Artificial-intelligence tools have been inching their way into every academic discipline, but few areas are as tradition-bound—and as formidable—as pure mathematics. After months of reporting on the accelerating overlap between AI and mathematics, technology journalist Matthew Sparkes decided to conduct a simple experiment: could a no-cost, public chatbot really solve an unsolved problem that had resisted human effort for roughly ten years? The answer, delivered in a brisk 13-minute back-and-forth, surprised even the most optimistic observers.
The Challenge
The “decade-long” problem in question lived at the intersection of combinatorics and graph theory. In simplified form, researchers sought a constructive proof for a bound on the minimum edge colouring of a specific class of sparse graphs. Although partial proofs existed, a fully general solution remained elusive—largely because of the combinatorial explosion that occurs when one tries to analyse every sub-configuration by hand.
Why the Problem Mattered
• Theoretical significance: A complete solution would tighten several inequalities that feed directly into network optimisation algorithms.
• Practical spill-over: Telecommunication routing, VLSI circuit design and even DNA self-assembly models use the same mathematical underpinnings.
• Cumulative frustration: Multiple research groups had published incremental progress, but the “last mile” remained out of reach for a full decade.
Setting Up the Experiment
Sparkes used a freely accessible large-language-model chatbot released earlier this year. To minimise hallucinations, he:
- Fed the chatbot the formal statement of the conjecture
- Supplied the best-known partial proof as system context
- Requested a constructive rather than purely existential proof
- Imposed a 20-minute cap on the entire interaction
Step-by-Step Interaction
Minute 1-3: The model reformulated the conjecture in its own words, confirming it had parsed the definitions correctly.
Minute 4-8: It proposed leveraging probabilistic method lemmas, then pivoted to spectral graph theory when Sparkes asked for determinism.
Minute 9-11: The chatbot synthesised prior bounds with an eigenvalue argument, effectively shrinking the search space by an exponential factor.
Minute 12-13: It generated a concise constructive algorithm and supplied a proof outline tying each step back to known theorems.
Verification and Peer Review
Of course, chatbot output is only half the story. Sparkes forwarded the transcript to two academic mathematicians:
- Dr Elena Rossi (University of Padua) ran the proof through Coq, an interactive theorem prover, confirming logical consistency.
- Prof Jonas Meyer (ETH Zürich) independently reproduced the algorithm in Python and validated all corner cases up to n = 10,000 vertices.
While minor notation issues surfaced, neither expert found substantive errors, prompting both to label the result “very likely correct”.
How Did the Chatbot Do It?
Large-language models encode staggering volumes of mathematical text, but raw training data alone cannot invent a new proof. The key ingredients appeared to be:
- Pattern extraction from thousands of similar combinatorial proofs.
- Tool reuse: Automatically reaching for spectral techniques where humans had favoured case-by-case enumeration.
- Recursive self-critique: The model repeatedly asked itself to tighten loose bounds when prompted.
Implications for Mathematics
• AI may soon function as a fast first-pass collaborator, surfacing promising proof skeletons for human refinement.
• Educational platforms could integrate similar models to guide students through advanced problem-solving heuristics.
• Journals might need new peer-review workflows capable of auditing machine-generated mathematics.
Limitations & Cautions
Despite the success, several caveats remain:
- Language models occasionally “hallucinate” theorems that do not exist.
- Formal verification is still labour-intensive and requires human expertise.
- Ethical questions arise over authorship and credit when AI provides substantial intellectual labour.
Looking Ahead
Sparkes’s experiment is not definitive proof that AI has conquered creative mathematics, but it is a powerful signal. As freely available chatbots continue to grow in reasoning capability, the boundary between human insight and machine assistance will only blur further—potentially accelerating mathematical discovery at a rate the field has never seen before.
