AI能否自主做研究？听听这些专家怎么说
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR — 专家认为，AI正从辅助单项任务走向自主执行迭代式科研流程，但“得出答案”不等于完成研究。真正可靠的自主科研还取决于高质量数据、领域知识、长期推理与实验闭环，以及人类对问题价值、证据解释和研究规范的把关。
- 在材料和药物研发中，生成模型、仿真、自动化实验与现实反馈已被组合起来，用于提出候选方案并持续验证，展示了计算模型与物理实验闭环的潜力。
- 科研智能体仍难以应对开放式评估、长任务链、记忆不完整、工具调用不可靠等问题，也常难区分真实机制、表面相关性和单纯的数据拟合。
- 高质量实验数据是核心瓶颈；仿真可弥补湿实验数据不足，却不能取代机理理解和物理验证。数学领域也面临类似的验证与吸收瓶颈：AI产生结果的速度可能超过人类核验、解释和纳入知识体系的能力。
- 25位菲尔兹奖得主警告，以攻克重大难题和刷新基准为目标的“答案优先”路线，可能忽视概念理解、方法发展、知识传承及有价值新问题的提出，并带来抢先发布、归属与署名不清、未公开想法被使用等治理问题。
- 专家并非主张停止AI辅助研究，而是强调建立研究规范与监督机制；人类科学家仍需负责提出值得研究的问题、设计实验、解释证据、传授知识并监督重大决策。
注： 雷峰网侧重实验科学中的技术闭环与工程瓶颈，量子位侧重数学研究的目标错位、知识生态和治理风险。
Sources (2)
AI能否自主做研究？听听这些专家怎么说
TL;DR - Experts and companies at the 2026 Inclusion·Bund Conference examined how AI could evolve from assisting isolated research tasks to running iterative scientific-discovery workflows. They agreed that autonomous research depends on reliable data, domain knowledge, long-horizon agents, and closed loops between computational models and physical experiments.
- Materials and drug-development teams described systems combining generative models, simulations, automated experiments, and real-world feedback to propose and validate candidates.
- Scientific agents still struggle with open-ended evaluation, long task chains, incomplete memory, unreliable tool use, and distinguishing genuine mechanisms from correlations or good data fits.
- High-quality experimental data is a central constraint; simulations can supplement sparse wet-lab results but cannot replace mechanistic understanding and physical validation.
- Human scientists remain necessary for framing worthwhile questions, designing experiments, interpreting evidence, and overseeing consequential decisions.
陶哲轩邓煜究竟在反对什么：AI暴力解题摧毁人类数学精神
TL;DR - Twenty-five Fields Medalists, including Terence Tao and Yu Deng, warn that AI companies’ benchmark-driven pursuit of solving major mathematical problems is misaligned with mathematics’ goals of conceptual understanding, method development, and knowledge transmission. They seek urgent governance and research norms rather than a halt to AI-assisted mathematics.
- AI may now solve open problems faster than mathematicians can verify, explain, connect, and incorporate the results into established knowledge.
- The statement raises concerns about premature announcements, inadequate attribution, possible use of researchers’ prompts or unpublished ideas, and unclear authorship.
- Large-scale agentic search could exhaust valuable open problems without generating equally meaningful new questions or reusable insights.
- The signatories warn that answer-first systems may weaken mathematical training, open collaboration, and the human process that turns proofs into teachable knowledge.