A panel of leading mathematicians has drawn a blunt line around what today's large language models can actually do with math: they are dependable at computation and pattern-matching, but they falter the moment a problem calls for a genuinely new idea rather than a known technique applied harder.

That distinction matters because the past two years of AI marketing have leaned hard on math as proof of reasoning ability — gold-medal claims at international olympiads, benchmark leaderboards, demos of models solving textbook proofs in seconds. The mathematicians' verdict cuts through that narrative: solving a hard problem with a known playbook is not the same as inventing the playbook.

According to The Decoder, the mathematicians frame current LLMs as strong calculators — fast, tireless, and reliable within the boundaries of established methods — but poor creative thinkers when a problem has no template to follow. That framing should shape how builders think about deploying these models on anything math-adjacent.

Where the line actually falls

Computation is procedural: apply a known algorithm, follow the steps, check the arithmetic. LLMs, trained on enormous volumes of worked examples, are very good at this — symbolic manipulation, plugging numbers into formulas, recognizing which of a few thousand standard techniques applies to a given problem shape. That's most of what shows up in textbooks and most of what shows up in benchmark suites built from competition problems with known solution paths.

Creativity in mathematics is different. It means noticing a connection nobody has written down, choosing a strange substitution, or reframing a problem in a domain it doesn't obviously belong to. There's no corpus of "how to have this specific insight" for a model to pattern-match against, because by definition the insight didn't exist until someone had it. That's the gap the mathematicians are pointing at, and it lines up with what researchers have quietly noted for a while: models do well on problems that resemble something in their training distribution, however distantly, and do worse as problems move toward genuinely open or under-explored territory.

Why the gap is easy to miss

The confusion happens because both modes of math produce the same surface output: a clean, confidently written solution. A model that has correctly applied a memorized technique and a model that has genuinely reasoned its way to a novel one both hand you a tidy proof or a final number. Without checking the reasoning step by step — or without already knowing the answer — it's hard to tell which one you got. That ambiguity is exactly what has let benchmark scores and demo videos stand in for a harder question: is the model actually thinking, or retrieving?

What this means for anyone building on top of LLMs

For teams shipping AI products in math-adjacent domains — tutoring, engineering calculations, financial modeling, scientific research assistants — the mathematicians' distinction is a practical design constraint, not an abstract philosophical point:

• Trust the model for computation, verification, and translating a known method into code or a worked solution — that's where the error rate is lowest.
• Don't trust it to originate a genuinely new approach, a novel proof strategy, or an unprecedented mathematical framing without a domain expert reviewing the reasoning, not just the answer.
• Watch for confident wrong answers on out-of-distribution problems — the model's tone doesn't change when it's guessing versus when it's applying a technique it actually understands.
• Use LLMs as a fast first pass on hard problems — generating candidate approaches for a human to prune — rather than as the final authority on whether a proof or a model is correct.

AiiN's takeaway

The mathematicians' framing is useful precisely because it doesn't dismiss LLMs — it locates their value correctly. A calculator that never gets tired, never makes arithmetic slips, and can hold a full derivation in its head is genuinely useful to a working mathematician or engineer. The mistake is expecting that same tool to also supply the spark that starts a new line of research. In our estimation, the more realistic near-term role for LLMs in math and adjacent technical fields is as an amplifier for human creativity rather than a replacement for it — a fast collaborator that handles the grinding middle of a proof while a person still has to supply the opening move.