Mathematicians and AI in behind-the-scenes battle over what’s true
AI models are solving mathematics problems with increasing pace, and a technique called formalisation is key to demonstrating that their claimed solutions are indeed correct. But can we trust the formalisation process?
Hidden Truths · AI Analysis
Mainstream Narrative
AI systems are rapidly advancing in mathematical problem-solving, but their solutions require verification through "formalisation" — a rigorous process that translates proofs into computer-checkable logic. The emerging question is whether this verification method itself is trustworthy.
Missing Context
This story sits at the intersection of several decades-long developments:
1. **Proof assistants** (software like Lean, Coq, Isabelle) have existed since the 1970s-80s, used by mathematicians to verify complex proofs like the Four Color Theorem and Kepler Conjecture.
2. **The replication crisis** in mathematics is minimal compared to other sciences — peer review has traditionally been robust, though highly complex modern proofs (like the classification of finite simple groups) have raised concerns about human verification limits.
3. **DeepMind's AlphaProof and similar systems** recently solved International Mathematical Olympiad problems, marking genuine breakthroughs but representing narrow problem domains, not general mathematical creativity.
4. **Formalism vs. intuition debate** has raged since the early 20th century (Hilbert vs. Brouwer) — whether mathematics is purely symbolic manipulation or requires human insight.
Bias Analysis
New Scientist typically favors **tech-optimist framing** with cautionary undertones — "AI does amazing thing, but here's the wrinkle." The headline's "battle" language dramatizes what is actually a methodological debate within mathematics and computer science communities. The framing assumes readers find AI progress both exciting and slightly threatening, which drives engagement but may overstate actual controversy. The publication tends toward accessibility over technical precision, potentially oversimplifying the distinction between AI-generated conjectures and AI-verified proofs.
Counter-Narratives
1. **The "nothing new here" perspective**: Mathematicians like Timothy Gowers argue formalisation simply makes explicit what peer review already does — it's quality assurance infrastructure, not a philosophical crisis. AI tools are assistants, not replacements.
2. **The human-proof primacy view**: Some mathematicians insist that formal proofs lack the explanatory power and insight of traditional proofs. A correct proof isn't just mechanically valid — it should illuminate *why* something is true. Formalisation may verify correctness while obscuring understanding.
3. **The "AI hype cycle" interpretation**: Critics note that AI mathematical achievements remain narrowly scoped. Most claimed "breakthroughs" involve well-defined competition problems, not the creative conjecture-generation and paradigm-shifting insights that drive real mathematics forward.
Alternative Angles (Speculative)
Some technoskeptics speculate that **trust in formalisation creates new vulnerabilities** — if proof-checking software contains subtle bugs or backdoors, false results could be certified as true, and the perceived rigor might prevent human scrutiny. This echoes broader concerns about "automation bias."
Fringe theorists in the "AI existential risk" community worry that **mathematically sophisticated AI could discover exploitable logical inconsistencies** in formal systems themselves (echoing Gödel's incompleteness theorems), potentially undermining foundations of mathematics or cryptography in ways humans couldn't detect.
Some critics argue this represents **credentialism transfer** — shifting mathematical authority from human expert communities to corporate-owned AI systems and their proprietary training processes, with implications for who controls mathematical knowledge production.
Fact-Check Flags
What To Read Next
1. **Primary technical sources**: Papers from the Lean mathematical community and DeepMind's AlphaProof publications to understand actual capabilities versus media framing.
2. **Philosophical perspectives**: Articles by mathematicians like William Thurston (on proof and progress) or Kevin Buzzard (leading Lean advocate) discussing what proofs are *for* beyond mere verification.
3. **Critical computer science literature**: Research on software verification limits, formal methods, and documented cases where proof assistants failed or required correction — provides crucial reality-check on reliability claims.