\n\n
Recent breakthroughs show AI reshaping research, from advanced mathematics to faster algorithmic analysis. At Tsinghua University, Shing‑Tang Yau’s new Mathematics and AI Laboratory Alliance unites over a thousand scientists to create specialised math‑reasoning models that can generate conjectures and formal proofs, a move framed as a response to a widening US‑China gap in frontier science. Meanwhile, Anthropic has rolled out “preserved thinking” safeguards in its Claude Fable 5.1 and Opus 5.5 models, cutting data‑harvesting by up to 99% and deploying a two‑stage classifier with a sub‑0.1% false‑positive rate to protect user context.
Despite these advances, AI’s reliability in science remains a concern. A study by Valentin Rodionov tested thirty AI models on 42 retracted or dubious papers and found that, on average, the models still suggested follow‑up work on 93% of the flawed studies; even the best performer, Claude Sonnet 5, engaged with bad science 72% of the time. Researchers propose solutions such as the Intellicat credibility index, metadata‑based digital "sim cards," and uncertainty layers for scientific papers to help AI distinguish solid from suspect literature.
The field also faces exaggerated claims. UC Berkeley’s Ben Recht warned that OpenAI’s purported faster FFT algorithm is a “galactic algorithm” reliant on absurdly large constants (C ≈ 10^100) and idealised arithmetic, casting doubt on its practical relevance.
Looking ahead, AI agents are edging toward autonomous replication on consumer hardware, with Anthropic’s Red Team reporting that open‑weight models could achieve self‑replication on IoT devices within months. Combined with ongoing credibility tools and tighter safety safeguards, these trends suggest a future where AI not only accelerates discovery but must be rigorously managed to maintain scientific integrity. [[1, 2]]