The European General Data Protection Regulation (GDPR), which governs the processing of personal data in all EU Member States, contains an exemption for “purely personal or household activities”. Whether this so-called household exemption covers the setup of an online communication forum, particularly on a social media or chat platform, is a question of the individual case. We present a case-based reasoning approach to automatically assessing a scenario provided by the user and generating a tailored legal recommendation.
up to 4 per cent of the total worldwide annual turnover of the preceding financial year, whichever is higher.
The material scope of application of the GDPR, however, is not restricted to enterprises and organizations: Virtually every individual handles personal data of others incessantly and can therefore be subject to data protection law. Seemingly innocuous examples of data processing include private blogging, the use of social media accounts, taking and sharing pictures of others and a variety of similarly common activities.
Legal uncertainty entails negative consequences beyond mere lack of compliance, specifically so-called “chilling efects” 5: Individuals may be deterred from socially desirable activities even though these could have been implemented in conformity with data protection laws. The need for high-quality, low-threshold legal expertise is therefore a worthy area of application for artificially intelligent systems.
In the interdisciplinary INWEND6 project at the University of Trier, we are developing a software prototype that can elicit relevant factual information from non-lawyers and automatically generate a (preliminary) legal assessment using case-based reasoning (CBR) techniques. The focus of this project is the so-called household exemption in Article 2(2)(c) GDPR which excludes from the scope of the regulation the processing of personal data “by a natural person in the course of a purely personal or household activity”.
This project has a pilot character for the research at the intersection of artiifcial intelligence and European law. In an interdisciplinary cooperation between legal experts and computer scientists the understanding of the respective other discipline is expected to be improved, which should also enable the expansion of research activities to other common interdisciplinary fields of research. The scientific relevance of the development of a structured legal case base which facilitates automated legal reasoning lies in the possibility to work with consistent and structured case data in the future in order to develop artificially intelligent systems.
The remainder of the paper is organized as follows. In section 2 we outline related work on the use of CBR systems in a legal context as well as recent but distinct research in legal information technology. In section 3 we provide an overview of the relevant legal domain and the scope of the prototype (in the interest of brevity called INWEND hereinafter). In section 4 we proceed to describe our approach to modelling this knowledge domain by using an initial case situation and approximately 200 case variations generated using a set of 5 The term is defined by the Oxford Dictionary as “a discouraging or deterring efect on the behaviour of an individual or group, especially the inhibition of the exercise of a constitutional right, such as freedom of speech, through fear of legal action.”, https://www.lexico.com/definition/chilling_efect. 6 INWEND stands for “Intelligente Wissensbasierte Entscheidungsunterstützung für juristische Fragestellungen am Beispiel des Datenschutzrechtes”, which translates to “Intelligent knowledge-based decision support for legal questions on the example of the data protection law”. parameters. We conclude with some remarks on the feasibility of our approach in civil law jurisdictions. 2
The automation of legal tasks in the broadest sense is currently a topic of intense
economic and scientific interest, with a variety of projects and products being
discussed under the broad umbrella term of “Legal Tech” [
The general suitability of CBR techniques for handling legal tasks was discovered early, partly because the CBR approach exhibits structural similarities to legal reasoning that are particularly apparent in the common law with its doctrine of precedent and the principle of stare decisis7. We will first outline some pioneering research work conducted in the 1980s and 1990s and then contrast the INWEND project to this and other contemporary research with a diferent outlook. 2.1
The possibility to associate the CBR approach with legal reasoning and
argumentation was discovered in as early as the 1980s [
More recently, the ARGUMENTUM [
The INWEND project takes a diferent approach, setting aside argument mining and starting from highly structured legal case information provided by domain experts. 3
We begin our account with a brief introduction to the legal domain of our projected CBR system, in order to both provide some context for our development strategy and highlight the practical need for the solution we propose. 3.1
According to Article 2(1) of the GDPR, the regulation applies, inter alia, to the
“processing of personal data wholly or partly by automated means”, comprising
practically any use of data processing systems [
This clause excludes from the scope of application the processing of personal
data “by a natural person in the course of a purely personal or household
activity”. This means that any such activity may be carried out without having
to observe any of the obligations under the GDPR. These comprise a variety of
legal duties, such as the principles relating to the processing of personal data set
forth in Article 5, the need to justify the processing according to Article 6 and
to obey the rights of the data subject set forth in Articles 12 et seqq. Being an
exception to the entire data protection regime, there is general agreement that
the clause must be interpreted restrictively [
A general problem of legal interpretation, including the construction of
statutory law, is the open-textured nature of legal concepts [
The INWEND system is designed to address the need for an automated legal assessment of the question: Does a natural person opening a communication forum on an online platform comply with the household exemption? The practical implication of this system is that a user may enter the circumstances of their 8 Paal/Pauly-Ernst, Art. 2 GDPR margin number 5. 9 Kühling/Buchner-Kühling/Raab, Art. 2 GDPR margin number 21. (intended) use of such a forum and be immediately provided with (preliminary) legal feedback. The availability of such a tool may, as a secondary efect, incentivise platform providers to make available privacy-friendly options to attract more users. In the following section we will describe our strategy for modelling this legal problem and the first results of our research. 4
In case-based reasoning the task of solving problems is based on previous
experience, which is stored in the form of cases in a case base [
The CBR approach relies on the existence of a case base, from which relevant reference experiences can be retrieved. With the GDPR being a novel legal framework, the availability of pertinent court decisions – particularly by the authoritative European Court of Justice (ECJ) – is low. Thus, there is little documented experience available that could be turned into a case base.
We therefore took a diferent approach in acquiring knowledge from the knowledge domain: Based on an interpretation of the household exemption, we developed an exemplary initial case situation. This is a short factual description of the scenario in which the question concerning the applicability of the household exemption arises. In particular, we consider the situation in which a user wants to open a forum on a communication platform and invite others to join the discussion.
This allowed us to identify a number of parameters which pertain to the applicability of the household exemption: Who is the intended circle of participants (such as family members, close friends, acquaintances), and how many people will be accessing the forum? How much personal information of other participants in the discussion will the user gain? Are there technical asymmetries by which the user will have more insight into other participants’ personal data than vice versa?
Since legal ontologies are commonly used in legal informatics as a formal
knowledge model [
With the help of domain experts we created a documentation of the selected parameters and their features, providing strategies for the assessment of problematic cases as well as sets of positive and negative examples. In the next step, we used these parameters to synthetically generate a case base of approximately 200 cases. Since not all combinations of the parameter features were meaningful from a legal standpoint, the case base was slightly smaller than the product of all features: With five parameters, each of which ranging between two and ifve features, a total of 240 combinations would have resulted; since 48 of these combinations involved irreconcilable case details, there were only 192 legally meaningful cases.
We then assessed the applicability of the household exemption in the cases of the case base. It is worth noting that, rather than judging cases on an individual basis, legal domain experts employ two strategies to determine the outcome of cases by groups: Firstly, they try to identify so-called “edge cases”, these being parts of the case base where the legal assessment reaches a “tipping point” between two contrary results. Secondly, they extensively employ reasoning a fortiori10, arguing for instance that a case cannot qualify for the household exemption where a comparable case with a more restricted circle of participants has already been rejected: Since a wider circle of participants is an argument against the applicability of the household exemption, such a decision would be inconsistent.
The consistency of a legal case base is a notion which was discussed earlier in
the context of the common law, specifically with regard to the doctrine of
precedent [
Drawing on the aforementioned domain knowledge, we then implemented the
case representation using the CBR system ProCAKE [
Considering the argumentation pattern mentioned in 4.2, whereby the applicability of the household exemption can be determined by a fortiori reasoning, we chose to model the parameter circle of participants as a polyvalent attribute. Cases that difer merely with regard to this parameter can be summarized as one case by representing the values of the respective cases for this parameter in a set. This step reduces the case base by 30 cases. At the time of writing, further modifications of the model are still in progress.
Based on the case model and the case base, a first similarity model was created. For calculating the global similarity the individual parameters are weighted by their impact on the assessment of the case. For example, the aforementioned parameter circle of participants has a particularly strong bearing on the outcome of the assessment.
The similarity for the parameter circle of participants is computed with a set mapping. If the value of the query matches a value in the set of a case of the case base, the similarity is high, otherwise low. The other local similarities are implemented with simple similarity measures, with equal values being assigned a high similarity score and diferent values being assigned a low similarity score. Again, at the time of writing, further modifications of the model are in progress. 4.4
We developed a prototypical Graphical User Interface (an excerpt is shown in Figure 1), to visualize the intended interaction with the program. The user has to respond to questions by choosing the answer most fitting to the situation at hand. As mentioned under 4.3, the questions also include the option to abstain from answering any individual question.
Each question corresponds to an attribute in our case model, with the answers determining their possible values. As can be seen in Figure 1, we guide the user through the interaction by providing additional information on questions and answers as needed. For additional reference, the factual situation described by the selected parameters is summarized for the user, which can be seen in Figure 2. Specifically, the text seen underneath the headline “Sachverhalt” (facts of the case) is a summary of the factual situation automatically generated from the answers given by the user; the text underneath the headline “Einschätzung” (legal opinion) is an easily comprehensible assessment of this factual situation, pertaining to the plain result – whether the household exemption is applicable or not – as well as its legal certainty.
On the basis of the answers selected by the user a query case is created. When the user clicks the submit button, this query is transmitted to the back end as a JSON object and then relayed as a query to a ProCake server, where it is used for a retrieval on the case base. The retrieval result containing the most similar case is then returned to the web interface. The outcome of this retrieved case is displayed on the GUI (see Figure 2) and suggested to the user as the closest match to their query.
In summary, this first version of a GUI generates a legal assessment in the context of the household exemption for a user case, based on the presented parameters, case base and similarity model.
This paper presents a first CBR approach to modelling the legal assessment of a factual scenario in the context of the GDPR household exemption. The prototype system developed in the INWEND project is designed to elicit relevant facts from the user and to automatically generate a tailored legal recommendation. In the current stage of development, we have generated a substantial case base to reflect the structure of the GDPR household exemption. The system is controlled via a web interface and is able to retrieve reference cases from its case base.
In a next step, we plan to refine our ontology which represents our concept of the case representation. The ontology serves as a translation help between users and domain experts.
We judge the scalability of this approach to a wider field of law as rather dificult, because the expert knowledge required is expensive and the acquisition efort is high. Thus, a further interesting task to work on is to generate new cases semi-automatically: An algorithm suggests a scenario based on the case parameters and an expert decides the applicability of the household exemption. This algorithm should determine the potential edge cases and disregard cases that do not improve the ability to decide a case.
Further work needs to be undertaken to identify the most relevant cases in the current case base and to adapt the domain and similarity model. We would also like to evaluate our approach with test users in order to verify its reliability and practical utility.
We conclude that CBR systems are generally capable of providing initial legal assessments for non-lawyers in civil and EU law jurisdictions. We could imagine combining our approach with text mining strategies to extract case features from a plain text entered by the user instead of generating the query from a user’s questions and answers.
Acknowledgements. This work is funded by the German Federal Ministry of Education and Research (BMBF) under grant number 01UG1920.