Report Open Access

What to explain when explaining is difficult. An interdisciplinary primer on XAI and meaningful information in automated decision-making

Asghari, Hadi; Birner, Nadine; Burchardt, Aljoscha; Daniela Dicks; Faßbender, Judith; Feldhus, Nils; Hewett, Freya; Hofmann, Vincent; Kettemann, Matthias C.; Schulz, Wolfgang; Simon, Judith; Stolberg-Larsen, Jakob; Züger, Theresa

Automated decision making (ADM) systems have become ubiquitous in our everyday lives, enabling new business models and intensifying the datafication of our economies. Yet, the use of these systems entails risks on an individual as well as on a societal level. Explanations of how such systems make decisions (often referred to as explainable AI, or XAI, in the literature) can be considered a promising way to mitigate their negative effects. Explanations of the process and decision of an ADM system can empower users to legally appeal a decision, challenge developers to be aware of the negative side effects of the ADM system during the entire development process, and increase the overall legitimacy of the decision. However, it remains unclear what content an explanation has to include and how the explanation can be made to achieve an actual gain of knowledge for the recipient, especially if that recipient is not an expert. The GDPR provides a legal framework for explaining ADM systems. “Meaningful information about the logic involved” has to be provided to the person affected by the decision. Nonetheless, neither the text of the GDPR itself nor the commentaries on the GDPR provide details on what “meaningful information about the logic involved” precisely is. 

This interdisciplinary report discusses this question from a legal, design, and technical perspective. The paper proposes three questions to help formulate a good explanation: Who needs to understand what in a given scenario? What can be explained about the system in use? What should explanations look like in order to be meaningful to affected users? The outcomes could potentially not only advance the debate among legal scholars but also help developers and designers to understand the legal obligations when developing or implementing an ADM system.

Legally, the explanation has to enable the user to appeal the decision made by the ADM system. “The logic” can be understood as “the structure and sequence of the data processing”. This does not necessarily have to include complete disclosure of the entire technical functioning of the ADM system. Since the explanation is intended to balance the power of the ADM developer with those of the user, this balance has to be at the center of the explanation. The GDPR focuses on individual rather than collective rights. This is the subject of many discussions among scholars. However, the interpretation of the GDPR as protecting mainly individual rights is just the minimum requirement for an explanation. Any explanation going further and also having the protection of collective rights in mind, will be compliant with the GDPR as long as the individual rights are also protected. Therefore, we recommend putting the individual at the centre of the explanation in a first step in order to comply with the GDPR.

With regard to the question “what should explanations look like”, we argue that XAI is more than just a technical output. To our view, XAI has to be understood as a complex communication process between human actors and cannot be merely evaluated in terms of technical accuracy. Against this backdrop, evaluating the communication process should accompany evaluating the XAI system’s technical performance. Evaluating an explanation created by an ADM system cannot be achieved without involving the user receiving the explanation. Their assessment of what a meaningful explanation needs to entail is an essential prerequisite for XAI. For domain experts, the evaluation of the explanation must include information about potentials, risks, and limitations of ADM systems; explainability starts even before the system is in use and encompasses the complete socio-technical complex.

When it comes to the target group of an explanation, public or community advocates should play a bigger role. These advocate groups support individuals confronted with an automated decision. Their interest will be more in understanding the models and their limitations as a whole instead of only focussing on the result of one individual decision.

As we will demonstrate, there is a gap between how developers and legal experts define what explanations are. Developers aim to debug statements that help them understand their models, but these are less useful for individuals who need explanations to be able to challenge a decision. Also, from a technical perspective, the term “logic involved” as it is used in the GDPR is – at best – misleading. ADM systems, and data-based systems in particular, are complex and dynamic socio-technical ecosystems. Understanding “the logic” of such diverse systems therefore requires action from different actors and at numerous stages from conception to deployment. Developers have to explain to the ADM system how to explain. Methods to explain the explanation often involve using additional approximate models with potentially lower accuracy that raise the question of whether the goal is to explain (justify) the decision or to really understand how the original model arrived at the decision.

Furthermore, transparency at the input level is a core requirement for mitigating potential bias, as post-hoc interpretations are widely perceived as being too problematic to tackle the root cause. The focus should therefore shift to making the underlying rationale, design and development process transparent—documenting the input data as part of the “logic involved”. For example, the use of datasheets can lead to more transparency by enabling expert users to better understand the overall process and translate it to lay users. Ultimately, using such measures will help improve ADM systems. In other words, the overall XAI process should involve direct and indirect stakeholders from the very beginning rather than hoping that machine learning models will be able to provide human-compatible explanations post-hoc at the end of the development chain without prohibitive effort.

Files (441.3 kB)
Name Size
What-to-explain-when-explaining-is-difficult.pdf
md5:cf32298b82cff753923efaa36906b6f4
441.3 kB Download
78
57
views
downloads
All versions This version
Views 7878
Downloads 5757
Data volume 25.2 MB25.2 MB
Unique views 6464
Unique downloads 4949

Share

Cite as