Op-Ed: “The EU Artificial Intelligence Act and Access to Justice” by Melanie Fink
On 21 April 2021, the European Commission presented its long-awaited proposal for a Regulation laying down harmonised rules on artificial intelligence (AI) – the Artificial Intelligence Act. The proposed Regulation distinguishes AI systems according to the risk they pose to the fundamental rights of individuals or EU values. Those that present an unacceptable risk are prohibited, high-risk AI systems have to conform with a long list of obligations before and after they are put on the market, limited-risk AI systems are subject to transparency obligations, and minimal-to-no risk AI systems may be freely used. Overall, the proposal has been welcomed, with some common points of criticism being the limits and exceptions to the prohibition of certain AI systems, gaps in what is qualified as ‘high-risk’, and the hesitant approach to combatting algorithmic bias.
In this piece, I argue that the use of AI systems by the public administration raises specific challenges that should be addressed in the proposed Regulation. The exercise of state power, such as law enforcement or adjudication, brings particular fundamental rights risks. But for that reason, it is also subject to stronger safeguards against abuse of that power—transparency, accountability, oversight. However, as AI technologies become increasingly embedded in public bodies’ day-to-day decision-making, the possibilities for individuals to rely on these safeguards and meaningfully challenge decisions that affect them diminish. To fully guarantee individuals’ right to access to justice in the AI context, we need, first, more clarity on the benchmarks for AI-supported decision-making to comply with the right to a reasoned decision and, second, additional mechanisms for individuals to invoke their rights before an independent body.
The Right to a Reasoned Decision in the AI Context
Under EU law, public authorities are required to give reasons for their legal acts and decisions and communicate them on their own initiative. This duty finds its legal basis in Article 296 TFEU and Article 41(2c) of the Charter of Fundamental Rights of the EU. The latter is considered to entail an individual right to a reasoned decision, a breach of which may entitle a person to be compensated for the damage suffered. According to the Court of Justice (Elf Aquitaine, C‑521/09 P), the statement of reasons must be sufficiently clear and unequivocal so as to permit the Court to review legality, but also to provide the persons concerned with sufficient information to know whether the decision may be vitiated by an error and enable them to challenge its validity. The duty to give reasons is thus not only a transparency obligation in its own right, but meant to facilitate accountability and individual access to justice.
The right to a reasoned decision can be affected in two ways when public authorities rely on AI systems in their decision-making. First, there is an inherent tension between the duty of the administration to justify its decisions and the limited explainability of some AI systems. The process that translates input into output can be so complex or opaque that humans, even those who designed the system, are not able to understand what variables exactly determined the outcome. This is often referred to as the ‘black box’ problem and limits the ability of the AI system’s user to justify AI-enabled decisions. Second, there is the problem of ‘automation bias’. This refers to the phenomenon that humans tend to ascribe a certain authority to outcome suggested by an algorithm that leads them to neglect other available information or counter-indications. Even where an authority thus can give a justification, it may in substance boil down to: ‘because the machine said so’.
The proposed Regulation aims to address these issues through specific transparency and human oversight obligations. In relation to the problem of explainability, Article 13 specifies that high-risk AI systems shall be developed and designed to be sufficiently transparent to ensure the user’s ability to interpret and use the system’s output. However, it does not entail an obligation on the part of the user to communicate that information to persons subject to the AI-supported decision. The only transparency obligation vis-à-vis these persons is stipulated in Article 52, but limited to the duty to inform them about the fact that an AI system is used.
The proposed Regulation therefore does not include obligations of AI users to explain or justify the decisions they reach towards those affected by them, even less a corresponding right on the part of individuals to demand that. While individuals can rely on the general right to a reasoned decision under Article 41(2c) of the Charter to fill this gap, the specific challenges its application raises when public bodies rely on AI systems in their decision-making justify additional safeguards. First, to avoid any doubt, the applicability of this right in the AI context should be made explicit. Second, the benchmarks used to assess compliance with the right to a reasoned decision in the AI context should be clarified by answering two related questions. What does the right to a reasoned decision actually require in terms of the nature and depth of the communication of reasons by the public authority that relied on an AI system? And what does that in turn require from the AI system’s design: transparency, interpretability, explainability, contestability? Given these aspects are central to an individual’s possibility to challenge AI-supported decisions, they should not be left to be worked out through litigation.
In relation to the problem of ‘automation bias’, Article 14 of the proposed Regulation requires that human oversight is to be ensured in such a way to enable the person assigned that task to be able to correctly interpret output and be aware of the potential of ‘automation bias’. The explicit recognition of this problem is valuable in itself. Yet, combatting it more effectively might necessitate additional safeguards, for instance by requiring the public authority that relies on AI systems for their decision-making to communicate how other available information or alternative outcomes were considered in reaching a decision.
Access to Justice Through Individual Complaints Mechanisms
Article 47 of the Charter requires that persons whose rights and freedoms guaranteed by EU law are violated have a right to an effective remedy. Even though this right ultimately demands access to a tribunal, it does not exclude the possibility to set up additional individual complaints mechanisms that are complementary to the existing judicial avenues. Examples of such mechanisms exist in particular in technically complex or fundamental rights-sensitive areas, such as the possibility to challenge decisions of the European Chemicals Agency under Article 92 REACH Regulation; to lodge fundamental rights complaints against Frontex’s activities under Article 111 EBCG Regulation; or to lodge complaints before the European Data Protection Supervisor under Article 57 GDPR. Article 56 of the proposed Regulation does establish a European Artificial Intelligence Board and requires Member States to designate national supervisory authorities, but there is no individual complaints-possibility.
In the absence of specific mechanisms to challenge a public body’s AI-enabled decisions, persons affected have to make use of the avenues available in the EU’s general remedies system. The EU’s remedies system is based on a distribution of jurisdiction between EU and national courts and heavily relies on mechanisms provided at national level. However, where conduct of EU bodies is concerned, EU courts are exclusively competent to hear complaints. Since there is no specific fundamental rights complaints procedure, the two most important avenues for individual applicants who wish to challenge EU conduct are the action for annulment (Article 263 TFEU) and the action for damages (Articles 268 and 340 TFEU). The former is notorious for the strict conditions under which individuals are allowed as applicants, and the latter for the high threshold required for success on the merits.
Both actions also set out limits that may raise particular difficulties when public authorities rely on AI systems. In the context of the action for annulment, the EU Courts’ judicial review is limited – in areas where EU bodies enjoy a wide margin of discretion – to examining whether the contested act contains a manifest error of assessment. In the context of the action for damages, the Court of Justice has consistently held (Laboratoires Pharmaceutiques Bergaderm, C-352/98 P) that liability only arises for breaches that are sufficiently serious, meaning that the authorities in question ‘manifestly and gravely disregard the limits on their discretion’. The key question for these requirements of flagrancy or inexcusability is how the choice to follow (or not) an AI system’s recommendation would affect the assessment of the reprehensibility of the authority’s error. At least in liability law, the Court of Justice has held (here and here) that reasonably relying on the assessment of another authority is a factor that may exclude liability. Even though in those cases this was the Commission, not an AI system, the underlying idea that authorities may trust certain sources of information without double-checking may equally apply to the AI context. The answer to this question will have a substantial impact on the chances of success of affected individuals before the EU Courts.
For all its benefits in terms of speed and efficiency, the increasing use of AI systems in the public administration’s day-to-day decision-making also brings a number of challenges. It may reinforce biases, compound the problem of ‘many hands’ in allocating responsibility, and disrupt models of transparency and accountability. This last aspect has a major impact on individual access to justice. When the reasons why a certain decision was taken are not sufficiently clear, this affects the possibilities of individuals to bring arguments against it.
To meet this challenge, we can rely on established rights under EU law – the right to a reasoned decision and the right to effective judicial protection – and adapt them to the AI context. This involves, on the one hand, developing benchmarks to assess compliance of public authorities that use AI systems in their decision-making with the obligation to give reasons. On the other hand, it means creating mechanisms for individuals to invoke their rights before an independent body.
Melanie Fink is Assistant Professor at Leiden University.