AI Is Better at Refuting Conjectures Than Proving Them, Argues Post

GlenBradley · x · 2026-08-25

The poster argues refuting a conjecture only requires one counterexample, while proving it demands a generalized principle covering all cases—so the former is far easier. He claims AI is currently better at finding what's wrong than proving what's right, though it is evolving quickly. A reply extends the idea, proposing to characterize "structural" conjectures via counterfactual propagation: assume the conjecture false and measure how much of the surrounding epistemic dependency graph must be revised.

Related event: AI Is Better at Falsifying Conjectures Than Proving Them(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →