Realising Accountable Intelligent Systems (RAInS)

智能系统技术正在越来越多的场景中得到应用，包括自动驾驶汽车、智能家电、公共服务、零售和制造业。但是，如果这些系统出现故障，就像最近发生的涉及自动驾驶汽车的引人注目的事故一样，会发生什么呢？如果发现这些系统（及其开发人员）做出了有偏见或不公平的决定，该如何追究其责任？我们能否在部署智能系统之前对其进行审查，以确保它们适合使用？鉴于智能系统将越来越多地影响日常生活的许多方面，这些都是现实而及时的挑战。虽然所有新技术都有可能造成伤害，但对于智能系统来说，可能很难甚至不可能知道哪里出了问题，或者谁应该为此负责。有一个非常现实的担忧是，许多人工智能技术的复杂性，周围系统和工作流程之间的数据和交互，将减少相应决策的理由，“算法让我这样做”，或者实际上“我们不知道发生了什么”。然而，这些系统在决策的准确性甚至安全性方面都有超过人类的潜力，这表明使用它们的愿望将难以抗拒。因此，问题是我们如何努力做到两全其美。我们如何在不放弃对问责制、透明度和责任的渴望的情况下，从这些系统提供的超人的能力和效率中受益？我们如何才能避免在完全放弃自动化系统的好处或接受某种程度的随意性之间做出选择的僵局，这种随意性在社会上通常的人际关系中是不可想象的？我们将与公众、法律界和科技公司等持份者紧密合作，探讨如何实现透明和可问责的未来智能系统。问责结构是我们对未来支持此类系统审计的计算基础设施的设想——有点类似于（但比）与客机相关的“黑匣子”飞行记录器。我们的工作不仅将在事后增加透明度，而且还将以一种允许早期审讯和审计的方式增加透明度，这反过来可能有助于预防或减轻事前损害。在我们能够实现问责结构之前，需要调查几个关键问题：影响公民对智能系统信任和问责的看法的重要因素是什么？智能系统的法律责任应该采取什么形式？法律如何公平运作，并激励那些开发/使用这些系统的人做出最佳行为？我们如何制定一个适当的词汇来描述和表征智能系统、它们的背景、行为和偏见？记录智能系统行为的技术手段是什么，从使用的数据、部署的算法到所做决策的后续影响？我们能否实现智能系统的问责解决方案，跨越一系列技术和组织边界，能够支持第三方审计和评估？这些问题的答案（以及许多其他肯定会出现的问题）将引导我们开发将与项目合作伙伴一起评估的原型解决方案。我们的目标是创造一种方法，通过这种方法，智能系统的开发人员可以提供安全的、防篡改的系统特征和行为记录，这些记录可以在发生事件或投诉时（在受控情况下）与有关当局共享。

Intelligent systems technologies are being utilised in more and more scenarios including autonomous vehicles, smart home appliances, public services, retail and manufacturing. But what happens when such systems fail, as in the case of recent high-profile accidents involving autonomous vehicles? How are such systems (and their developers) held to account if they are found to be making biased or unfair decisions? Can we interrogate intelligent systems, to ensure they are fit for purpose before they're deployed? These are all real and timely challenges, given that intelligent systems will increasingly affect many aspects of everyday life.While all new technologies have the capacity to do harm, with intelligent systems it may be difficult or even impossible to know what went wrong or who should be held responsible. There is a very real concern that the complexity of many AI technologies, the data and interactions between the surrounding systems and workflows, will reduce the justification for consequential decisions to "the algorithm made me do it", or indeed "we don't know what happened". And yet the potential for such systems to outperform humans in accuracy of decision-making, and even safety suggests that the desire to use them will be difficult to resist. The question then is how we might endeavour to have the best of both worlds. How can we benefit from the superhuman capacity and efficiency that such systems offer without giving up our desire for accountability, transparency and responsibility? How can we avoid a stalemate choice between forgoing the benefits of automated systems altogether or accepting a degree of arbitrariness that would be unthinkable in society's usual human relationships?Working closely with a range of stakeholders, including members of the public, the legal profession and technology companies, we will explore what it means to realise future intelligent systems that are transparent and accountable. The Accountability Fabric is our vision of a future computational infrastructure supporting audit of such systems - somewhat analogous to (but more sophisticated than) the 'blackbox' flight recorders associated with passenger aircraft. Our work will increase transparency not only after the fact, but also in a manner which allows for early interrogation and audit which in turn may help to prevent or to mitigate harm ex ante. Before we can realise the Accountability Fabric, several key issues need to be investigated:What are the important factors that influence citizen's perceptions of trust and accountability of intelligent systems? What form ought legal liability take for intelligent systems? How can the law operate fairly and incentivize optimal behaviour from those developing/using such systems?How do we formulate an appropriate vocabulary with which to describe and characterise intelligent systems, their context, behaviours and biases?What are the technical means for recording the behaviour of intelligent systems, from the data used, the algorithms deployed, and the flow-on effects of the decisions being made?Can we realise an accountability solution for intelligent systems, operating across a range of technologies and organisational boundaries, that is able to support third party audit and assessment? Answers to these (and the many other questions that will certainly emerge) will lead us to develop prototype solutions that will be evaluated with project partners. Our ambition is to create a means by which the developer of an intelligent system can provide a secure, tamper-proof record of the system's characteristics and behaviours that can be shared (under controlled circumstances) with relevant authorities in the event of an incident or complaint.