We believe it is crucial that an AI-application is justified if it is to serve the public interest. To serve a public, this public must be able to get informed and involved. A justification, as we understand it, should make transparent how the AI application addresses the challenges it is designed for and why this AI system is the best way to do so. A justification should be public (or publicly accessible) and go beyond individual commercial interests.

Often, the principles related to equity in the AI-context are enhancing fairness and reducing bias. But equity should not be reduced to that: More fundamentally, a question that should be asked is whether the AI system is truly necessary and, in addition, how the system is located in current power relations – for example whether it impacts vulnerable groups negatively. Additionally, it needs to be clear that an AI system without bias does not automatically mean that it serves the public interest – think of a social media algorithm that works just as intended, but nevertheless increases inequality through misinformation spreading.

If AI is to succeed in supporting equity amongst all citizens, the design process should promote inclusive principles: First, equity in access is necessary, meaning especially open systems and open source where it is possible. Second, equity in knowledge should be promoted, meaning the civic tech literacy should be strengthened. Third, the equity between generations should be considered as well, especially regarding the impacts AI systems can have on the future.

A critical aspect of an application serving the public interest is that there is a certain “public” defined that has a stake in the development process. No single entity – no single person or organisation – can define what the public interest is. We believe that a deliberative development process must be encouraged to achieve this. Otherwise, it is very likely that the interests of “the public” will only be assumed, but never actually met.

There are various paths which a deliberative process can take: From transparent access to information, surveys, design research, co-design approaches, ethics boards or regular consultations with engaged citizens –  many approaches are possible.

Of course, participatory design is no guarantee for a democratic process, or a working product, not even a guarantee to serve the public interest. For example if forms of “participatory washing” allow only very exclusive participation and the process is getting inscrutable instead of opened up, “participation” can be misused.

A technical system that does not meet basic safety standards cannot serve the public interest. There are three distinct areas of technical safety measures.

First, the data quality and system accuracy must be considered. It is necessary to provide transparency about the data sources as well as their exact use and documentation of their inadequacies. Biases and incorrect outcomes can disqualify an application from truly serving the  public interest.

Second, data privacy is necessary, as privacy is understood as a prerequisite for the realization of an autonomous life.

Third, the overall system safety and security is important. As a starting point, failures should be monitored to ensure efforts are put in the right spots. In general, the system needs to be designed in a safe and robust way, so that, for example, malfunctions endanger neither public safety nor security.

As seen with the other conditions, the consideration of an application as serving the public interest does not end once an application is deployed but is an ongoing process. This is especially true for this fifth condition which states that an AI application needs to be open for validation by others.

This is the case because transparency is necessary, first, to avoid unintended societal harms, such as the perpetuation of existing stereotypes. Second, openness for validation is understood as a fundamental democratic norm, as civil society does not only have the right to understand how technology works, but also needs to be able to validate and confirm its mechanisms.

Finally, this is closely related to accountability, which is understood as a clear attribution of responsibility and liability. Thus, openness for validation indicates a direct feedback channel for citizens to those entities which are responsible for the AI-application and have the possibility to implement changes or terminate the use of the application.

As AI does not exist without consequences but quite materially and as its applications have tangible ecological, social and economic impacts, it is an important aspect of the development of an AI-solution to keep its sustainability in mind. Not only can this include ecological decisions which can have an impact on the CO2 emissions – as for example which model is deployed, where, how and when it is trained, and for which use cases it is made – but also questions of social and economic sustainability, which might tackle questions like how does this application impact working conditions. Economic sustainability should not be understood as serving private commercial interests, but as the ability of an AI project to sustain itself. Another question spanning all three fields of ecological, social and economic sustainability is the question of interoperability and the reusability of models and data, which are key factors to sustain small AI projects and to reimplement them for example in other domains as well.