JohnXie的个人博客分享 http://blog.sciencenet.cn/u/JohnXie

博文

澳大利亚统计学会 关于使用生成式人工智能的声明

已有 106 次阅读 2026-7-15 15:09 |个人分类:海外生活|系统分类:海外观察

Statistical Society of Australia (SSA) = 澳大利亚统计学会在今年3月份发表了一份关于在统计学与数据科学中使用生成式人工智能的声明,我把其声明的英文原文及中文翻译都放在这个帖子里供大家参考。

另外,澳大利亚统计学会最近还组织了一场关于数据质量及可靠性(Data that we can trust)的专家讨论会。其中有专家认为目前的人工智能工具在为我们进行分析或回答我们的问题时所依据的数据或语料或信息源的质量及可靠性是一个值得我们关注的问题。该专家认为只有用户能够看得到的数据才是可信的(data that we can see is the data we can trust)。而目前人工智能工具所依据的信息源/语料的质量与可靠性都是不透明或无法被我们这些终端用户所确认的。人工智能工具的开发商也没有或者无法向用户保证信息源/语料的质量与可靠性。因此对于任何由AI为我们提供的答案都要保持必要的谨慎的态度,必须经过某种人类智能所能够确认的检验才应该被采信或应用。

澳大利亚统计学会关于在统计学与数据科学中使用生成式人工智能的声明 (ChatGPT 5.5 基础付费版中文译文)

生成式人工智能(Generative Artificial Intelligence,简称AI)正迅速发展,为各学科研究人员在数据分析的各个阶段提供了新的机遇。与任何技术一样,生成式 AI 必须以负责任的方式使用。下文概述了在利用生成式 AI 工具辅助统计工作时应遵循的重要原则。这些指导原则应结合各机构针对具体科研活动及其他应用所制定的相关政策一并理解和应用。

保存对话记录与代码,确保研究可重复性

可重复性(Reproducibility)是科学研究的核心要求之一。尽管可以将数据上传至 AI 平台,并要求其完成分析或返回分析结果,但其他研究人员按照相同步骤重复操作时,未必能够得到相同的结果。因此,AI 工具生成的输出仅应视为思路启发(brainstorms)或初稿(drafts)。研究人员必须获取、保存并报告能够重复执行分析的代码,以确保分析过程和结果可以被验证和复现。此外,在条件允许的情况下,分析过程中使用 AI 工具时的对话记录(chat)、提示词(prompts)以及所使用的AI 模型信息,也应一并保存,作为科研记录(scientificrecord)的组成部分。

核实分析结果的正确性

AI 可能在数据分析的不同阶段引入错误。研究负责人有责任在力所能及的范围内,理解并核查 AI 所完成分析步骤的正确性。例如:若使用 AI 进行数据整理(data manipulation),应检查最终生成的数据集是否存在错误;若使用 AI 生成分析代码,应确认代码能够正确运行,并且确实实现了预期的分析目的。

重视知识产权与保密要求

使用 AI 工具意味着需要共享一定的信息。如果需要将数据或分析流程中的任何输出上传至 AI 平台,应首先确认此举是否符合:数据共享(data sharing)相关规定;保密(confidentiality)要求;知识产权(Intellectual Property,IP)政策;以及项目合作伙伴所享有的知识产权权益。

在咨询统计学专家时

无论是否采用 AI 辅助分析,在研究设计、数据分析及研究结果解释过程中,向具有专业资质的统计学家寻求建议始终是良好的科研实践。这一点不会因 AI 的应用而改变。

为了更有效地开展统计咨询,建议研究人员:

  • 在咨询前明确界定研究问题;

  • 若在咨询前已经使用过 AI,应携带与 AI 的对话记录(chat)以及相关生成成果(如代码等)参加咨询;

  • 与统计学家讨论数据的背景和研究设计,以便在 AI 对话和提示词中正确使用统计学术语;

  • 明确适当的统计分析方法及其准确表述,从而有助于在 AI 对话中正确描述分析需求。

统计学家通常无法仅依据 AI 的分析输出判断其结果是否正确,而往往需要独立开展分析加以验证。因此,独立于 AI 工具开发分析脚本,仍然是开展严谨统计分析的重要组成部分。然而,如果能够提供 AI 的对话记录及其生成结果,将有助于统计学家了解分析工作的整体背景,并可能提出如何改进提示词(prompts),从而获得更高质量 AI 输出的建议。

其它的相关信息链接:

1.美国统计学会(American Statistical Association,ASA)

ASA Statement on Ethical AI – Principles for Statistical Practitioners

https://www.amstat.org/docs/default-source/amstat-documents/asa-statement-on-ethical-ai-principles-for-statistical-practitioners.pdf 

2.美国政府问责局(U.S. Government Accountability Office,GAO)

Artificial Intelligence: An Accountability Framework for Federal Agencies andOther Entities

https://www.gao.gov/products/gao-21-519sp 

3.澳大利亚政府(Australian Government)

Interim guidance for Australian Government agencies: Introducing generative AI

https://www.digital.gov.au/policy/artificial-intelligence/interim-guidance-introducing-generative-ai

致谢:SSA 衷心感谢**统计咨询网络(Statistical Consulting Network)**对本声明制定工作所作出的贡献。

------------------------------------------------------------------------------------------------------------------------------------

(以下为声明的英文原文:)

SSA statement on the use of generative AI in statistics and data science

Generative Artificial Intelligence (AI) is advancing rapidly and provides opportunities for investigators across many disciplines in all stages of data analysis. Like any technology, it needs to be used responsibly and below, we outline some important principles to consider when using generative AI tools to aid statistical work. These guidelines should be considered in the context of institutional policies governing specific research and other applications.

Store chats and code for reproducibility

Reproducibility is a core requirement in scientific investigations. While it is possible to upload data to an AI platform and ask it to return results of an analysis, another person repeating this process might get different results. Output from AI tools should only be considered as “brainstorms” or “drafts”. It is essential that investigators obtain and report reproducible code to conduct analyses. Similarly, where possible, “chats” and prompts used with AI tools in the analysis process should also be stored (together with details on which AI model used) as part of your scientific record.

Check correctness

AI can introduce errors at different stages of analysis. The onus is on the responsible investigator to understand and check the correctness of AI-led steps in analysis, as far as possible. For example, if using AI for data manipulation, the final dataset should be checked for errors; if using AI to generate code, that code should be checked that it runs correctly and does what it was intended to do.

Consider intellectual property and confidentiality

The use of AI involves sharing information. If uploading data or any outputs from the analysis workflow to AI, consider if this is consistent with the data sharing, confidentiality and IP policies governing the data, and the IP rights of any collaborators on the project.

When consulting with a statistician

It is always good practice to seek advice from a qualified statistician when planning a research study or analysing and interpreting findings, and this remains true for AI-assisted analysis.

To benefit effectively from consultation with a statistician:

• An investigator should have a clearly defined research question;

• If using AI prior to seeing a statistician, bring the chat as well as any relevant products (code, etc.) to the consultation;

• Discuss the context and design of data for analysis to support appropriate use of statistical terminology in any AI chat or prompts;

• Clarify the appropriate method of analysis and how it is accurately described to support correct use in AI chats.

It may not be possible for a statistician to verify that an analytic output from AI is correct without carrying out an independent analysis. Independent development of analysis scripts separate from AI tools is an essential element of rigorous analysis. However, access to both the chat and output can help the statistician understand the broader context around the product and may result in advice on how better to prompt AI.

Useful resources:

The Americal Statistical Association Statement on Ethical AI

The US Government Accountability Office report on AI

Australian Government advice to staff on use of public generative AI

SSA acknowledges the Statistical Consulting Network for the development of this statement.



https://blog.sciencenet.cn/blog-3503579-1543734.html

上一篇:我与澳大利亚统计学会的故事




    
收藏 IP: 49.187.150.*| 热度|

5 崔锦华 宁利中 杨正瓴 郑永军 蒋大和

该博文允许注册用户评论 请点击登录 评论 (1 个评论)

数据加载中...

Archiver|手机版|科学网 ( 京ICP备07017567号-12 )

GMT+8, 2026-7-16 16:19

Powered by ScienceNet.cn

Copyright © 2007- 中国科学报社

返回顶部