Publications

The Ada Lovelace Institute (Ada), AI Now Institute (AI Now), and Open Government Partnership (OGP) have partnered to launch the first global study to analyse the initial wave of algorithmic accountability policy for the public sector. As governments are increasingly turning to algorithms to support decision-making for public services, there is growing evidence that suggests that these systems can cause harm and frequently lack transparency in their implementation. Reformers in and outside of government are turning to regulatory and policy tools, hoping to ensure algorithmic accountability across countries and contexts. These responses are emergent and shifting rapidly, and they vary widely in form and substance – from legally binding commitments, to high-level principles and voluntary guidelines. This report presents evidence on the use of algorithmic accountability policies in different contexts from the perspective of those implementing these tools, and explores the limits of legal and policy mechanisms in ensuring safe and accountable algorithmic systems.

During the past decade, a small but rapidly growing number of Law&Tech scholars have been applying algorithmic methods in their legal research. This Article does it too, for the sake of saving disclosure regulation failure: a normative strategy that has long been considered dead by legal scholars, but conspicuously abused by rule-makers. Existing proposals to revive disclosure duties, however, either focus on the industry policies (e.g. seeking to reduce consumers’ costs of reading) or on rulemaking (e.g. by simplifying linguistic intricacies). But failure may well depend on both. Therefore, this Article develops a `comprehensive approach', suggesting to use computational tools to cope with linguistic and behavioral failures at both the enactment and implementation phases of disclosure duties, thus filling a void in the Law & Tech scholarship. Specifically, it outlines how algorithmic tools can be used in a holistic manner to address the many failures of disclosures from the rulemaking in parliament to consumer screens. It suggests a multi-layered design where lawmakers deploy three tools in order to produce optimal disclosure rules: machine learning, natural language processing, and behavioral experimentation through regulatory sandboxes. To clarify how and why these tasks should be performed, disclosures in the contexts of online contract terms and privacy online are taken as examples. Because algorithmic rulemaking is frequently met with well-justified skepticism, problems of its compatibility with legitimacy, efficacy and proportionality are also discussed.

This article argues that the poor design and implementation of experimental regulations and regulatory sandboxes can have both methodological and legal implications. First, the internal validity of experimental legal regimes is limited because it is unclear whether the verified positive or negative results are the direct result of the experimental intervention or other circumstances. The limited external validity of experimental legal regimes impedes the generalization of the experiment and thus the ability to draw broader conclusions for the regulatory process. Second, experimental legal regimes that are not scientifically sound make a limited contribution to the advancement of evidence-based lawmaking and the rationalization of regulation. Third, methodological deficiencies may result in the violation of legal principles (e.g., legality, legal certainty, equal treatment, proportionality) which require that experimental regulations follow objective, transparent, and predictable standards. This article contributes to existing comparative public law and law and methods literature with an interdisciplinary framework which can help improve the design of experimental regulations and regulatory sandboxes. This article starts with an analysis of the central features, functions, and legal framework of these experimental legal regimes. It does so by focusing on legal scholarship, policy reports, and case law on experimental regulations and regulatory sandboxes from France, United Kingdom, and The Netherlands. While this article is not strictly comparative in its methodology, the three selected jurisdictions illustrate well the different facets of experimental legal regimes. This article draws on social science literature on the methods of field experiments to offer novel methodological insights for a more transparent and objective design of experimental regulations and regulatory sandboxes.

Imagine that two doctors in the same city give different diagnoses to identical patients – or that two judges in the same court give different sentences to people who have committed matching crimes. Now imagine that the same doctor and the same judge make different decisions depending on whether it is morning or afternoon, or Monday rather than Wednesday, or they haven’t yet had lunch. These are examples of noise: variability in judgements that should be identical. In Noise, Daniel Kahneman, Olivier Sibony and Cass R. Sunstein show how noise produces errors in many fields, including in medicine, law, public health, economic forecasting, forensic science, child protection, creative strategy, performance review and hiring. And although noise can be found wherever people are making judgements and decisions, individuals and organizations alike commonly ignore its impact, at great cost. Packed with new ideas, and drawing on the same kind of sharp analysis and breadth of case study that made Thinking, Fast and Slow and Nudge international bestsellers, Noise explains how and why humans are so susceptible to noise and bias in decision-making. We all make bad judgements more than we think. With a few simple remedies, this groundbreaking book explores what we can do to make better ones.

As intuitive statisticians, human beings suffer from identifiable biases, cognitive and otherwise. Human beings can also be “noisy,” in the sense that their judgments show unwanted variability. As a result, public institutions, including those that consist of administrative prosecutors and adjudicators, can be biased, noisy, or both. Both bias and noise produce errors. Algorithms eliminate noise, and that is important; to the extent that they do so, they prevent unequal treatment and reduce errors. In addition, algorithms do not use mental short-cuts; they rely on statistical predictors, which means that they can counteract or even eliminate cognitive biases. At the same time, the use of algorithms, by administrative agencies, raises many legitimate questions and doubts. Among other things, they can encode or perpetuate discrimination, perhaps because their inputs are based on discrimination, perhaps because what they are asked to predict is infected by discrimination. But if the goal is to eliminate discrimination, properly constructed algorithms nonetheless have a great deal of promise for administrative agencies.

The Recommendation aims at further strengthening regulatory governance by helping to update and enhance relevant instruments, processes, and institutions. More precisely, it seeks to provide a conceptual framework and relevant guidance for using and adapting regulatory policy and governance in the face of the regulatory challenges and opportunities arising from innovation.

Artificial Intelligence (AI) is a peculiar case of General Purpose Technology that differs from other examples in history because it embeds specific uncertainties or ambiguous character that may lead to a number of risks when used to support transformative solutions in the public sector. AI has extremely powerful and, in many cases, disruptive effects on the internal management, decision-making and service provision processes of public administration. Over the past few years, the European Union and its Member States have designed regulatory policies and initiatives to mitigate the AI risks and make its opportunities a reality for national, regional and local government institutions. ‘AI Watch’ is one of these initiatives which has, among its goals, the monitoring of European Union’s industrial, technological, and research capacity in AI and the development of an analytical framework of the impact potential of AI in the public sector. This report, in particular, follows a previous landscaping study and collection of European cases, which was delivered in 2020. This document first introduces the concept of AI appropriation in government, seen as a sequence of two logically distinct phases, respectively named adoption and implementation of related technologies in public services and processes. Then, it analyses the situation of AI governance in the US and China and contrasts it to an emerging, truly European model, rooted in a systemic vision and with an emphasis on the revitalised role of the member states in the EU integration process, Next, it points out some critical challenges to AI implementation in the EU public sector, including: the generation of a critical mass of public investments, the availability of widely shared and suitable datasets, the improvement of AI literacy and skills in the involved staff, and the threats associated with the legitimacy of decisions taken by AI algorithms alone. Finally, it draws a set of common actions for EU decision-makers willing to undertake the systemic approach to AI governance through a more advanced equilibrium between AI promotion and regulation. The three main recommendations of this work include a more robust integration of AI with data policies, facing the issue of so-called “explainability of AI” (XAI), and broadening the current perspectives of both Pre-Commercial Procurement (PCP) and Public Procurement of Innovation (PPI) at the service of smart AI purchasing by the EU public administration. These recommendations will represent the baseline for a generic implementation roadmap for enhancing the use and impact of AI in the European public sector.