The paper examines the legal consequences of using AI in criminal justice decision-making, covering predictive policing, investigation, risk assessment, evidence analysis, sentencing and judicial decision support. It assesses these uses against due process, equality before the law and judicial independence, and discusses algorithmic bias, transparency, privacy, accountability and procedural fairness, including the possibility that risk assessment tools trained on biased data repeat past injustices. The author concludes that AI can be adopted in criminal justice without sacrificing core legal protections provided specific measures, including meaningful human oversight, are in place.
L'apprendista istruttore. L'intelligenza artificiale sperimentata nell'ambito dei controlli sulla SCIA edilizia
The article draws on EDIL.I@, a multi-agent AI system designed to support the review of SCIA building filings (certified notices of the start of building works) in two pilot Italian municipalities. Using the project as a case, the author examines whether technical building rules can be translated into computable form, whether the checks are mandatory, how administrative powers are allocated, the duty to give reasons, liability and remedies. The author concludes by proposing that jurists take part in the design of such systems rather than only reviewing them afterwards.
L'adozione di strumenti di IA potrebbe agevolare la redazione di pubblicazioni economiche come il Bollettino economico?
Banca d'Italia · · Italian title and abstract; paper text in English
The paper asks whether AI and language models could support the drafting of economic publications, specifically the Bank of Italy's quarterly Economic Bulletin. The author proposes an AI-based methodology that draws on earlier editions of the Bulletin and economic data available online, designed to limit the risk of statements unsupported by data or economic logic. The author reports that the approach can link AI-generated statements to their sources with high reliability and produce text grounded in empirical evidence and in the economic reasoning of previous editions, and states that specialists remain essential to prevent errors in AI-generated content.
Absorptive Capacity, AI-Based Decision Support, and Administrative Decision-Making Effectiveness in Public Sector Organizations: A Systematic Review and Integrative Framework for the Saudi Context
The paper asks whether public agencies' absorptive capacity is associated with adoption of AI-based decision support systems and with administrative decision-making effectiveness, and whether such systems carry that effect. It reports a systematic review following PRISMA 2020 using one database (OpenAlex), in which 61 studies were included from 8,260 records; no included study tested the full chain, only five measured absorptive capacity directly, evidence linking AI or analytics use to decision outcomes was mixed, and only three studies were from Saudi Arabia. The author proposes a conditional framework with testable propositions for the Saudi context and lists limitations including the single database, unavailable full texts and the absence of a second reviewer.
Infinite Potential: Insights On Artificial Intelligence Loss of Control Across Five Scenarios
The report synthesises observations from 26 Infinite Potential games featuring AI loss-of-control dynamics, conducted between June 2025 and February 2026 across five scenarios (Business and Civil Society, Cyber Surprise, European Energy, Geopolitics of Energy on the Korean Peninsula, and Remote Control). The authors report that governments may struggle to recognise and interpret loss-of-control events in time to respond, may lack practical ways to stop or contain AI systems operating outside human control, may need new communication and coordination channels among governments and operational actors, and may struggle to maintain continuity, public trust and social resilience. Recommendations address detection and crisis recognition, containment (including isolation and shutdown capabilities and emergency authorities), intergovernmental coordination and public-private cooperation, fallback mechanisms that work without AI systems, and decision support on accountability, deterrence and escalation.
Minaccia Cyber nell'era dell'A.I. generativa. I possibili strumenti di intervento normativi nel sistema del libero mercato
The article examines the regulatory response to cyber risks arising from applied and generative AI in markets already affected by unlawful interference, with reference to recent Italian legislation adapting national law to European rules and to security requirements against attacks by criminal actors. It analyses current legislation, including the NIS 1 and NIS 2 Directives, the EU AI Act and Italian Law 90/2024, together with proposals for future intervention. The author seeks to identify the weaknesses and merits of a regulatory architecture intended to mitigate the operational risk of the actors involved, and concludes that cybersecurity and AI must develop together, with regulatory frameworks acting as a necessary oversight mechanism.
The State of AI Global Governance and Its Implications for the U.S.-China Summit
The paper that brings geopolitics back into the frame. It maps the global governance landscape through the lens of great-power competition, and is candid about why China's offer resonates with middle powers: a governance model that supports state authority and does not condition cooperation on political or human rights commitments.
Public procurement and AI governance: developing a strategic framework for middle powers
The report examines how countries without major domestic AI developers ("AI middle powers") can use government purchasing power as a governance mechanism. It identifies three principal risks for state buyers: vendor lock-in, limited visibility into AI development, and data security, and proposes a framework of contractual obligations for governments to negotiate with AI providers. It concludes that procurement alone has limits, but that coordinated purchasing standards among middle powers could strengthen their bargaining position and their capacity to manage dependence on foreign frontier AI firms.
Before agentic AI scales in government: the democratic authorization gap
The authors introduce the concept of a "democratic authorization gap", defined as a break in the chain linking legal and democratic authority to actions carried out by AI agents in government. They identify four mechanisms through which the gap may arise: mandate translation, recursive delegation, action diffusion and contestability lag. The paper proposes five governance conditions for pilot programmes of agentic AI in public institutions.
Artificial intelligence and public policy: a systematic review with bibliometric insights
The review analyses 257 Scopus-indexed articles published between 1995 and 2025 on AI applications in public policy, using PRISMA methodology and bibliometric analysis. It identifies six thematic areas: AI-enabled governance across policy levels; regulatory and governance frameworks; ethical challenges of bias, accountability and inclusivity; socioeconomic effects on welfare systems and labour markets; public trust and acceptance of AI-driven decisions; and AI in crisis management and policy evaluation. The authors conclude that government use of AI requires ethical safeguards, accountability mechanisms and inclusive governance alongside technical efficiency.
Care and consideration: lessons for AI adoption in the public sector from a research collaboration with NICE
The report examines the conditions for responsible AI adoption in public-sector internal processes through a collaboration between the Ada Lovelace Institute and the National Institute for Health and Care Excellence (NICE). Methods included interviews with 11 NICE staff, analysis of workshops, evaluation of a feasibility-testing protocol for LLM use in literature screening, and consultation with 22 experts. The authors identify risks including vendor dependency and evidence quality, and conclude that adoption requires sociotechnical evaluation covering workforce effects, stakeholder engagement and alignment with public-sector values, not technical performance alone.
The contractual vacuum: AI procurement and the erosion of good administration
The article examines how the adoption of AI-as-a-service by governments transfers decision-making functions to private vendors while limiting public bodies' access to the operational data needed to verify or challenge those decisions. It reviews European and common law procurement rules against the principle of good administration. The author proposes "computational legal design", a framework that translates legal requirements into technical safeguards through audit mechanisms and a cost model for administrative remedies.
Characterizing agentic flooding of government services
The paper asks whether use of AI agents is producing surges in demand for government services and how governments respond. The authors compiled 84 documented cases across 11 jurisdictions and 13 service domains in which officials or credible sources attributed demand surges to AI, coded with an LLM-assisted workflow verified by humans. They report that the services most exposed are financially consequential ones with complex requirements previously held down by friction, that operational collapse appears unlikely, and that governments typically respond by adding friction, which reduces accessibility; they suggest capacity-building and digital identity integration as alternatives.
Can LLMs reason in a legally meaningful manner? A small-scale study on European Court of Human Rights cases
The study examines whether a large language model produces legally meaningful reasoning when forecasting European Court of Human Rights judgments on Article 10 (freedom of expression). The authors tested GPT-5.4 with three prompting strategies on 30 recent cases, with evaluation by law students and by other large language models acting as judges. They report reasoning that is structurally complete but substantively shallow, that expert-curated prompts produce fuller analysis without improving prediction accuracy, and that LLM evaluators are internally consistent but only weakly aligned with human assessment.
Generative AI experimentation in government: Learning from emerging guidelines
An empirical look at how public institutions are actually approaching generative AI: piloting, constraining, adapting internal guidance, and learning in real time. Its sharpest finding: guidance is everywhere, but it is fragmented, and very few governments are systematically measuring whether their experiments work.
Closing the Adoption Gap: How AI Is Being Built in the Global South
The necessary complement to the compute-concentration story. Across 53 case studies, compute was a major obstacle in just 13% of cases; data in 79%. Its sharpest claim is that the compute debate uses the wrong unit: most Global South deployments adapt and run models rather than train them. And it names a quieter bias: even well-built local models get misjudged by benchmarks translated from English.
One of the clearest comparative snapshots of where AI governance frameworks actually stand across 135 countries and jurisdictions. What it shows is that the gap is no longer simply whether countries have frameworks on paper, but whether those frameworks carry binding force, institutional follow-through, and enforcement capacity.
Preliminary Report of the Independent International Scientific Panel on AI: Evidence-based assessment of opportunities, risks and impacts of AI
The first global, independent scientific assessment designed to inform intergovernmental AI governance. Its significance is not just substantive, but institutional: this is the document intended to give the Global Dialogue on AI Governance a common evidence base. It has also brought two issues to the centre of the debate – the difficulty of securing reliable control over advanced AI agents, and the extreme concentration of compute capacity.
The narrative they form: the centre of gravity has shifted from principle-setting to asymmetries of implementation, compute, and institutional power. The next phase of AI governance is not just about defining rules – it's about who has the capacity to make them hold.
Suggest a research resource
Share a paper, report or other material that could help people understand how AI is being adopted in government and public services.