What Scientists Actually Use AI For: Data and Drafts, Rarely Ideas
科学家用 AI 干什么：分析数据和写作，很少用来提想法
Across 207,000 research sessions, Claude was used for analysis and writing far more than their share of a scientist's day — and for hypothesis generation far less.
Anthropic has published a study of how researchers actually use its models, based on 207,000 research sessions between 2 August and 30 September this year — 129,000 chats and 78,000 Claude Code sessions. The company kept only the tags attached to each session: discipline, sub-field, task type. It says it did not extract or store research data, hypotheses, methods or unpublished papers.
The first finding is that adoption follows the shape of the work. Computer science, mathematics and statistics use the models most; health and the physical sciences that need a laboratory use them least. Computer science accounted for 11.9% of research sessions, engineering 8.4% and medicine 7.7% — and relative to the number of researchers in each field, computer science, maths and statistics are clearly over-represented.
Claude Code sharpens the picture. Computer scientists are 4.3% of the world's researchers but account for 32.7% of Claude Code research sessions. AI and machine learning is more extreme still: 1.2% of researchers, 5% of Claude sessions, 15.2% of Claude Code sessions.
The task breakdown is the interesting part, because it can be set against how scientists say they spend their time. Data analysis takes 27.6% of a researcher's working hours and 33.7% of Claude research sessions — 48.1% of Claude Code sessions. Writing and communication takes 16% of their time and 32% of Claude sessions.
The tasks that are under-used are just as telling. Data collection is 28.6% of researchers' time but only 11.5% of sessions. Generating ideas and hypotheses is 7.1% of their time and 4.2% of sessions — 2.1% for Claude Code. The pattern is consistent: the model is used where output can be checked quickly and where the work is text or code, and much less where judgement or physical apparatus is involved.
A survey of 528 scientists — mostly in life and health sciences, and all already active AI users — matched the logs closely. Anthropic is careful about the limits: these are its own users rather than a random sample, and a tag describes what a session was about, not whether it helped.
中文翻译（全文）
Anthropic 发表了一项关于研究者实际如何使用其模型的研究，样本是今年 8 月 2 日至 9 月 30 日间的 207,000 次科研会话——其中 129,000 次对话与 78,000 次 Claude Code 会话。公司只保留了每次会话附带的标签：学科、子领域、任务类型。它表示未提取或保存研究数据、假设、方法或未发表论文。
第一个发现是：使用情况跟着工作的形状走。计算机科学、数学与统计学用得最多；需要实验室的医疗健康与物理科学用得最少。计算机科学占科研会话的 11.9%，工程学 8.4%，医学 7.7%——而相对于各领域的研究人员数量，计算机科学、数学与统计学明显被过度代表。
Claude Code 让这个图景更极端。计算机科学家占全球研究人员的 4.3%，却贡献了 Claude Code 科研会话的 32.7%。AI 与机器学习更甚：占研究人员 1.2%，占 Claude 会话 5%，占 Claude Code 会话 15.2%。
真正有意思的是任务分布，因为它可以和科学家自述的时间分配对照。数据分析占研究人员工作时间的 27.6%，占 Claude 科研会话的 33.7%——在 Claude Code 中高达 48.1%。写作与沟通占他们 16% 的时间，却占 Claude 会话的 32%。
用得太少的那些任务同样说明问题。数据收集占研究人员 28.6% 的时间，却只占会话的 11.5%；生成想法与假设占 7.1% 的时间，只占会话的 4.2%——Claude Code 中是 2.1%。规律很一致：模型被用在能快速核验产出、且工作本身是文字或代码的地方，而在涉及判断或实体设备时用得少得多。
对 528 位科学家的问卷调查——多为生命与健康科学领域，且都是已在使用 AI 的人——与日志高度吻合。Anthropic 也谨慎说明了局限：这些是它自己的用户，不是随机样本；而一个标签说明的是某次会话在做什么，而不是它有没有帮上忙。
重点词汇
- discipline/ˈdɪsəplɪn/n.— a branch of academic study学科Usage follows the shape of each discipline.
- sub-field/ˌsʌb ˈfiːld/n.— a specialised area within a discipline子领域Sessions were tagged by sub-field.
- over-represented/ˌəʊvə ˌreprɪˈzentɪd/adj.— appearing more often than its share would suggest被过度代表的Computer science is over-represented.
- breakdown/ˈbreɪkdaʊn/n.— an analysis of something divided into parts分布；拆解The task breakdown is the interesting part.
- hypothesis/haɪˈpɒθəsɪs/n.— a proposed explanation to be tested假设Hypothesis generation was rarely assisted.
- judgement/ˈdʒʌdʒmənt/n.— the ability to decide well判断It is used less where judgement is involved.
- log/lɒɡ/n.— an automatic record of activity日志记录The survey matched the logs.
- sample/ˈsɑːmpl/n.— a group selected to represent a larger one样本These are its own users, not a random sample.
- consistent/kənˈsɪstənt/adj.— staying the same across cases一致的The pattern is consistent.
- account for/əˈkaʊnt fɔː/phrasal v.— to make up a particular share占（比例）They account for 32.7% of sessions.
阅读测验
共 5 道题目
What was the study based on?
Which fields use the models most?
Data analysis takes a larger share of Claude Code sessions than of researchers' working hours.
Which tasks are under-used relative to time spent? (Select all that apply)
Anthropic notes that these users are not a random _____ of researchers.
评论
0 条讨论
登录后发表评论
立即登录暂无评论
成为第一个分享想法的人吧！