According to Ethan Mollick on X, a recurring AI pattern emerges: initial overstated claims, followed by minor research assists, and later verified breakthroughs; he cites Przemek Chojecki’s post claiming GPT-5.4 Pro helped solve multiple Erdős problems within 24 hours (source: Ethan Mollick on X; original claim by Przemek Chojecki on X). According to Mollick, last year’s flubbed Erdős problem claims illustrate the risk of premature announcements, while recent AI-aided discovery represents incremental but real value (source: Ethan Mollick on X). For AI leaders, the business takeaway is to require formal verification, peer review, and reproducible proofs before marketing frontier-model math wins, and to focus near term on validated use cases such as theorem search, lemma generation, and proof checking pipelines where commercial AI stacks can win in academic and enterprise R&D (source: Ethan Mollick on X; industry practice). As reported by Mollick, this hype-to-proof progression affects capability communication, suggesting vendors should publish benchmarks, third-party audits, and artifacts (code, proof scripts) to convert attention into enterprise trust in 2026 (source: Ethan Mollick on X).