those were cases where humans used LLMs as tools within rigorous frameworks, not as oracles dispensing wisdom.

The second camp is more cynical. They know the game they're playing. In the publish-or-perish world of academia, quantity often trumps quality, and what better way to pad your publication list than with an endless stream of LLM-generated "research"? The effort-to-output ratio is unbeatable. Why spend years on one groundbreaking paper when you can produce dozens of mediocre ones in the same timeframe?

However, this approach fundamentally misunderstands both how scientific breakthroughs happen and what LLMs actually do. Real theoretical physics advances don't come from randomly combining existing concepts, they come from deep understanding, careful observation, mathematical rigor, and often, years of wrestling with paradoxes and inconsistencies. Einstein didn't stumble upon relativity by asking enough questions; he spent a decade thinking deeply about the nature of space and time. Meanwhile, LLMs are essentially very sophisticated autocomplete systems. They predict what words should come next based on patterns in their training data. When you ask them about theoretical physics, they're not reasoning from first principles or accessing some hidden understanding of the universe, they're pattern-matching against every physics paper, textbook, and Wikipedia article they've seen. It's impressive mimicry, but it's still mimicry.

I love LLMs and think we are underestimating how they will change academic research. These tools have legitimate uses in research: literature reviews, code generation, exploring analogies, even brainstorming. But there's a difference between using them as tools and treating them as research partners. One enhances human intelligence; the other substitutes for it.

Will this approach eventually produce something genuinely novel? Maybe. Infinite monkeys and all that. The real breakthroughs in theoretical physics will probably come from the same place they always have: clever humans doing hard work. Until then, I'll be here, watching repositories fill up with papers asking whether dark matter might be conscious, and wondering if this is really the best use of our collective intellectual energy.