Developing safe and useful general-purpose AI systems will require us to make progress on scalable oversight: the problem of supervising systems that potentially outperform us on most skills relevant to the task at hand. Empirical work on this problem is not straightforward, since we do not yet have systems that broadly exceed our abilities. This paper discusses one of the major ways we think about this problem, with a focus on how to turn it into one that can be productively studied empirically. We first present an experimental design centered on choosing tasks for which human specialists succeed but unaided humans and current general AI systems fail. We then present a proof-of-concept experiment following meant to demonstrate a key feature of this experimental design and show its viability with two question-answering tasks: MMLU and time-limited QuALITY. On these tasks, we find that human participants who interact with an unreliable large-language-model dialog assistant through chat -- a trivial baseline strategy for scalable oversight -- substantially outperform both the model alone and their own unaided performance. These results are an encouraging sign that scalable oversight will be tractable to study with present models and bolster recent findings that large language models can productively assist humans with difficult tasks.

我们介绍了Sparrow，这是一个寻求信息的对话代理，与提示的语言模型基线相比，训练有素，更有帮助，正确和无害。我们使用从人类反馈中的强化学习来培训我们的模型，以帮助人类评估者判断代理人的行为。首先，为了使我们的代理人更有帮助和无害，我们将良好对话的要求分解为代理人应遵循的自然语言规则，并分别向评估者询问每个规则。我们证明，这种崩溃使我们能够收集对代理行为的更多针对性的人类判断，并允许更有效的规则条件奖励模型。其次，我们的代理商在收集对模型声明的偏好判决时提供了支持事实主张的来源的证据。对于事实问题，麻雀提供的证据支持了78％的时间。比基线比基线更享受麻雀，同时对人类的对抗性探测更具弹性，在探测时只有8％的时间违反了我们的规则。最后，我们进行了广泛的分析，表明尽管我们的模型学会遵守我们的规则，但它可以表现出分布偏见。

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses. In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences. We then train with RL using the preference model as the reward signal, i.e. we use 'RL from AI Feedback' (RLAIF). As a result we are able to train a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels.

Many real-world applications of language models (LMs), such as code autocomplete and writing assistance, involve human-LM interaction, but the main LM benchmarks are non-interactive, where a system produces output without human intervention. To evaluate human-LM interaction, we develop a framework, Human-AI Language-based Interaction Evaluation (H-LINE), that expands non-interactive evaluation along three dimensions, capturing (i) the interactive process, not only the final output; (ii) the first-person subjective experience, not just a third-party assessment; and (iii) notions of preference beyond quality. We then design five tasks ranging from goal-oriented to open-ended to capture different forms of interaction. On four state-of-the-art LMs (three variants of OpenAI's GPT-3 and AI21's J1-Jumbo), we find that non-interactive performance does not always result in better human-LM interaction and that first-person and third-party metrics can diverge, suggesting the importance of examining the nuances of human-LM interaction.

Incivility remains a major challenge for online discussion platforms, to such an extent that even conversations between well-intentioned users can often derail into uncivil behavior. Traditionally, platforms have relied on moderators to -- with or without algorithmic assistance -- take corrective actions such as removing comments or banning users. In this work we propose a complementary paradigm that directly empowers users by proactively enhancing their awareness about existing tension in the conversation they are engaging in and actively guides them as they are drafting their replies to avoid further escalation. As a proof of concept for this paradigm, we design an algorithmic tool that provides such proactive information directly to users, and conduct a user study in a popular discussion platform. Through a mixed methods approach combining surveys with a randomized controlled experiment, we uncover qualitative and quantitative insights regarding how the participants utilize and react to this information. Most participants report finding this proactive paradigm valuable, noting that it helps them to identify tension that they may have otherwise missed and prompts them to further reflect on their own replies and to revise them. These effects are corroborated by a comparison of how the participants draft their reply when our tool warns them that their conversation is at risk of derailing into uncivil behavior versus in a control condition where the tool is disabled. These preliminary findings highlight the potential of this user-centered paradigm and point to concrete directions for future implementations.

We are currently unable to specify human goals and societal values in a way that reliably directs AI behavior. Law-making and legal interpretation form a computational engine that converts opaque human values into legible directives. "Law Informs Code" is the research agenda capturing complex computational legal processes, and embedding them in AI. Similar to how parties to a legal contract cannot foresee every potential contingency of their future relationship, and legislators cannot predict all the circumstances under which their proposed bills will be applied, we cannot ex ante specify rules that provably direct good AI behavior. Legal theory and practice have developed arrays of tools to address these specification problems. For instance, legal standards allow humans to develop shared understandings and adapt them to novel situations. In contrast to more prosaic uses of the law (e.g., as a deterrent of bad behavior through the threat of sanction), leveraged as an expression of how humans communicate their goals, and what society values, Law Informs Code. We describe how data generated by legal processes (methods of law-making, statutory interpretation, contract drafting, applications of legal standards, legal reasoning, etc.) can facilitate the robust specification of inherently vague human goals. This increases human-AI alignment and the local usefulness of AI. Toward society-AI alignment, we present a framework for understanding law as the applied philosophy of multi-agent alignment. Although law is partly a reflection of historically contingent political power - and thus not a perfect aggregation of citizen preferences - if properly parsed, its distillation offers the most legitimate computational comprehension of societal values available. If law eventually informs powerful AI, engaging in the deliberative political process to improve law takes on even more meaning.

我们调整了大型语言模型，以使用行为克隆来编写自然语言批评（自然语言批判性评论）。关于基于主题的摘要任务，我们的模型所写的批评帮助人类在摘要中发现了本来会错过的漏洞。我们的模型有助于在模型和人类书面摘要中发现自然存在的缺陷，以及人类撰写的摘要中有意误导的摘要中的缺陷。我们研究批评的缩放特性，包括基于主题的汇总和合成任务。较大的模型写出更多有用的批评，在大多数任务上，尽管产生了更困难的输出，但在大多数任务上都更好地进行了自我关注。较大的模型还可以将自己的自我批评纳入反馈，将自己的摘要完善为更好的摘要。最后，我们激励并引入了一个框架，以比较批评能力的产生和歧视能力。我们的测量表明，即使是大型模型也可能仍然具有他们无法或不表达为批评的相关知识。这些结果是使用AI辅助的人类反馈来扩展机器学习系统的监督到人类直接评估的任务的概念证明。我们释放培训数据集以及批评援助实验的样本。

As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but the quality bar for medical and clinical applications is high. Today, attempts to assess models' clinical knowledge typically rely on automated evaluations on limited benchmarks. There is no standard to evaluate model predictions and reasoning across a breadth of tasks. To address this, we present MultiMedQA, a benchmark combining six existing open question answering datasets spanning professional medical exams, research, and consumer queries; and HealthSearchQA, a new free-response dataset of medical questions searched online. We propose a framework for human evaluation of model answers along multiple axes including factuality, precision, possible harm, and bias. In addition, we evaluate PaLM (a 540-billion parameter LLM) and its instruction-tuned variant, Flan-PaLM, on MultiMedQA. Using a combination of prompting strategies, Flan-PaLM achieves state-of-the-art accuracy on every MultiMedQA multiple-choice dataset (MedQA, MedMCQA, PubMedQA, MMLU clinical topics), including 67.6% accuracy on MedQA (US Medical License Exam questions), surpassing prior state-of-the-art by over 17%. However, human evaluation reveals key gaps in Flan-PaLM responses. To resolve this we introduce instruction prompt tuning, a parameter-efficient approach for aligning LLMs to new domains using a few exemplars. The resulting model, Med-PaLM, performs encouragingly, but remains inferior to clinicians. We show that comprehension, recall of knowledge, and medical reasoning improve with model scale and instruction prompt tuning, suggesting the potential utility of LLMs in medicine. Our human evaluations reveal important limitations of today's models, reinforcing the importance of both evaluation frameworks and method development in creating safe, helpful LLM models for clinical applications.

Recent developments in natural language generation (NLG) using neural language models have brought us closer than ever to the goal of building AI-powered creative writing tools. However, most prior work on human-AI collaboration in the creative writing domain has evaluated new systems with amateur writers, typically in contrived user studies of limited scope. In this work, we commissioned 13 professional, published writers from a diverse set of creative writing backgrounds to craft stories using Wordcraft, a text editor with built-in AI-powered writing assistance tools. Using interviews and participant journals, we discuss the potential of NLG to have significant impact in the creative writing domain--especially with respect to brainstorming, generation of story details, world-building, and research assistance. Experienced writers, more so than amateurs, typically have well-developed systems and methodologies for writing, as well as distinctive voices and target audiences. Our work highlights the challenges in building for these writers; NLG technologies struggle to preserve style and authorial voice, and they lack deep understanding of story contents. In order for AI-powered writing assistants to realize their full potential, it is essential that they take into account the diverse goals and expertise of human writers.

鉴于大型语言模型的广泛能力，应该有可能朝着一般的文本的助手工作，这些助手与人类价值一致，这意味着它是有帮助，诚实的和无害的。在此方向上的初始遗传，我们研究简单的基线技术和评估，例如提示。我们发现，从模型规模增加适度的干预措施的好处，概括为各种对准评估，并不会损害大型模型的性能。接下来，我们调查与对齐，比较仿制，二进制歧视和排名偏好建模相关的几个培训目标的缩放趋势。我们发现排名优先级模型比模仿学习更好地表现得多，并且通常以模型大小更有利地缩放。相比之下，二进制歧视通常与模仿学习非常类似地执行和缩放。最后，我们研究了一种“偏好模型预训练阶段的培训阶段，其目的是在对人偏好的芬明时提高样本效率。

语言模型可以根据给定的文化背景产生有害和偏置的输出并表现出不良行为。我们提出了一种将语言模型适应社会（PALM）与值目标数据集的过程，以通过在反映预定的一组目标值集合的数据集上进行制备和微调来显着地改变模型行为的迭代过程。我们使用三个指标评估我们的进程：具有人类评估的定量指标，将输出遵守目标值，毒性评分对产出;和定性度量分析与给定社会类别相关的最常见的单词。通过每次迭代，我们根据来自评估的观察到的缺点添加其他培训数据集示例。与基线和控制模型相比，PALMS在所有指标上显着更好地为广泛的GPT-3语言模型尺寸进行了基线和控制模型，而不会影响能力完整性。我们发现PALMS的有效性随模型规模而增加。我们表明，显着调整语言模型行为与小型手腕策划数据集是可行的。

随着人工智能系统变得越来越强大和普遍，人们对机器的道德或缺乏道德的关注变得越来越关注。然而，向机器讲授道德是一项艰巨的任务，因为道德仍然是人类中最激烈的争论问题之一，更不用说AI了。但是，部署到数百万用户的现有AI系统已经在做出充满道德影响的决策，这构成了一个看似不可能的挑战：教学机器的道德意义，而人类继续努力努力。为了探索这一挑战，我们介绍了Delphi，这是一个基于深层神经网络的实验框架，直接训练了描述性道德判断，例如，“帮助朋友”通常是不错的，而“帮助朋友传播假新闻”不是。经验结果提供了对机器伦理的承诺和局限性的新见解。面对新的道德情况，德尔菲（Delphi）表现出强大的概括能力，而现成的神经网络模型表现出明显差的判断，包括不公正的偏见，证实了对明确教学机器的道德意义的必要性。然而，德尔菲并不完美，表现出对普遍性偏见和不一致的敏感性。尽管如此，我们还是展示了不完美的Delphi的积极用例，包括在其他不完美的AI系统中将其用作组件模型。重要的是，我们根据著名的道德理论来解释Delphi的运营化，这使我们提出了重要的未来研究问题。

为了实现长文档理解的构建和测试模型，我们引入质量，具有中文段的多项选择QA DataSet，具有约5,000个令牌的平均长度，比典型的当前模型更长。与经过段落的事先工作不同，我们的问题是由阅读整个段落的贡献者编写和验证的，而不是依赖摘要或摘录。此外，只有一半的问题是通过在紧缩时间限制下工作的注释器来应答，表明略读和简单的搜索不足以一直表现良好。目前的模型在此任务上表现不佳（55.4％），并且落后于人类性能（93.5％）。

我们研究语言模型是否可以评估自己主张的有效性，并预测他们能够正确回答的问题。我们首先表明，当以正确的格式提供时，较大的模型在多样化的多项选择和True/False问题上进行了很好的校准。因此，我们可以通过要求模型首先提出答案，然后评估其答案正确的概率“ p（true）”来对开放式采样任务进行自我评估。我们发现在各种任务中，P（true）的表现，校准和缩放令人鼓舞。当我们允许模型考虑自己的许多样本之前，在预测一种特定可能性的有效性之前，自我评估的性能进一步改善。接下来，我们研究是否可以培训模型来预测“ P（ik）”，即“我知道”问题的概率，而无需参考任何特定提出的答案。模型在预测P（IK）方面表现良好，并且在跨任务中部分概括，尽管它们在新任务上的P（IK）校准方面遇到了困难。预测的p（IK）概率在存在相关的原始材料的情况下以及对数学单词问题解决方案的提示也适当增加。我们希望这些观察结果为培训更诚实的模型提供了基础，并研究了诚实对模型模仿人类写作以外的其他目标培训的案例的普遍性。

如果未来的AI系统在新的情况下是可靠的安全性，那么他们将需要纳入指导它们的一般原则，以便强烈地认识到哪些结果和行为将是有害的。这样的原则可能需要得到约束力的监管制度的支持，该法规需要广泛接受的基本原则。它们还应该足够具体用于技术实施。本文从法律中汲取灵感，解释了负面的人权如何履行此类原则的作用，并为国际监管制度以及为未来的AI系统建立技术安全限制的基础。

我们微调GPT-3使用基于文本的Web浏览环境来回答长形问题，允许模型搜索和导航Web。通过建立任务，以便通过人类执行，我们能够使用模仿学习培训在任务上的模型，然后通过人体反馈优化答案质量。为了使人为评估事实精度更容易，模型必须在浏览支持答案时收集引用。我们在ELI5上培训并评估我们的模型，Reddit用户提出的问题数据集。我们的最佳模型是通过使用行为克隆进行微调GPT-3获得的，然后对训练训练的奖励模型进行拒绝采样来获得以预测人类偏好。这种模式的答案是人类56％的答案，我们的人类示威者的时间和69％的时间到Reddit的最高投票答复。

There has been a recent resurgence in the area of explainable artificial intelligence as researchers and practitioners seek to make their algorithms more understandable. Much of this research is focused on explicitly explaining decisions or actions to a human observer, and it should not be controversial to say that looking at how humans explain to each other can serve as a useful starting point for explanation in artificial intelligence. However, it is fair to say that most work in explainable artificial intelligence uses only the researchers' intuition of what constitutes a 'good' explanation. There exists vast and valuable bodies of research in philosophy, psychology, and cognitive science of how people define, generate, select, evaluate, and present explanations, which argues that people employ certain cognitive biases and social expectations towards the explanation process. This paper argues that the field of explainable artificial intelligence should build on this existing research, and reviews relevant papers from philosophy, cognitive psychology/science, and social psychology, which study these topics. It draws out some important findings, and discusses ways that these can be infused with work on explainable artificial intelligence.

我们提出了一项探索性定性研究，以了解作家如何与下一页建议相互作用。尽管对建议系统对写作的影响进行了一些定量研究，但几乎没有定性的工作来理解作家如何与建议系统互动及其如何影响他们的写作过程 - 特别是针对非本地但英国作家的。我们进行了一项研究，要求业余作家分别写两部电影评论，一本没有建议。我们发现作家以各种复杂的方式与下一页建议互动 - 作家能够抽象建议的多个部分并将其纳入他们的写作中 - 即使他们不同意整个建议。建议系统对写作过程也有各种影响 - 以独特的方式为写作过程的不同方面做出了影响。我们提出了一种用于与GPT-2写作的作家 - 探索互动模型，用于电影评论写作任务，然后是该模型可用于未来研究的方式，并概述了研究和设计的机会。

在线众包平台使对算法输出进行评估变得容易，并提出诸如“哪个图像更好，A或B？”之类的问题的调查，在视觉和图形研究论文中的这些“用户研究”的扩散导致了增加匆忙进行的研究充其量是草率且无知的，并且可能有害和误导。我们认为，在计算机视觉和图形论文中的用户研究的设计和报告需要更多关注。为了提高从业者的知识并提高用户研究的可信度和可复制性，我们提供了用户体验研究（UXR），人类计算机互动（HCI）和相关领域的方法论的概述。我们讨论了目前在计算机视觉和图形研究中未利用的基础用户研究方法（例如，需要调查），但可以为研究项目提供宝贵的指导。我们为有兴趣探索其他UXR方法的读者提供了进一步的指导。最后，我们描述了研究界的更广泛的开放问题和建议。我们鼓励作者和审稿人都认识到，并非每项研究贡献都需要用户研究，而且根本没有研究比不小心进行的研究更好。