Delivering data and AI with care for ethics
Delivering data and AI with care for ethics
Data solutions and artificial intelligence can bring big efficiencies and benefits for users and the public at large. For example, by introducing intelligent document comparison, HM Land Registry was able to greatly improve efficiency by removing manual comparisons from their work process and to simultaneously reduce the risk of indemnity claims.
However, such structured and intelligent solutions do not always come without risk, and several incidents have shown the potential for (often unintended) negative impacts. For instance, a fraud detection algorithm used in the Dutch government’s benefit system targeted neighbourhoods with mostly low-income and minority residents. The District Court of The Hague ruled that the system conflicted with EU human rights.
Reviewing 106 such incidents that took place between 2011-2022, Holweg, Younger and Wen found that about half of ‘AI failure’ cases were related to privacy, just under a third to bias, and an additional 14% to explainability failures (not being clear why a decision was reached by AI). The range of problems that can occur illustrate how building ethical solutions requires careful consideration of multiple risks for harm throughout the building of data and AI solutions.
For this reason, data ethics is a focus area for Kainos when it comes to delivering our data and AI services. Ethics informs our ways of working in any project, in addition to dedicated subject matter expertise in data ethics being available for selected projects. From mobilisation to go live, we work together with our clients to consider and investigate ethical questions and risks, guided by five data and AI ethical delivery principles:
- Mitigating Harms & Seeking Benefits: Within projects, we support our clients to minimise harms and maximise societal benefits of data and AI solutions
- Fairness: We minimise bias in data and AI solutions and scrutinise them for fairness
- Transparency & Explainability: We communicate with transparency about data and AI solutions and explore the reasons behind their conclusions
- Respect for the Human Behind the Data: We design data and AI solutions with respect for the autonomy, dignity and privacy of those affected by the technology
- Responsibility: We believe in taking accountability for data and AI solutions and will encourage clients to establish appropriate oversight
In case of dedicated data ethics support, activities include an explicit tracking of ethical considerations throughout project delivery and the running of ethics and harm workshops. In these workshops, we seek to identify what unintended harms could occur as a result of using the solution and how to best mitigate those harms while maximising potential benefits. When mitigations are identified that can be built directly into the system architecture, that information will be fed to the relevant members of the team to consider incorporating within the project.
Conclusions from the ethics and harm workshop, supplemented by insights from a Data Ethicist and summaries of deployed risk and harm mitigations are brought together in an algorithmic impact assessment document. In addition, transparency documentation is shared, in the form of a model card of a completed template of the CDDO and CDEI’s algorithmic transparency recording standard.
Our principles and approach recognise both key themes in the data ethics landscape (see e.g. Jobin, Ienca & Vayena, 2019; CDDO's data ethics framework) and the developing legislative and policy landscape. Noting that every project is different, we will tailor the exact components and focus of the data ethics work to the needs of the project.
If you have any questions about AI or data ethics and how it can help your organisation, please get in touch.