AI当"科研搭子"，你真的放心吗？

中国农业大学、北京大学等科研机构已将人工智能融入科研工作的多个环节，包括选题、实验规划和论文撰写，但这一应用趋势也暴露出诸多风险。1在选题阶段，虽然AI能批量生成候选研究问题——中国农业大学朱晨系统生成的79个候选研究问题中87%符合基本条件——但研究表明这些选题多停留在初级水平，无法提出具有原始创新性的问题。1在实验分析中，AI容易产生"看似合理"的假说却难以验证，苏黎世大学Yanagizawa-Drott团队评估了200余篇AI生成论文，发现大多数仅达到研究生水平。1

论文写作成为AI应用最突出的问题所在。1学术文献中虚假引文泛滥，2025年arXiv、bioRxiv、SSRN、PubMed Central四大平台存在近15万条虚假参考文献。1另一方面，北京大学干细胞研究中心的智能平台在实验执行层面展现出优势，独立完成120多个步骤，将实验误差从10%降至1%以下，但谷歌DeepMind预测的220万种新晶体结构中，虽有38万被判定为稳定，实际验证的比例不到0.2%。1

专家普遍认为，AI应严格定位为辅助工具。1华大集团首席研究员徐讯判断当前AI处于L3阶段，能完成既定的发现任务但不能自主提出原始问题。1北京大学副校长初晓波表示，应让AI处理数据广度，让人类守住思想深度与价值温度，关键决策权和原创性思维必须由人类保留。1

Artificial intelligence has become increasingly integrated into scientific research across multiple Chinese institutions, from topic selection and experimental design to manuscript writing, yet emerging evidence suggests significant limitations remain in its ability to advance original discovery.1 Research institutions including China Agricultural University and Peking University have embedded AI systems into key stages of the research pipeline, with mixed results that highlight both the technology's utility and its fundamental constraints.1

At China Agricultural University, AI systems generated 79 candidate research questions, with 87% meeting basic criteria, though researchers note that AI-generated topics typically remain at an elementary level and cannot formulate truly original problems.1 Similarly, when researchers at ETH Zurich examined over 200 papers generated by AI, the vast majority reflected only graduate-level research quality.1 In contrast, Peking University's stem cell research center deployed an intelligent platform that independently executed more than 120 experimental steps, reducing experimental error from 10% to below 1%, demonstrating AI's strengths in precision execution.1 Yet Google DeepMind's prediction of 2.2 million new crystal structures revealed a fundamental verification gap: while 380,000 were flagged as potentially stable, fewer than 0.2% have received experimental confirmation.1

A critical challenge plaguing academic publishing is the proliferation of fabricated citations. As of 2025, major platforms including arXiv, bioRxiv, SSRN, and PubMed Central contain approximately 150,000 false references, a consequence of AI systems generating plausible but fictitious sources.1 Experts and institutional leaders have converged on a consensus framework for AI deployment in research. Xu Xun, chief researcher at BGI Genomics, assessed that current AI has reached Level 3 capability—capable of making discoveries but unable to independently propose original research questions.1 Chu Xiaopo, vice president of Peking University, articulated this boundary explicitly: AI should handle the breadth of data processing while humans must preserve intellectual depth and values.1 The emerging view emphasizes that critical decision-making authority and originality in thinking must remain with human researchers, with AI functioning as a supporting tool rather than an autonomous agent.1