Anthropic研究团队发布了一份关于人工智能辅助科学发现的报告,通过分析207,049个来自消费者计划的Claude及Claude Code科研会话样本,揭示了科学家在实际研究中如何使用这些AI工具。1 该数据分析的时间跨度为2026年8月2日至9月30日,结果显示AI在计算机科学和数学等形式科学领域的应用最为广泛,主要用于数据分析和写作,而在需要物理操作或高度创意构思的任务中使用较少。1
报告指出,数据分析是科研人员使用AI最频繁的场景,此类任务占据了Claude会话总量的33.7%,在Claude Code会话中更是高达48.1%,这一比例显著高于研究人员原本在此类任务上的时间投入占比(27.6%)。1 基于模型预测,在假设“完全自动化非物理步骤”的情境下,计算机科学和统计学领域的研究产出理论上可能实现多达15倍的增长。1 然而,受限于生物和化学等领域必须进行的物理实验瓶颈,这些学科的加速幅度相对较小。1
此外,调查数据显示约40%的受访科研人员预计,到2030年前AI将具备独立完成完整研究项目的能力。1 这种预期与不同学科对AI能力的评估密切相关,例如计算机科学与人文学科相比,其认知步骤在AI能力评分上更高,从而导致了两者在使用率上的显著差异。1 报告还引入了经济成本作为参考背景,假设AI代理的运行成本为每小时30美元,而人类研究人员的时薪约为95美元。1
Anthropic’s research team has released a comprehensive analysis of 207,049 scientific sessions conducted via Claude and Claude Code between August 2 and September 30, 2026, offering new insights into how artificial intelligence is reshaping academic workflows 1. The report indicates that the technology is currently most effective in formal sciences such as computer science and mathematics, where it serves primarily as a tool for data analysis and writing rather than physical experimentation or highly creative conceptualization 1.
A significant finding from the study highlights a disparity in adoption rates compared to traditional time allocation; data analysis tasks accounted for 33.7% of general Claude interactions and 48.1% of Claude Code sessions, substantially higher than the 27.6% share these tasks hold in researchers' overall time budgets 1. This trend is driven by superior AI capabilities in handling cognitive steps within fields like computer science, leading to notably higher usage rates in these disciplines compared to others 1.
Looking toward future impacts, the study projects that if non-physical steps in research were fully automated, productivity in computer science and statistics could theoretically increase up to fifteenfold 1. However, gains in biology and chemistry are expected to remain modest due to inherent bottlenecks associated with physical experiments 1. Despite these constraints, approximately 40% of surveyed scientists anticipate that AI will be capable of independently completing full research projects by 2030 1. The economic implications are also noted, with the report assuming an hourly cost of $30 for AI agents versus $95 for human researchers 1.
评论
还没有评论,欢迎留下第一条。