The Ethics of Environmental AI: Modelling Public Perceptions

人工智能系统将在实现环境可持续发展的未来方面发挥重要作用。实时处理大量数据集的需求导致了这样的建议：这些系统必须成为我们在关键领域优化效率的主要工具；例如，农业产出、能源使用、运输供应和水资源管理。然而，人工智能本身也存在潜在的风险和环境成本，在我们开始将这些系统投入使用之前，需要解决重大的伦理和社会影响。当采用广泛的系统来提高效率时，有一种风险，即总体成本的降低（无论这些成本是什么）将以相对成本的不成比例分配为代价。我们已经看到人工智能可以学习人类的偏见，以及算法的效率如何最终加剧了已经存在的社会不平等。人工智能的干预可能会带来净环境或健康效益，但总会有权衡，我们的成本效益计算需要考虑到潜在成本可能采取的不同形式。我们在这一领域面临的成本效益分析将很难付诸实施。我们应该如何平衡预期的环境收益与潜在的伦理、社会和环境风险？考虑到的许多因素需要不同的评估度量，常常使权衡不可通约。考虑一下“智能灌溉”计划的影响：我们应该如何权衡水稻产量预期增加50%与生态学家警告的一种独特的当地物种灭绝的风险？（如果危险的动物“超级可爱”，情况会发生变化吗？）这些复杂问题只会增加伦理方面的担忧：如果预计每个月都会出现一次严重低估，我们是否应该引入一个通过模拟整个人口的饮食需求来减少食物浪费的系统？这可能以较小的净成本获得净收益，但这些成本是否会对弱势群体（例如最贫穷的家庭）造成不成比例的伤害？使用历史消费数据会提高准确性还是只会使现有的不平等永久化？公众对这些问题的看法是这个难题的重要组成部分，因为这些问题永远不会有绝对正确的答案。它为监管和政策制定提供了重要的投入，其好处不仅仅是做出“正确”的决定。众包偏好赋予人工智能系统的使用道德防御能力；它提倡自由/平等的民主价值观，承认每个人都有权就这些问题进行协商，并使人们能够轻松地表达自己的关切。我的研究将集中在公平和风险感知上，旨在更好地理解公众对环境收益的道德权衡的可接受性的偏好。环境人工智能系统可能被设计和实施的方式（由私营公司、政府部门、环保团体等）使得不可能在个案的基础上直接获得对个别提案的道德判断。因此，这个项目的最终目标是使用我收集的更一般的数据来构建一个计算框架，该框架可以准确地模拟人们偏好的机制并提供预测。在新的情况下，我们没有公众意见的先验指示，这个工具可以提供一定程度的道德指导，因为至少可以考虑公众（建模）的观点。不平等和权力不平衡为环境人工智能的实施带来了重大的伦理问题，其深远的影响源于设计系统的个人和制定系统的权威人士所做的决定。这个拟议中的项目有可能在这些问题上提供帮助，并在一定程度上减轻“气候不公正”的风险。

AI systems will play a large role in enabling an environmentally sustainable future. The need to process vast datasets in real time has led to proposals that such systems must become our primary tools for optimising efficiency in critical domains; for example, agricultural output, energy usage, transport provision, and water resource management. However, AIs present their own potential risks and environmental costs and there are major ethical and societal implications that need to be addressed before we start to put these systems to work. When employing wide-reaching systems for improved efficiency there is a risk that overall cost reduction (whatever those costs are) will come at the expense of a disproportionate distribution of relative costs. We have already seen that AIs can learn human biases and how algorithmic efficiency can end up reinforcing pre-existing social inequality. The interventions of AIs may bring net environmental or health benefits but there will always be trade-offs and our cost-benefit calculations need to account for the different forms that potential costs may take. The cost-benefit analyses that we face in this area are going to be very hard to operationalise. How should we balance projected environmental gains against potential ethical, social, and environmental risks? Many factors under consideration require different evaluation metrics, often making the trade-offs incommensurable. Consider the implications of a 'smart irrigation' initiative: how should we weigh an expected 50% increase to rice yield against ecologist's warnings that it risks the extinction of a unique local species? (Does this change if the animal at risk is "super cute?") These complications only increase for ethical concerns: should we bring in a system that reduces food wastage, by modelling the dietary needs across the population, if it is expected to give a major underestimate approximately once a month? This might give a net gain for a small net cost, but will the costs cause disproportionate harm to vulnerable groups (e.g., the poorest families)? Will using historical consumption data improve accuracy or just perpetuate existing inequalities? Public perception of these issues is an important part of this puzzle as such questions will never have categorically correct answers. It provides a crucial input for regulation and policy making with benefits beyond making the 'right' decision. Crowdsourcing preferences gives the use of AI systems ethical defensibility; it promotes democratic values of freedom/equality, recognising that everyone has a right to consultation on these matters and enabling people to easily voice their concerns. My research will focus on fairness and risk perception, aiming to better understand public preferences regarding the acceptability of moral trade-offs for environmental gains. The manner in which environmental AI systems are likely to be designed and implemented (by private companies, governmental departments, environmental groups etc.) make it impossible to directly acquire moral judgments of individual proposals on a case-by-case basis. Accordingly, the ultimate goal of this project is to use the more general data I collect to build a computational framework that can accurately model the mechanisms underlying people's preferences and provide predictions. In novel scenarios, where we have no a priori indication of public opinion, this tool could provide some level of ethical guidance as at least the public's (modelled) views can be considered. Inequality and power imbalances pose major ethical concerns for the implementation of environmental AIs, with far-reaching ramifications originating in the decisions made by individuals designing the systems and those in authority who enact them. This proposed project has the potential to aid in such matters and go some way towards mitigating the risks of 'climate injustice.'