Headlines about AI "solving the unsolvable" in mathematics make for great clicks. They also make for terrible engineering decisions.
Every time a model clears a high-profile benchmark, scores medal-equivalent marks on a competition test, or spits out a formal proof for a nasty little problem, the lazy consensus kicks into overdrive. Suddenly, the story goes that artificial intelligence is storming the citadel of pure math, cracking open problems that stumped humanity for generations, and turning centuries of human intuition into a commodity.
That story is backward. Wrong framing. Wrong metric. Wrong diagnosis.
AI is not solving the unsolvable. What it is doing—often with terrifying speed and brute computational grit—is chewing through hard problems where the rules of the game are already crystal clear, the search space is vast, and the bottleneck is either tedious casework, pattern recognition over massive symbol trees, or mechanical verification. Calling routine hard problems "unsolvable" because a human hasn't bothered to grind through ten thousand cases yet is a category error dressed up as a miracle.
I have watched research labs and enterprise teams burn millions of dollars chasing the hallucination that a reasoning model can simply "think up" a genuinely new mathematical field the way a human mathematician shifts the ground beneath an entire discipline. They build bloated pipelines around brittle proof assistants, hire expensive prompt engineers to coax out step-by-step derivations, and act stunned when the system hits a wall the moment a problem requires conceptual invention rather than clever search.
To understand what is actually happening when AI collides with mathematics, you have to stop treating "hard" and "unsolvable" as synonyms. More importantly, you have to ask whether you are even looking at the right question in the first place.
You Are Asking the Wrong Question About AI and Math
People love asking: Can AI solve the hardest problems in mathematics?
That question is a trap. It treats mathematics like a giant locked safe where all you need is a strong enough crowbar or a smart enough locksmith to pop the tumblers. Real mathematics doesn't work like a safe. Often, nobody even knows which room the safe is in, what language the combination is written in, or whether the whole building needs to be torn down and rebuilt on different foundations.
A hard problem has a well-posed target. You know what an answer looks like. You may not know the path yet, or the path may be astronomically long, or the trick might require spotting an obscure bridge between two areas, yet the objective is defined. Think of searching through a colossal maze where the exit exists and every wall is solid. A machine that explores ten million corridors a second has a distinct advantage over a human pacing the halls with a flashlight.
An unsolvable problem, by contrast, splits into two completely different beasts:
- Provably undecidable or uncomputable problems: Things like the Halting Problem, established by Alan Turing in 1936, or Hilbert's tenth problem on Diophantine equations, resolved negatively by Yuri Matiyasevich in 1970 (building on work by Martin Davis, Hilary Putnam, and Julia Robinson). No algorithm—not today, not with a trillion parameters, not with a Dyson sphere full of GPUs—can decide every instance in general. More compute does not repeal mathematical impossibility.
- Problems blocked by missing concepts or paradigm shifts: Situations where progress isn't waiting on faster tree search; it is waiting on someone to invent a new object, a new invariance, or a new structural lens. When Andrew Wiles proved Fermat's Last Theorem in the 1990s (with crucial work from Richard Taylor), the breakthrough wasn't just grinding through modular forms faster. It was stitching together deep machinery—elliptic curves, modular forms, Galois representations—and forging new mathematical bridges to attack a statement that had sat open for over three centuries.
When a competitor article claims AI is "solving the unsolvable," it blurs those boundaries into mush. A benchmark problem that takes a bright undergraduate three hours under exam pressure is not an unsolvable mystery of the cosmos. Even a formidable open conjecture solved with machine assistance is still a tractable hard problem once the right representation or search strategy is found.
Asking whether AI can solve "unsolvable" math is the wrong question because it mistakes computational power for conceptual creation. The real question is far more practical—and far more uncomfortable for hype merchants: Where does machine search and formal verification actually beat human bottlenecks, and where does it hit a brick wall because there is no clean target to optimize?
How Search and Formal Proof Actually Work Under the Hood
Strip away the sci-fi gloss and look at the mechanics. When AI makes real mathematical progress, it rarely wins by having a mystical epiphany in a quiet room. It wins by pairing statistical intuition with strict formal rules or massive search.
Let's define the terms cleanly before people start throwing buzzwords around:
- Informal math is how humans usually talk and write proofs: prose, diagrams, leaps of intuition, implied assumptions, and "it is easy to see that..." statements. Easy to read, dangerously easy to smuggle a subtle bug into.
- Formal math translates statements and proofs into a machine-checkable language—using proof assistants like Lean, Coq (now often called Rocq), or Isabelle—where every single logical step must follow strict inference rules from accepted axioms. No hand-waving allowed.
- Heuristic search / reinforcement learning explores a space of candidate moves, conjectures, or proof tactics, using a score or verifier signal to prune dead ends and double down on promising branches.
Think about why that combination is potent in specific regimes. In pure natural language, a large language model can sound extraordinarily convincing while quietly inventing a false lemma on line four and building a castle of nonsense on top of it. Anyone who has spent serious time testing LLMs on Olympiad-style or research-adjacent problems has seen that exact failure mode: fluent prose, plausible notation, fatal logical rot in the middle.
Pair a model with a formal verifier or a structured search loop, and the game changes. Now the machine cannot bluff its way past a type mismatch or an unjustified step. Every candidate step gets checked against ironclad rules. If the system searches over proof states, tests millions of combinations, or discovers narrow counterexamples and edge cases faster than a human team can type them, it can crack problems that look intimidating on paper.
We have seen variations of this dynamic in computer science and math for years—from automated theorem proving and SAT/SMT solvers to AlphaProof and AlphaGeometry-style systems and exploratory math tools used alongside researchers like Terence Tao and Timothy Gowers. Notice what makes those wins real: verifiability and searchability.
Imagine a scenario where a research team wants to find a tight bound or a counterexample in a combinatorial structure with billions of small configurations. A human mathematician might spend weeks testing dozens of cases by hand or writing quick scripts to spot a pattern. A search-guided AI system can probe millions of structured variations, flag anomalies, propose candidate lemmas, and let a formal checker or script test whether the pieces hold. That isn't magic. That is industrial-scale hypothesis generation and verification.
Now flip that scenario. Imagine a scenario where a team hands an AI a sprawling conjecture in algebraic geometry or number theory and says, "Invent the right cohomology theory or deformation argument needed to bridge this gap." Suddenly there is no neat local reward signal. A dead end ten steps deep might look completely fine at step two. Worse, the concept needed to make the proof natural might not even have a standard name yet.
A machine trained to predict tokens or optimize within a defined search space excels at navigating known terrain fast. It struggles mightily when the map itself is missing a continent.
Dismantling the "People Also Ask" Myths
Whenever a splashy math-and-AI headline hits the feed, the exact same set of questions pops up everywhere. Most of them rest on shaky premises. Let's take them apart one by one.
"Will AI replace human mathematicians?"
No—at least not the ones doing actual mathematical creation. If your job consists of routine algebraic manipulation, standard calculation, or grinding through predictable cases, you should already be nervous. If your job is figuring out what is worth proving, spotting hidden analogies between unrelated fields, questioning hidden assumptions in a definition, and inventing new frameworks, you are not being replaced anytime soon.
Replacing human mathematicians with pure AI assumes mathematics is just a giant pile of answers waiting to be dug up. In reality, much of the hardest work in math is problem formulation and concept design—deciding which questions are deep rather than merely hard, and building the language in which the answer becomes transparent. A tool can accelerate proof discovery without replacing the human judgment that gives the work meaning or direction.
"Does beating math benchmarks mean AI truly understands mathematics?"
Benchmark scores measure performance on a test set, not deep conceptual mastery. A model can score sky-high on competition problems—which are often designed to have clever tricks within a bounded timeframe and known syllabus—while still failing basic consistency checks on slightly altered problem statements or struggling with open-ended research tasks where there is no answer key at the end of the book.
Treating benchmark wins as proof of general mathematical "understanding" is like assuming a sprinter who breaks a record on a straight track automatically knows how to design a better suspension system for a rally car. Fast execution on a benchmark is a signal worth tracking, not a blank check to declare victory.
"Can AI solve Millennium Prize Problems or famous open conjectures?"
Don't hold your breath for a silver-bullet announcement where a single prompt drops a clean, indisputable proof of the Riemann Hypothesis or P vs NP tomorrow morning. Problems that have resisted top minds for decades—even centuries—usually sit at intersections of deep theory where progress requires years of conceptual synthesis, community scrutiny, and fresh mathematical machinery.
Could AI play a supporting role—surfacing a helpful lemma, exploring a sub-case, formalizing a portion of a known argument, or helping spot a pattern? Yes, potentially. Could it single-handedly bulldoze through every deep open problem by sheer scale alone? Nothing in computer science or mathematics supports that bet.
"Is formal verification a silver bullet that fixes AI math hallucinations?"
Formal verification fixes logical validity of the checked proof, not relevance, beauty, or alignment with what you actually meant to formalize. If you formalize the wrong statement—say, a slightly weaker version of a theorem or a definition with an unintended edge case—a proof assistant will happily verify a rock-solid proof of the wrong thing.
Formal tools are guardrails, not steering wheels. They catch bad steps; they don't tell you whether you are driving toward a breakthrough or a ditch.
Battle Scars: Where Teams Waste Money on AI Math Hype
I've seen companies and labs blow serious budget on "AI for math and reasoning" initiatives because leadership confused press releases with engineering reality.
One common mistake is buying a shiny reasoning model or agentic wrapper and expecting it to act like an autonomous research mathematician out of the box. Teams feed it dense PDFs, open-ended research notes, or messy domain specifications, then act shocked when it produces plausible-sounding garbage with subtle sign errors, missing edge cases, or circular logic. Without a tight verification loop—whether a formal proof checker, a unit-test harness, a symbolic engine, or domain experts checking every step—raw generation scales plausible error just as fast as it scales useful drafts.
Another expensive trap is over-engineering formalization pipelines for problems that don't actually need heavy formal machinery. Translating complex mathematical ideas into formal proof languages takes real expertise and serious compute. If your actual goal is numerical simulation, optimization, algorithm tuning, or data analysis, forcing everything through a rigid theorem-proving stack can burn months of engineering time for marginal payoff.
Even when formal proof assistants make complete sense—such as high-assurance software, cryptographic protocols, hardware verification, or checking intricate combinatorial arguments—teams sometimes forget the human bottleneck: reading and trusting the output. A formal proof that passes a checker might still be ugly, brittle, or hard for humans to extract conceptual insight from. If nobody on your team can read the proof or verify that the formal specification matches reality, you haven't eliminated risk; you've just moved it into the specification layer.
Where My Contrarian Take Has Real Blind Spots
Being contrarian is easy if you ignore every counter-argument. It gets honest when you admit where your own skepticism could bite you.
First, progress in search and formal reasoning is moving fast, and what looks like a hard wall today can become a routine workflow in a few years. Systems that combine neural intuition, tree search, program synthesis, and formal proof assistants are improving at tackling multi-step reasoning and finding non-obvious proof paths. Underestimating compounding engineering improvements is a classic rookie mistake.
Second, even "non-creative" or search-heavy mathematical work has huge practical value. You don't need an AI to invent a brand-new branch of topology for it to save researchers weeks of grunt work, uncover subtle edge cases in algorithms, or help formalize complex proofs. Dismissing something because it isn't a paradigm shift misses how much real-world work happens in the trenches of checking, searching, and refining.
Third, human intuition itself is often overrated in hindsight. Mathematicians are brilliant at telling a clean, heroic story after a proof is finished, smoothing over dead ends, lucky guesses, and brute-force checks. Sometimes a machine finding a messy, non-intuitive proof—or spotting a pattern humans overlooked because of ingrained habits—is genuinely valuable even if it doesn't match our romantic picture of "genius."
What to Do Instead of Chasing Hype
If you are building with AI, funding research, or trying to separate real capability from marketing noise, stop asking whether AI is "solving the unsolvable." Start running a sharper checklist on every claim and every project:
- Demand a verifier, not just a confident answer. Never trust unverified multi-step mathematical or logical output in high-stakes work. Pair generation with a formal checker, a symbolic solver, deterministic code execution, or strict human audit.
- Match the tool to the structure of the problem. Use search-and-verification systems where rules are explicit, states can be checked, and candidate paths can be pruned or scored. Don't expect magic on vague, ill-defined problems with no clear objective.
- Audit the specification before celebrating the proof. Check that the formal statement or problem setup actually captures what you care about. A verified proof of a subtly wrong model is still a failure in the wild.
- Measure time-to-verified-truth, not benchmark bragging rights. Track how fast your team gets a checked, reliable, usable result—including human review time—rather than getting hypnotized by isolated test scores.
- Keep humans focused on framing, taste, and synthesis. Let machines grind through search spaces, propose candidates, and check tedious steps. Keep human experts steering the questions, checking conceptual coherence, and deciding what matters.
AI is reshaping how we explore hard problems, check proofs, and search vast spaces of possibilities. That is already a big deal—big enough that you don't need to lie to yourself and call it solving the unsolvable.
Check the verifier. Question the headline. Build for what actually holds up under pressure.