Principled requirements of ML workflows (ESR2)

In an effort to address major socio-ethical challenges that arise from the development, implementation and usage of algorithmic systems, various design and auditing strategies have been suggested in recent years. On the one hand, solutions relying on technical fixes have been impactful to deal with harmful algorithmic bias but fall short when dealing with other forms of harmful algorithmic behavior. On the other hand, ethical guidelines that deal with harmful algorithmic behavior have been accused of being toothless and non-actionable. This research project focuses on overcoming these limitations and translating ethical guidelines into actual system requirements.

We, therefore, seek to answer the following research question:

How can human values be accounted for in algorithmic systems so that they drive the design and evaluation processes of such systems?

To this end, we are exploring a combination of approaches coming both from the design and computational fields. We started by performing a literature review of ethics in AI documents and designed a framework for operationalizing values (See our publication at the 2022 ACM Conference on Fairness, Accountability and Transparency -FAccT’22-). We translated values into specific criteria and their manifestations. We also mapped stakeholders that should be part of value negotiation processes with tools that aim at enacting value-specific manifestations.

Currently, we are exploring design strategies and methodologies to enable and facilitate multi-stakeholder negotiation of values in algorithmic systems. These will allow us to conduct design workshops with relevant stakeholders and test our theoretical framework in synthetic and real-world scenarios.

For our next steps, we will adopt a more computational standpoint and explore practices and tools coming from software engineering to enable effective and systematic value-driven design and evaluation of algorithmic systems. This involves mapping the decision-making nodes along the design, data procurement, building, testing and deployment of Machine Learning-driven systems. It also involves devising computational tools for making values encoded along the pipeline explicit and traceable.