10/26/2026, 11:57 AM
The idea is fire, per se, IMHO. The problem is a personal one: I still don’t know what to make of all the “plastic-fraud” jargon that LLMs inject into anything remotely technical-looking.
I recently read two papers that were clearly AI-generated by free or low-cost models. Everything sounded so good that it couldn’t be true. Some parts of the analysis and the overall inferences seemed fishy, especially in areas of mathematics I wasn’t familiar with. But when the papers got to something I knew I was strong in, I immediately saw how shallow and incomplete the reasoning was, dressed up in perfectly conventional mathematical jargon.
It’s already hard enough to spot errors in formal proofs written by humans. It’s even easier to miss subtle details in “pre-formalized proofs” or discursive proofs, that is, writing that’s half math and half mathematical English and those proofs make 90% of preprints and peer-reviewd literature. I wonder how many errors are hidden behind LLMs’ pristine English grammar, perfectly stochastic style, and good-taste writing practices.
That said, I’m fine with AI tools summarizing and automating semantic classification and tagging tasks, as well as many others tasks. But an expert eye should always check the final result and take responsibility for what it represents and what it’s meant to serve. For example, I’m looking at the fractional-ranks section. I’m sure I’m not smart enough to follow everything, but I can say for sure that the material isn’t explained very well. A more intuition–definition–proposition–proof approach would benefit the “human” reader. Another weak point in the recent use of AI tools to produce mathematical discourse has often been the lack of pedagogical and intuitive friendliness. Just my two cents.
I recently read two papers that were clearly AI-generated by free or low-cost models. Everything sounded so good that it couldn’t be true. Some parts of the analysis and the overall inferences seemed fishy, especially in areas of mathematics I wasn’t familiar with. But when the papers got to something I knew I was strong in, I immediately saw how shallow and incomplete the reasoning was, dressed up in perfectly conventional mathematical jargon.
It’s already hard enough to spot errors in formal proofs written by humans. It’s even easier to miss subtle details in “pre-formalized proofs” or discursive proofs, that is, writing that’s half math and half mathematical English and those proofs make 90% of preprints and peer-reviewd literature. I wonder how many errors are hidden behind LLMs’ pristine English grammar, perfectly stochastic style, and good-taste writing practices.
That said, I’m fine with AI tools summarizing and automating semantic classification and tagging tasks, as well as many others tasks. But an expert eye should always check the final result and take responsibility for what it represents and what it’s meant to serve. For example, I’m looking at the fractional-ranks section. I’m sure I’m not smart enough to follow everything, but I can say for sure that the material isn’t explained very well. A more intuition–definition–proposition–proof approach would benefit the “human” reader. Another weak point in the recent use of AI tools to produce mathematical discourse has often been the lack of pedagogical and intuitive friendliness. Just my two cents.
Mother Law \(\sigma^+\circ 0=\sigma \circ \sigma^+ \)
\({\rm Grp}_{\rm pt} ({\rm RK}J,G)\cong \mathbb N{\rm Set}_{\rm pt} (J, \Sigma^G)\)