21 September 2023
The transformative future potential of AI alongside its ethical risks are frequently talked about, receiving particular attention since the advent of powerful Large Language Models (LLMs) such as ChatGPT. But AI in government is already a reality. Our role is to ensure it is used responsibly, ethically, and in the right places. There exist many frameworks which outline the principles of ethical/trustworthy AI, but these are often theoretical and leave designers without practical lessons they can apply to their own work.
So, how can designers avoid ethical nightmares when integrating AI into services? Let’s first consider an example of an ethical nightmare that has arisen from attempting to integrate AI into government services.
Dutch tax authorities employed an AI system based on a self-learning algorithm to try and weed out benefits fraud at early stages. This system would generate risk profiles using indicators drawn from vast arrays of data. These risk profiles would often disproportionately suspect those from ethnic minorities or lower economic brackets of committing fraud given the types of data and proxy variables used. Authorities would then penalise those who were suspected of fraud based solely on this system’s outputs.
The consequences of this were dire: tens of thousands of families were pushed into poverty as a result of huge debts owed to tax authorities and more than a thousand children were taken into foster care. It is important to note here the disparate impact this will have had on minority ethnic communities given how this system created these risk profiles.
Use of this system was halted when external, government-backed auditors concluded that it was insufficiently transparent or accountable: citizens could not trace how or why a particular decision had been made and had insufficient recourse to challenge such a decision.
This case illustrates a number of the ethical concerns with integrating AI into public services: which groups does it affect? Does it affect them differently? Can they understand how it affects them? Can they challenge any decisions it makes?
Appropriately and comprehensively assessing these risks is a complex task. But that does not mean that it is an insurmountable one. In fact, there are a number of practical tools that organisations have developed to help implement AI solutions in a considered, human-centred and responsible way. These tools include:
Stakeholder impact assessments are perhaps the most useful tool in a service designers arsenal when looking to help teams think through the potential ways in which the application of AI could cause benefits and harm to different groups of people. SIAs encourage teams to think about the different user groups it is important to test the system with, and to think about potential negative consequences of the system and how to mitigate against them.
In their publication on “Understanding Artificial Intelligence Ethics and Safety,” the Alan Turing Institute presents an outline of a stakeholder impact assessment for an AI project in its alpha phase. Inspired in part by this work, and our own trustworthy AI assessment, the Oxford Insights abridged algorithmic impact assessment offers teams questions they can use to have an initial conversation about the types of stakeholders that could be affected by an AI project and prompts teams to think through how they could mitigate against any negative consequences of the system.
Data ethics frameworks are intended to help teams involved in a project which involves the use and collection of data to explore potential ethical issues related to the collection, storage, transfer and use of data, both at the beginning of a project and throughout the project lifecycle.
The Open Data Institute’s Data Ethics Canvas “provides a framework to develop ethical guidance that suits any context,” while the UK Government’s Central Digital and Data Data Office’s Data Ethics Framework gives “a set of principles to guide the design of appropriate data use in the public sector.” Both of these resources can be helpful to designers as they seek to navigate the ethical issues related to data collection and use, and think through these questions early on in their work.
Designers can also benefit from following AI ethics guidelines which are tailored to local context, to ensure that particular cultural or legislative concerns are addressed. A list of national governments which have started to adopt their own ethical guidelines around the use of algorithms has been mapped in the AI Ethics Guidelines Global Inventory by Algorithm Watch.
Ensuring that examples of AI in government are published to open registries for algorithms and AI applications also represents a positive stride toward increased transparency, allowing residents to track how algorithms and AI are used in their cities or countries and provide feedback. At the city level, Amsterdam and Helsinki are notable examples of locales that have already implemented such registries. On the national level, the Public Law Project has also created a repository of automation examples in the UK government, ranking them in terms of transparency.
At Oxford Insights, we have also developed a Trustworthy AI assessment, intended to help government officials and researchers evaluate the extent to which AI is being used in a given country in a trustworthy and ethical manner. The assessment includes a framework and assessment questions under the five pillars of public purpose, human-centred values, transparency and explainability, robustness, security and safety, and accountability.
Using the assessment questions covered under each pillar, a user can determine a given government’s score in each of the pillars, showing the areas that the government is doing well in and areas in which they could improve in their implementation of trustworthy AI systems in the public sector.
Popular discourse around AI can range from unbounded optimism to total risk aversion. We believe it’s important to chart a balance between these two perspectives, recognising that AI can have positive benefits while keeping a realistic view of potential negative impacts. Designers have an important role to play in designing through this complexity. By co-designing AI applications alongside those who will be affected by them, we can advocate for a human-centred approach to artificial intelligence, which is vigilant to the risk of discrimination, and champions transparency and accountability.
Building trustworthy and ethical AI systems is not something that should be discussed once, but instead necessitates an ongoing process of dialogue with the community of people affected by the system. We will explore these ideas in more detail in our next blog where we’ll discuss what we learned about AI ethics during our time at the Service Design in Government conference in Edinburgh.