AI治理之他山之石 | AI Now：大型AI模型正巩固大型科技企业的力量

编者按

「AI治理之他山之石」系列旨在通过快速的AI机器翻译，将海外最前沿的关于AI治理的思考和动态引入国内讨论。

此篇是我们系列的第四篇，其由AI Now 研究院发布，作为其2023年度科技治理全景报告的一个章节。在此章节中，AI Now针对大型AI模型的治理尖锐地指出大型科技企业正在无序地通过宣称大模型的基础性，掠夺式垄断和汲取数据价值，这是AI治理应当立刻关注并予以限制的。

行业正试图避开监管，但大规模AI需要更多而不是更少的审查。

大规模通用AI模型（如GPT-4及其面向用户的应用程序聊天GPT）被工业界宣传为“基础”和该领域科学进步的主要转折点。它们也经常与“开源”的模糊定义联系在一起。

这些说法分散了我们所说的“规模病态”的注意力，这种病态每天都变得更加根深蒂固：大规模AI模型仍然主要由大型科技公司控制，因为它们需要巨大的计算和数据资源，并且还提出了关于歧视、隐私和安全漏洞以及负面环境影响的有据可查的担忧。

在过去的一年里，像大型语言模型（LLM）这样的大规模AI模型受到了最多的炒作和恐惧。围绕这些系统的兴奋和焦虑1都强化了这样一种观念，即这些模型是“基础”，是该领域进步的一个主要转折点，尽管这些系统未能对提示做出有意义的反应。2但是与这些系统相关的叙述分散了我们所说的“规模病态”，这种新兴框架既突出又分散了我们的注意力。例如，斯坦福大学在2022年初宣布一个同名的新中心时引入了“基础”一词3，此前发表了一篇文章，列出了与LLM相关的许多生存危害。4值得注意的是，这些模型作为“基础”的引入旨在将它们（以及支持它们的人）等同于毫无疑问的科学进步，这是通往“通用人工智能”道路上的垫脚石5（另一个模糊的术语，唤起科幻小说中取代或取代人类智能的概念），从而使它们的大规模采用成为必然。6随着开放AI的推出，这些论述又回到了前台s最新的基于LLM的聊天机器人chatGPT。

另一方面，“通用AI”（GPAI）一词正在欧盟的AI法案等政策工具中使用，以强调这些模型没有明确的下游用途，可以微调以适用于特定情况。7它被用来提出论点，例如，由于这些系统缺乏明确的意图或明确的目标，它们应该受到不同的监管或根本不监管——实际上在法律中造成了一个重大漏洞（更多信息见下文第2节）8。

这些术语故意掩盖了这些模型的另一个基本特征：它们目前需要的计算和数据资源规模最终只有资源最充足的公司才能承受。9.从数字的角度来看，一些估计表明，运行聊天GPT 每月 将花费300万美元。10和2000万美元的计算成本来训练路径语言模型（PaLM），这是谷歌最近的LLM11目前只有少数资源极其丰富的公司能够构建它们。这就是为什么大多数现有的大规模AI模型几乎完全由大科技公司开发，尤其是谷歌（谷歌大脑、深度思维）、元和微软（及其投资对象OpenAI）。这包括许多现成的、预训练的AI模型，这些模型作为云AI服务的一部分提供，这个市场已经集中在大科技公司，如AWS（亚马逊）、谷歌云（字母表）和Azure（微软）。即使随着这些系统的大规模部署，成本降低或下降（这是一个备受争议的说法12），大型科技公司也可能保持先发优势，因为它们拥有磨练底层语言模型和开发宝贵内部专业知识所需的时间和市场经验。因此，较小的企业或初创企业可能难以成功进入这一领域，让LLM的巨大处理能力集中在少数大型科技公司手中。13

这种市场现实贯穿了越来越多的叙述，这些叙述强调了“开源”和“社区或中小型企业（SME）驱动”GPAI项目的潜力，甚至将GPAI与开源（正如我们在围绕欧盟的 AI 法案的讨论中看到的那样）混为一谈。14例如，2022年9月，由软件联盟（或BSA）领导的十个行业协会发表了一份声明，反对将GPAI模型的开发者纳入任何法律责任。15他们的头条论点是，这将“严重影响欧洲的开源开发”，并“破坏AI吸收、创新、和数字化转型。“16该声明依赖于假设的例子，这些例子讽刺了GPAI模型是如何工作的，以及监管干预将需要什么——引用的经典案例是一个个人开发人员创建了一个开源文档阅读工具，并被监管要求背负在它既无法预测也无法控制的未来用例上。

这里的讨论是将具有与权限和许可制度相关的特定含义的“开源”与“开放”的直观概念混为一谈，因为它们可供下游使用和适应（通常通过应用程序编程接口或API）。后者更类似于“开放访问”，尽管从这个意义上说，它们仍然有限，因为它们只共享API，而不是模型或训练数据源。17事实上，在OpenAI宣布其GPT-4模型的论文中，该公司表示不会提供用于开发GPT-4的架构、模型大小、硬件、训练计算、数据构造或训练方法的细节，只是指出它使用了从人类反馈中强化学习的方法，主张竞争和安全问题。与当前增加公司留档流程的努力背道而驰的是，18这样的举措加剧了基于机器学习的科学中已经被描述为可重复性危机的问题，其中关于基于AI的模型的能力的主张无法被其他人验证或复制。19

最终，这种形式的部署只会增加大型科技公司的收入，巩固其战略业务优势。20虽然有合理的理由考虑与使此类系统广泛可访问相关的潜在下游危害21，即使项目可能公开其代码并满足开源的其他定义，这些系统的巨大计算需求意味着这些项目和商业市场之间的依赖关系可能会持续存在。22

大规模AI模型必须接受紧急监管审查，特别是考虑到向公众推出的疯狂速度。在模型开发阶段记录和审查数据和相关设计选择是浮出水面和减轻伤害的关键。

这不是一张白纸。必须扩大和加强关于算法问责制的立法提案，并创造性地应用现有的法律工具来引入摩擦并塑造创新方向。

生成性AI模式的例外论越来越多，这种模式低估了固有风险，并证明将其排除在AI监管范围之外是合理的。我们应该从欧洲正在进行的关于将通用AI纳入即将出台的AI法案的“高风险”类别的辩论中吸取教训。

随着对AI未来潜力的大肆宣传，聊天GPT的发布（以及随后对微软搜索聊天机器人的改编）立即出现了棘手的法律问题，例如，谁拥有这些系统生成的内容并对其拥有权利？25生成AI是否受到保护，免受与它们可能在第230条等中介责任保护下生成的非法内容有关的诉讼？26

显而易见的是，已经有适用于大型语言模型的现有法律制度，我们不是从头开始构建它们。事实上，暗示这是必要的言论主要符合行业的最大利益，因为它减缓了执法和法律更新的速度。

该FTC最近发表的一篇博客文章概述了该机构当局已经适用于生成AI系统的几种方式：如果它们被用于欺诈、造成重大伤害或对系统能力提出虚假声明，FTC有理由介入。还有许多其他法律制度可能适用的领域：知识产权法、反歧视条款和网络安全法规就是其中之一。

还有一个前瞻性的问题，即哪些规范和责任应该适用于这些系统。围绕AI系统的公认危害（特别是不准确、偏见和歧视）越来越多的共识导致了过去几年围绕数据和算法设计实践的更大透明度和勤奋的一系列政策运动。这些新兴的AI政策方法需要得到加强，以应对这些模型带来的特殊挑战，目前公众对AI的关注正准备在缺乏势头的地方激发势头。

在欧盟，这个问题不是理论问题。这是一场激烈争论的核心，争论的焦点是所谓的“通用AI”（GPAI）模型的原始开发者是否应该受到即将出台的AI法案的监管要求。27欧盟委员会于2021年4月提出，委员会的原始提案（第52a条）有效地免除了GPAI的开发者遵守法律中的留档范围和其他问责要求。28因此，这意味着表面上没有预定用途或背景的GPAI将不符合“高风险”的条件——另一项条款（第28条）证实了这一立场，暗示GPAI的开发者只会承担责任如果他们大幅修改或调整AI系统以用于高风险用途，则符合要求。欧洲理事会的立场有所不同，即GPAI的原始提供者将受到法律中某些要求的约束，尽管正在制定将这些要求委托给委员会的具体细节。最近的报告表明，欧洲议会也在考虑原始GPAI提供者的具体义务。

随着欧盟机构间谈判在这个问题上的转变，辩论似乎已经演变成一种无益的二元方案，最终用户或原始开发者都承担责任29，而不是在不同阶段都承担不同类型的责任。30据报道，最近泄露的一份非官方US政府立场文件指出，给GPAI的原始开发者施加负担可能“非常繁重，技术上困难，在某些情况下是不可能的”31

这些说法忽略了大规模AI模型需要监督的两个最重要原因：

在开发阶段做出的数据和设计决策决定了模型的许多最有害的下游影响，包括偏见和歧视的风险。32越来越多的研究和宣传主张严格留档和问责制要求对大型模型开发人员的好处。33

这些模型的开发者，其中许多是大技术公司或由大技术公司资助的，通过与下游行为者的许可交易从这些模型中获得商业利益。34为特定用途许可这些模型的公司当然应该对在应用这些模型的特定背景下进行尽职调查负责，但是让它们对原始开发阶段的数据和设计选择产生的风险承担全部责任，将导致不公平和无效的监管结果。

参考阅读：

1. Future of Life Institute. “Pause Giant AI Experiments: An Open Letter.” Accessed March 29, 2023; Harari, Yuval, Tristan Harris, and Aza Raskin. “Opinion | You Can Have the Blue Pill or the Red Pill, and We’re Out of Blue Pills.” The New York Times, March 24, 2023, sec. Opinion. ↩

2. See Greg Noone, “‘Foundation models’ may be the future of AI. They’re also deeply flawed,” Tech Monitor, November 11, 2021 (updated February 9, 2023); Dan McQuillan, “We Come to Bury ChatGPT, Not to Praise It,” danmcquillan.org, February 6, 2023; Ido Vock, “ChatGPT Proves That AI Still Has a Racism Problem,” New Statesman, December 9, 2022; Janice Gassam Asare, “The Dark Side of ChatGPT,” Forbes, January 28, 2023; and Billy Perrigo, “Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic,” Time, January 18, 2023. ↩

3. See the Center for Research on Foundation Models, Stanford University; and Margaret Mitchell (@mmitchell_ai), “Reminder to everyone starting to publish in ML: ‘Foundation models’ is *not* a recognized ML term; was coined by Stanford alongside announcing their center named for it; continues to be pushed by Sford as the term for what we’ve all generally (reasonably) called ‘base models’,” Twitter, June 8, 2022, 4:01 p.m. ↩

4. Emily Bender, Timnit Gebru, Angelina McMillan-Major, Shmargaret Shmitchell, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT ’21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, March 2021. ↩

5. See Sam Altman, “Planning for AGI and beyond”, March 2023. ↩

6. See National Artificial Intelligence Research Resource Task Force, “Strengthening and Democratizing the U.S. Artificial Intelligence Innovation Ecosystem: An Implementation Plan for a National Artificial Intelligence Research Resource,” January 2023; and Special Competitive Studies Project, “Mid-Decade Challenges to National Competitiveness,” September 2022. ↩

7. The EU Council’s draft or “general position” on the AI Act text defines General Purpose AI (GPAI) as an AI system that “that – irrespective of how it is placed on the market or put into service, including as open source software – is intended by the provider to perform generally applicable functions such as image and speech recognition, audio and video generation, pattern detection, question answering, translation and others; a General Purpose AI system may be used in a plurality of contexts and be integrated in a plurality of other AI systems.” See Council of the European Union, “Proposal for a Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Acts – General Approach,” November 25, 2022; see also Future of Life Institute and University College London’s proposal to define GPAI as an AI system “that can accomplish or be adapted to accomplish a range of distinct tasks, including some for which it was not intentionally and specifically trained.” Carlos I. Gutierrez, Anthony Aguirre, Risto Uuk, Claire C. Boine, and Matija Franklin, “A Proposal for a Definition of General Purpose Artificial Intelligence Systems,” Future of Life Institute, November 2022. ↩

8. Alex C. Engler, “To Regulate General Purpose AI, Make the Model Move,” Tech Policy Press, November 10, 2022. ↩

9. See Ben Cottier, “Trends in the dollar training cost of machine learning systems”, Epoch, January 31, 2023; Jeffrey Dastin and Stephen Nellis, “For tech giants, AI like Bing and Bard poses billion-dollar search problem”, Reuters, February 22, 2023; Jonathan Vanian and Kif Leswing, “ChatGPT and generative AI are booming, but the costs can be extraordinary”, CNBC, March 13, 2023; Dan Gallagher, “Microsoft and Google Will Both Have to Bear AI’s Costs”, WSJ, January 18, 2023; Christopher Mims, “The AI Boom That Could Make Google and Microsoft Even More Powerful,” Wall Street Journal, February 11, 2023; and Diane Coyle, “Preempting a Generative AI Monopoly,” Project Syndicate, February 2, 2023. ↩

10. See Tom Goldstein (@tomgoldsteincs), “I estimate the cost of running chatGPT is $100K per day, or $3M per month. This is a back-of-the-envelope calculation. I assume nodes are always in use with a batch size of 1. In reality they probably batch during high volume, but have GPUs sitting fallow during low volume,” Twitter, December 6, 2022, 1:34 p.m; and MetaNews, “Does ChatGPT Really Cost $3M a Day to Run?” December 21, 2022 ↩

11. Lennart Heim, “Estimating ��PaLM’s training cost,” April 5, 2022; Peter J. Denning and Ted G. Lewis, “Exponential Laws of Computing Growth,” Communications of the ACM 60, no. 1 (January 2017):54–65. ↩

12. Andrew Lohn and Micah Musser, “AI and Compute”, Center for Security and Emerging Technology ↩

13. Richard Waters, “Falling costs of AI may leave its power in hands of a small group”, Financial Times, March 9, 2023. ↩

14. Ryan Morrison, “EU AI Act Should ‘Exclude General Purpose Artificial Intelligence’ – Industry Groups,” Tech Monitor, September 27, 2022. ↩

15. See BSA | The Software Alliance, “BSA Leads Joint Industry Statement on the EU Artificial Intelligence Act and High-Risk Obligations for General Purpose AI,” press release, September 27, 2022, ; and BSA, “Joint Industry Statement on the EU Artificial Intelligence Act and High-Risk Obligations for General Purpose AI,” September 27, 2022. ↩

16. BSA, “BSA Leads Joint Industry Statement on the EU Artificial Intelligence Act and High-Risk Obligations for General Purpose AI.” ↩

17. Peter Suber, Open Access (Cambridge, MA: MIT Press, 2019). ↩

18. Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, Timnit Gebru, “Model Cards for Model Reporting,” arXiv, January 14, 2019; Emily Bender and Batya Friedman, “Data Statements for Natural Language Model Processing: Toward Mitigating System Bias and Enabling Better Science”, Transactions of the Association for Computational Linguistics, 6 (2018): 587-604; Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé Iii, and Kate Crawford, “Datasheets for Datasets.” Communications of the ACM 64, no. 12 (2021): 86-92. ↩

19. Sayash Kapoor and Arvind Narayanan. “Leakage and the Reproducibility Crisis in ML-based Science.” arXiv, July 14, 2022. ↩

20. A report by the UK’s Competition & Markets Authority (CMA) points to how Google’s “open” approach with its Android OS and Play Store (in contrast to Apple’s) proved to be a strategic advantage that eventually led to similar outcomes in terms of revenues and strengthening its consolidation over various parts of the mobile phone ecosystem. See Competition & Markets Authority, “Mobile Ecosystems: Market Study Final Report,” June 10, 2022. ↩

21. Arvind Narayanan and Sayash Kapoor, “The LLaMA is Out of the Bag. Should We Expect a Tidal Wave of DIsinformation?” Knight First Amendment Institute (blog), March 6, 2023. ↩

22. See Coyle, “Preempting a Generative AI Monopoly.” ↩

23. Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? ��.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–23. FAccT ’21. New York, NY, USA: Association for Computing Machinery, 2021. ↩

24. Metz, Cade, and Daisuke Wakabayashi. “Google Researcher Says She Was Fired Over Paper Highlighting Bias in A.I.” The New York Times, December 3, 2020, sec. Technology. ↩

25. James Vincent, “The scary truth about AI copyright is that nobody knows what will happen next”, The Verge, November 15, 2022. ↩

26. Electronic Frontier Foundation, “Section 230”, Electronic Frontier Foundation, n.d. ↩

27. Creative Commons, “As European Council Adopts AI Act Position, Questions Remain on GPAI”, Creative Commons, December 13, 2022; Corporate Europe Observatory, “The Lobbying Ghost in the Machine: Big Tech’s covert defanging of Europe’s AI Act”, February 2023; Gian Volpicelli, ‘ChatGPT broke the EU plan to regulate AI’, Politico, March 3. ↩

28. European Commission, “Proposal for a Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Acts,” April 21, 2021. ↩

29. An article by Brookings Fellow Alex Engler, for example, argues that regulating downstream end users makes more sense because “good algorithmic design for a GPAI model doesn’t guarantee safety and fairness in its many potential uses, and it cannot address whether any particular downstream application should be developed in the first place.” See Alex Engler, “To Regulate General Purpose AI, Make the Model Move”, Tech Policy Press, November 10, 2022; See also Alex Engler, “The EU’s attempt to regulate general purpose AI is counterproductive”, Brookings, August 24, 2022. ↩

30. The Mozilla Foundation’s position paper on GPAI helpfully argues in favor of joint liability. See Maximilian Gahntz and Claire Pershan, “Artificial Intelligence Act: How the EU Can Take on the Challenge Posed by General-Purpose AI Systems,” Mozilla Foundation, 2022. ↩

31. Luca Bertuzzi, “The US Unofficial Position on Upcoming EU Artificial Intelligence Rules,” Euractiv, October 24, 2022. ↩

32. Sasha Costanza-Chock, Design Justice: Community-Led Practices to Build the Worlds We Need. Cambridge: MIT Press. ↩

33. See Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford, “Datasheets for Datasets,” arXiv:1803.09010, December 2021, ; Mehtab Khan and Alex Hanna, “The Subjects and Stages of AI Dataset Development: A Framework for Dataset Accountability,” Ohio State Technology Law Journal, forthcoming, accessed March 3, 2023, ; and Bender, Gebru, McMillan-Major, and Shmitchell, “On the Dangers of Stochastic Parrots.” ↩

34. See for example Madhumita Murgia, “Big Tech companies use cloud computing arms to pursue alliances with AI groups”, Financial Times, February 5, 2023; Leah Nylen and Dina Bass, “Microsoft Threatens Data Restrictions In Rival AI Search”, Bloomberg, March 25, 2023; OpenAI, Pricing; Jonathan Vanian, “Microsoft adds OpenAI technology to Word and Excel”, CNBC, March 16, 2023; and Patrick Seitz, “Microsoft Stock Breaks Out After Software Giant Adds AI To Office Apps”, Investor’s Business Daily, March 17, 2023. ↩

原标题：《AI治理之他山之石 | AI Now：大型AI模型正巩固大型科技企业的力量》

阅读原文