剑桥大学人文与社会科学学院艾拉·麦克弗森博士与马太·坎代亚教授写到，撇开一切关于生成式人工智能的炒作，该文章（宣言或原则书）为学者及学生提供了一个框架，用以厘清生成式人工智能是否（而不是如何）真正有益于他们的研究。
This manifesto and principles cut through the hype around generative AI to provide a framework that supports scholars and students in figuring out if, rather than how, generative AI contributes to their scholarship, writes Dr Ella McPherson and Prof Matei Candea, School of the Humanities and Social Sciences, University of Cambridge.

这份研究警示我们目前岌岌可危的正是我们的教育价值观。

This approach reminds us that what is at stake is nothing less than our educational values, they argue.

生成式人工智能（AI）在疫情期间席卷了高等教育界，那时新冠带来的生活及工作上的不幸和需求还在继续，也面临着巨大的工作量。它的出现并没有受到来自暴利行业的引导或资源的影响。

Generative artificial intelligence (AI) has stormed higher education at a time when we are all still recovering from the tragedies and demands of living and working in a pandemic, as well as facing significant workload pressures. It has landed without any significant guidance or resources from a rampant-revenue sector. 

比如，ChatGPT官网上挂着题为“教育者常见问题”的8条问题链接，ChatGPT通过该链接邀请教育从业者无偿提供想法，探索出AI技术如何能够“帮助教育者和学生”。“有很多办法能得到答案，但最佳答案将会来自教育界。”

For example, ChatGPT’s website provides an eight-(8!)-question ‘Educator FAQ’ which asks for free labour from those who teach to figure out how their technology can ‘help educators and students’: ‘There are many ways to get there, and the education community is where the best answers will come from.’

尽管如此，教学和教学辅助人员仍在努力花时间仔细考量生成式AI对我们教学和研究的影响，以及如何应对这些影响。

Still, teaching and teaching support staff have scrambled to find time to carefully think through generative AI’s implications for our teaching and research, as well as how we might address these. 

在研究方面，我们被推销各种技术方案以加快关键研究过程，例如总结阅读材料、撰写文献综述、进行主题分析、数据可视化、写作、编辑、引用参考文献和同行评审。

On the research side, we are being pushed various techno-solutions meant to speed up crucial research processes, such as summarising reading, writing literature reviews, conducting thematic analysis, visualising data, writing, editing, referencing and peer reviewing. 

有时，这些生成式AI的功能会在我们已使用的工具中弹出，就像定性分析软件ATLAS.ti推出的“AI Coding Beta”，该功能是“世界上首个用于定性研究的AI解决方案”，“为您节省了大量时间”。

Sometimes these options pop up within tools we already use, as when the qualitative analysis software ATLAS.ti launched AI Coding Beta, ‘the beginning of the world’s first AI solution for qualitative insights’, ‘saving you enormous time’.

省时——提效——这当然很好。但是效率并不总是，或者说并不常常是推动学术研究的关键标准。评估生成式AI对主流价值观的影响是了解其是否以及如何应用于教学、学习与研究的唯一办法。

Saving time – efficiency – is all well and good. But efficiency is not always, or even often, the core norm driving scholarship.  The only way to know if and how to adopt generative AI into our teaching, learning and research is through assessing its impact on our main values.

我们常听说学术卓越是学术研究的核心价值观。使用生成式AI来促进知识的产生是符合该核心价值观的，但只有在不危及到激发我们教学、学习及研究动力的其他价值观时，才有效果。 

We often hear of academic excellence as the core value of scholarship. The use of generative AI, where it facilitates knowledge generation, can be in line with this core value, but only productively if it doesn’t jeopardise the other values animating our teaching, learning and research.

如下所说，所谓其他价值观包括教育、伦理以及尤里卡时刻。我们必须详细讲讲这些其他价值观，由此才能对生成式AI如何能对学术研究产生潜在影响有全面的了解。

As described below, these include education, ethics and eureka. We have to broaden the conversation to these other values to fully understand how generative AI might impact scholarship. 

教育

Education

Education is at the heart of scholarship.  As students and scholars, understanding the how of scholarship is just as important as the what of scholarship, yet it often gets short shrift. Methodology is emphasised less than substantive topics in course provision, and teaching and learning often focuses more on theories than on how they were made and on how the makings shaped the findings. 

这种关注点的错位意味着我们一直不能及时发现使用生成式AI带来的问题，那就是它可能会让我们错失机会，没有办法学习以及展示进行学术研究所必须的关键技能。 

This misattention means we have been slower to notice that the adoption of generative AI may take away opportunities to learn and demonstrate the key skills underpinning the construction of scholarship. 

“在做中学”的教学法不仅适用于学生，同样适用于学术有成的学者。这个方法通常很慢、让人困惑不解、充满错误以及低效，但是对于创造出新的学术研究成果以及培养新一代学者来说是必要的。

Learning-by-doing is a pedagogical approach that applies just as much to the student as the established scholar. It is often slow, discombobulating, full of mistakes and inefficiencies, and yet imperative for creating new scholarship and new generations of scholars.

尽管生成式AI的确在某些方面能够辅助学术研究，我们必须保证自己充分理解并且有能力完成生成式AI能够首先取代的研究过程，比如说总结与整合文本内容，创建参考文献目录，分析数据以及构建论据。

Though generative AI can support scholarship in some ways, we should be sure that we understand and can undertake the processes generative AI replaces first, such as summarising and synthesising texts, generating bibliographies, analysing data and constructing arguments.

如果我们使用生成式AI，我们还需要思考它是如何影响获得教育的平等性的。一方面，有支付能力的使用者可以通过付费获得更强大的工具。另一方面，教育专家正在研究生成式AI辅助残疾学生的潜能。虽然过去的经验告诉我们不经三思就在课堂中使用AI，如抄写笔记，会产生严重的负面影响。

If we allow generative AI, we also have to think about how it impacts the equality of access to education. On the one hand, users who can pay have access to more powerful tools. On the other, educators are investigating the potential for generative AI to support disabled students, though past experience shows us that rushing into AI adoption, like transcription, in the classroom has had significant negative repercussions. 

生成式AI最初的魅力也分散了我们的注意力，使我们没有注意到在研究中使用生成式AI所涉及的复杂伦理问题，包括生成式AI从知识领域和人类环境提取内容的天性，以及生成式AI对重要研究价值观的负面影响，比如同理心、诚信和效度。 

The initial enchantment of generative AI also distracted us from the complex ethical considerations around using generative AI in research, including their extractive nature vis-à-vis both knowledge sectors and the environment, as well as the way they trouble important research values like empathy, integrity and validity. 

这些问题适用于研究伦理中更宽广的架构，即必须追求利益最大化和伤害最小化。

These concerns fit into a broader framework of research ethics as the imperative to maximise benefit and minimise harm.

我们更需要明白的是，许多大型语言模型的训练并没有得到允许或承认，而这些训练素材则来源于许多知识行业富有创造力和表现力的出色作品，包括艺术、文学、新闻业及学术界。 

We are ever more aware that many large language models have been trained, without permission or credit, on the creative and expressive works of many knowledge sectors, from art to literature to journalism to the academy. 

考虑到人文社科树大根深的引用文化规范——这种规范认可他人观点，展示观点之间的联系，并支持读者理解我们写作的背景——依赖于不会标注作品引用出处的研究和写作工具是令人不适的虚伪行为。 

Given the well-entrenched cultural norm of citation in our sector – which acknowledges the ideas of others, shows how ideas are connected, and supports readers in understanding the context of our writing – it is uncomfortably close to hypocritical to rely on research and writing tools that do not reference the works on which they are built. 

可持续性正逐渐成为大学和研究的核心价值观。使用生成式AI意味着使用云数据中心，也就是在使用稀缺的淡水资源并同时排放二氧化碳。 

Sustainability is increasingly a core value of our universities and our research. Engaging generative AI means calling on cloud data centres, which means using scarce freshwater and releasing carbon dioxide.  

与ChatGPT进行十到五十个来回的一段普通对话就需要半升水来冷却服务器，而要求一个大型生成式AI模型为您创建一幅图像所需的能量相当于将您的智能手机充至满电。我们很难忽略这些环境后果，这也让我们意识到在使用生成式AI以完成我们自己能做的工作时应当停下来想一想。

A typical conversation with ChatGPT, with ten to 50 exchanges, requires a half-litre of water to cool the servers, while asking a large generative AI model to create an image for you requires as much energy as charging your smartphone’s battery up all the way. It’s difficult to un-know these environmental consequences, and they should give us pause at using generative AI when we can do the same tasks ourselves.

Research ethics are about conducting research with empathy and pursuing validity, namely producing research that represents the empirical world well, as well as integrity, or intellectual honesty and transparency. Generative AI complicates all of these. Empathy is often created through proximity to our data and closeness to our subjects and stakeholders. Generative AI as the machine-in-the-middle interferes with opportunities to build and express empathy. 

生成式AI的黑盒属性可能会干扰研究的效度，因为我们无法确切知道它是如何从数据中识别出符合主题的代码，也无法了解它是如何在写作中产出观点的——更不用说生成式AI产出的这些观点和所作的引用可能都是杜撰的。 

The black box nature of generative AI can interfere with the production of validity, in that we cannot know exactly how it gets to the thematic codes it identifies in data, nor to the claims it makes in writing – not to mention that it may be hallucinating both these claims and the citations on which they are based. 

黑盒属性还给学术透明性带来了问题，因此也影响了学术诚信；至少，保持研究诚信意味着要诚实地说明我们如何以及何时使用了生成式AI。学术机构目前正在制定用于承认使用AI的模式声明和说明。

The black box also creates a problem for transparency, and thus integrity; at a minimum, maintaining research integrity means honesty about how and when we use generative AI, and scholarly institutions are developing model statements and rubrics for AI acknowledgements.  

此外，我们必须认识到生成式AI可能是在精华数据集中进行训练的，因此它会排除边缘化的观点并再创知识的等级制度，同时也会重现这些数据中固有的偏见——这引发了关于使用它所产生的伤害是否会持续存在的问题。 

Furthermore, we have to recognise that generative AI may be trained on elite datasets, and thus exclude minoritised ideas and reproduce hierarchies of knowledge, as well as reproduce biases inherent in this data – which raises questions about the perpetuation of harms arising from its use. 

与所有新技术一样，伦理框架正在迎头赶上对新领域的研究实践。在这段空白期内，遵循互联网研究者的建议是明智的：跟随你的直觉（如果感觉不对，那它可能确实有问题），并与你的研究伙伴和同事进行讨论、商榷和辩论。

As always with new technologies, ethical frameworks are racing to catch up with research practices on new terrains. In this gap, it is wise to follow the advice of internet researchers: follow your instinct (if it feels wrong, it possibly is) and discuss, deliberate and debate with your research collaborators and colleagues.

It’s not just our education and our ethics that generative AI challenges, but also our emotions. As academics, we don’t talk enough about how research and writing make us feel, yet those feelings animate much of what we do; they are the reward of the job. 

想象一下，当一堆巧妙但杂乱的定性数据逐渐有序形成理论时的那一刻，或者在实验室里发现数据证实了假设的那一瞬间，或者当一个用来解决问题的原型成功运作的时候，又或者当与同事一起午餐时突然涌现出的灵感。 

Think of the moment a beautiful mess of qualitative data swirls into theory, or the instant in the lab when it becomes clear the data is confirming the hypothesis, or when a prototype built to solve a problem works, or the idea that surfaces over lunch with a colleague. 

跟着找到这些数据的尤里卡时刻而来的，是写作中的尤里卡时刻，这些“瞬间”的感受也许在人文社科中具有特殊的意义（对于我们的一些同事来说，写作实际上是方法论的一部分）：通过书写形成一个论点所带来的满足感，用一句话准确描述实证世界的兴奋感，以及文字游戏带来的书呆子般的自豪感。当然，轮番去感受这些巨大的快乐是很痛苦的，它们往往互相依存。 

These data eurekas are followed by writing eurekas, ones that may have special relevance in the humanities and social sciences (writing is literally part of the methodology for some of our colleagues): the satisfaction of working out an argument through writing it out, the thrill of a sentence that describes the empirical world just so, the nerdy pride of wordplay. Of course, running alongside these great joys are great frustrations, the one dependent on the other. 

关键在于，这些学术研究带来的情感是学术研究中人性和满足感的核心所在。如果以符合伦理道德地方式使用生成式AI，它可以为我们创造追求这些情感的空间，从而创造知识。

The point is that these emotions of scholarship are core to scholarship’s humanity and fulfilment. Generative AI, used ethically, can make space for us to pursue them and in so doing, create knowledge.

但是，生成式AI也可以使我们对研究和写作的情感流失殆尽，把我们的作用瓜分成更狭隘和更机械式的检查和编辑工作。而且这种事情可以在毫不察觉的情况下发生，光鲜亮丽的效率承诺正逐渐蚕食这些尤里卡时刻。

But generative AI can also drain the emotions out of research and writing, parcelling our contributions into the narrower, more automatic work of checking and editing. And this can happen by stealth, with the shiny promise of efficiency eclipsing these fading eureka moments.

当然这种疏远的过程在将科技引入到工作中时也发生过，工人们一直在抵制这种情况，从19世纪英国纺织工人到如今的亚马逊仓库工人。像卢德派一样，当代运动经常因为抵制变革而受到批评，但这种批评忽略了关键点。像我们谈论的这个情况，这些拒绝和抵制运动的核心在于让我们注意到在技术进步中失去了什么。

Of course, this process of alienation is nothing new when it comes to the introduction of technologies into work, and workers have resisted it throughout time, from English textile workers in the 1800s to Amazon warehouse workers today. As the Luddites were, contemporary movements are often criticised for being resistant to change, but this criticism misses the point. Core to these refusal and resistance movements, as in this case, is noticing what we lose with technology’s gain.

在科技行业势如破竹般地不断使用新科技的背景之下，我们认为学术界应当冷静下来，花时间思考我们是否应该在学术研究中使用生成式AI，而非思考何时使用或如何使用。

In the context of a tech-sector fuelled push to adopt new technologies, we argue that the academy should take its time and question not when or how but if we should use generative AI in scholarship.

与其被学术不端的威胁所逼迫，对于生成式AI和学术研究的判断应该由我们的价值观所驱动。

Rather than being motivated by the stick of academic misconduct, decisions around generative AI and scholarship should be motivated by the carrot of our values.

生成式AI已经在学术研究界引发了一场知识论战。其突然的出现使我们习以为常的事物变质，并为我们提供了反思和更新价值观的机会——这些是我们决定是否以及如何将生成式AI纳入我们的教学、学习和研究的最佳衡量标准。

Generative AI has created a knowledge controversy for scholarship. Its sudden appearance has denaturalised the taken-for-granted and has created opportunities for reflection on and renewal of our values – and these are the best measure for our decisions around if and how we should incorporate generative AI into our teaching, learning and research.

鉴于上文所述，我们提倡关于AI的以下五大关键原则：

Based on the considerations above, we propose these five key principles on AI:

N0.1

思考它，谈论它。