西北大学计算机科学教授及Ginni Rometty政策研究员Jessica Hullman发表文章,探讨在人工智能辅助科学研究中,“理解”与“可验证的输出”之间的张力 1。Hullman批评了冷战时期形成的“价值中立理想”,指出若仅追求正确结果而忽视人类的理解和问责机制,将削弱科学知识的根基 1。她认为,科学知识的生产依赖于社区内的批判性对话和责任归属,而当前AI缺乏持久身份和社会利害关系,无法完全替代人类作为认知主体的角色 1。
为论证上述观点,Hullman引用多位学者理论。她援引数学家Bill Thurston关于数学研究成功标准的观点,即衡量标准在于能否让人们更清晰地思考和理解数学 1。此外,她提及Terry Tao的思想实验:如果AI生成的癌症疗法有形式化证明且通过三期临床试验,但无人理解其原理,患者是否接受?Hullman指出,批评者认为这忽略了关键点 1。哲学家Helen Longino的观点也被引用,说明客观公共知识的基础是共享语言和对独立于感知的描述存在的信念;同行评审的功能不仅是核查数据,而是引入不同视角促使作者修正思考 1。依据哲学家Robert Brandom的理论,知识是一种社会行为,涉及承诺(commitment)和权利(entitlement)的归属;可靠性本身不足以构成知识,因为缺乏理由的支持 1。
针对AI技术的局限性,Hullman指出现代AI代理缺乏持久性和身份连续性,难以像人类科学家那样在社区中建立信誉并承担撤稿后果 1。Tom Dietterich曾指出arXiv上存在大量由看似不具备专业理解能力的人类提交的论文问题,这也反映了相关风险 1。综上所述,文章的核心论点在于:科学价值不仅在于正确性(输出),还在于理解、问责和社区批判性互动 1。
Jessica Hullman, a computer science professor at Northwestern University and Ginni Rometty Policy Fellow, has published an article arguing that the value of scientific knowledge extends beyond the mere production of correct outputs 1. She contends that the current trend in AI-assisted research, which prioritizes verifiable results over human understanding, threatens the foundations of science by neglecting essential mechanisms of accountability and critical dialogue within the scientific community 1. This perspective challenges the Cold War-era ideal of "value-neutral" science, suggesting that true knowledge requires more than just accurate data generation 1.
Hullman supports her argument with philosophical frameworks, citing Robert Brandom’s theory that knowledge is a social practice involving commitments and entitlements, noting that reliability alone does not constitute knowledge without supportive reasoning 1. She also references Helen Longino’s view that objective public knowledge relies on shared languages and peer review processes designed to introduce diverse perspectives rather than simply verifying data 1. Additionally, she highlights Bill Thurston’s assertion that the success of mathematical research is measured by its ability to clarify thinking, rather than just producing proofs 1.
The article further explores hypothetical scenarios to illustrate the limitations of AI in scientific contexts. It mentions Terry Tao’s thought experiment questioning whether patients would accept a cancer therapy if it were formally proven and clinically effective but remained completely unintelligible to humans; critics argue this scenario ignores the fundamental role of understanding in trust 1. Hullman points out that modern AI agents lack persistent identities and social stakes, preventing them from building reputation or facing consequences such as retractions, unlike human scientists 1. She notes that issues regarding professional comprehension are already present in human submissions to platforms like arXiv, where Tom Dietterich has observed papers submitted by individuals who appear to lack professional understanding of their own work 1.
评论
还没有评论,欢迎留下第一条。