On 11 February 2025, the European Commission quietly withdrew the proposed Artificial Intelligence Liability Directive (AILD) from its legislative programme, citing a "lack of foreseeable agreement" among member states. The directive had been designed to bridge the "liability gap" — the structural difficulty that individuals face when attempting to prove fault and causation in cases involving opaque AI systems. Its withdrawal left a void that no subsequent instrument has filled.
This data brief maps the accountability deficit that has emerged in the wake of that withdrawal. It examines three domains where AI systems are making consequential decisions about individuals' lives — predictive policing and risk scoring, judicial decision-making, and automated administrative decisions — and assesses the legal, technical, and institutional mechanisms available to those who seek to contest those decisions. The picture that emerges is one of structural asymmetry: the systems making decisions are increasingly sophisticated; the mechanisms for challenging them are increasingly inadequate.
The Liability Landscape: Key Data Points
The Regulatory Void
- February 2025: EU AI Liability Directive withdrawn. No replacement instrument proposed as of September 2026.
- August 2026: EU AI Act reaches full applicability for most provisions. High-risk AI applications in employment, education, and law enforcement deferred to December 2027.
- 22 of 27 EU member states had fully transposed the NIS2 Directive by mid-2026; five — including France, Ireland, and the Netherlands — remained in the legislative process.
- 0 horizontal EU instruments currently govern non-contractual civil liability for AI-related harm. Claimants must rely on the revised Product Liability Directive, the EU AI Act's transparency obligations, and national tort law — a fragmented patchwork that produces divergent outcomes across jurisdictions.
Government AI Deployment
- 57% of observed government AI deployments focus on automating and streamlining services (OECD, 2025).
- 45% aim at enhancing decision-making, forecasting, and "sense-making."
- 30% are specifically designed to improve accountability and anomaly detection.
- In October 2025, the Democracy Forward Foundation filed a complaint in the US District Court for the District of Columbia against federal agencies including OPM, GSA, and HUD for failing to respond to FOIA requests regarding the use of AI in federal rulemaking.
The withdrawal of the EU AI Liability Directive in February 2025 left a structural void: no horizontal EU law governs non-contractual civil liability for AI-related harm, forcing claimants into a fragmented patchwork of national tort systems.
Domain 1: Predictive Policing and Risk Scores
The Scale of Deployment
The withdrawal of the EU AI Liability Directive in February 2025 left a structural void: no horizontal EU law governs non-contractual civil liability for AI-related harm, forcing claimants into a fragmented patchwork of national tort systems.
Predictive policing tools — systems that use historical crime data, demographic information, and algorithmic models to forecast where crimes are likely to occur or which individuals are likely to offend — have been deployed by law enforcement agencies across North America, Europe, and Asia for over a decade. The tools range from place-based systems that identify "hot spots" for patrol allocation to person-based systems that generate individual risk scores used in bail, sentencing, and parole decisions.
The regulatory response has diverged sharply across jurisdictions. In Europe, the EU AI Act's Article 5 includes an explicit prohibition on AI systems designed to predict the probability of an individual committing a criminal offence. This prohibition entered into force in February 2025 and applies to all member states. In the United States, no equivalent federal prohibition exists. Predictive policing tools continue to be deployed by municipal and state law enforcement agencies in what legal scholars describe as a "regulatory vacuum," where vendors introduce capabilities before policymakers have established civil liberties protections.
The Bias Problem: Documented Evidence
The constitutional and civil rights concerns surrounding predictive policing are not theoretical. They are grounded in documented evidence of systematic bias in the data on which these systems are trained.
The core problem is what researchers call "dirty data" — historical law enforcement records that reflect decades of discriminatory policing practices. When algorithms are trained on this data, they create a feedback loop: neighbourhoods historically targeted by police are flagged as "high-risk," leading to increased police presence, more stops, more arrests, and further data collection that reinforces the initial bias. The algorithm does not introduce the bias; it amplifies and legitimises it.
Research on biometric and predictive tools has consistently documented an "accuracy gap" across demographic groups. Facial recognition systems exhibit significantly higher error rates for Black, female, and elderly individuals compared to white male subjects. Gait recognition systems show similar disparities. These accuracy gaps are not incidental technical imperfections — they are structural features of systems trained on non-representative data.
The legal implications are significant. In the United States, the use of biased algorithmic outputs as grounds for police intervention raises Fourth Amendment questions about probable cause and Fourteenth Amendment questions about equal protection. Legal experts have documented instances where police treated uncorroborated AI results as definitive "100% matches," leading to wrongful arrests. The opacity of many predictive systems — which operate as "black boxes" — complicates defendants' ability to challenge the evidence used against them, potentially violating the disclosure requirements established by Brady v. Maryland.
The Automation Bias Problem
Beyond the data quality issue lies a more fundamental human factors problem: automation bias. Research documents a consistent tendency among decision-makers — including trained law enforcement officers and judges — to trust computer-generated conclusions over human judgement or contradictory evidence. This is not a failure of individual rationality; it is a predictable response to the perceived authority of algorithmic outputs, particularly when those outputs are presented with numerical precision and institutional backing.
Automation bias is particularly dangerous in high-stakes contexts where the cost of error falls asymmetrically on the individual subject to the decision. A risk score that incorrectly classifies an individual as high-risk for reoffending may result in pre-trial detention, a longer sentence, or denial of parole — consequences that are difficult to reverse even when the error is subsequently identified.
The EU AI Act prohibits AI systems designed to predict the probability of an individual committing a criminal offence — yet in the United States, predictive policing tools continue to be deployed in a regulatory vacuum, with constitutional challenges still working through the courts.
Domain 2: AI in the Courtroom
Current Deployment Patterns
AI is currently used in judicial contexts for a range of functions: summarising evidence and case documents, classifying case filings, predicting recidivism for bail and sentencing decisions, assisting in legal research, and — in some jurisdictions — generating draft judgements for routine matters. The OECD's 2025 governance report found that AI deployment in judicial contexts is growing rapidly, driven by case backlogs and resource constraints in court systems across both developed and developing economies.
The risks are well-documented. AI models trained on historical judicial decisions replicate past patterns of racial, gender, and socioeconomic discrimination, potentially institutionalising bias under a veneer of scientific legitimacy. Generative AI tools used for legal research may produce factually incorrect information — "hallucinations" — creating risks when used to draft submissions or summarise evidence. Proprietary systems that operate as "black boxes" prevent parties from understanding the basis of a decision, which is a prerequisite for meaningful legal challenge.
The Right to Contest: Procedural Requirements
The EU AI Act prohibits AI systems designed to predict the probability of an individual committing a criminal offence — yet in the United States, predictive policing tools continue to be deployed in a regulatory vacuum, with constitutional challenges still working through the courts.
The right to contest AI-generated or AI-assisted output in judicial proceedings is a central focus of emerging legal discourse. International bodies have begun to articulate the procedural requirements that must be met to protect this right.
UNESCO's 2025 guidelines on AI in courts establish fifteen principles, including safety, auditability, human oversight, and explainability. The guidelines specify that when AI output is used as evidence, it must be subject to standard rules of admissibility, disclosure, and cross-examination. Parties must be informed when AI has materially influenced a decision. Judges must retain personal responsibility for all issued rulings and must verify AI-assisted research.
The Council of Europe's ethical charter on AI in judicial systems, adopted in 2018 and updated in 2024, maintains that AI must function only as a support tool, with human judges retaining ultimate responsibility. The charter explicitly prohibits the use of AI to replace judicial discretion — the capacity of a judge to weigh the specific circumstances of a case against the general requirements of the law.
These principles are not yet uniformly implemented. In many jurisdictions, there is no requirement to disclose when AI has been used in case management or decision support. Defendants — particularly those without resources to retain technical experts — frequently lack the means to challenge algorithmic evidence. The asymmetry between the sophistication of the systems making decisions and the mechanisms available to contest them is structural, not incidental.
When AI output is used as evidence in legal proceedings, it must be subject to standard rules of admissibility, disclosure, and cross-examination — a requirement that most current deployments cannot satisfy.
Domain 3: Automated Administrative Decisions
The Scale of Automated Decision-Making
Beyond policing and the courts, AI systems are making consequential decisions about citizens' lives across a wide range of administrative contexts: benefit eligibility, tax assessments, immigration applications, planning permissions, and regulatory compliance. These decisions are frequently made at scale, with limited human review, and with inadequate mechanisms for individuals to understand or contest the basis of an adverse outcome.
The EU AI Act classifies many of these applications as high-risk, subject to mandatory transparency, documentation, and human oversight requirements. However, the full enforcement of these provisions for high-risk applications is deferred to December 2027. In the interim, the regulatory framework is incomplete, and the accountability gap is real.
The October 2025 complaint filed by the Democracy Forward Foundation against US federal agencies illustrates the problem. The complaint alleged that agencies including OPM, GSA, and HUD had failed to respond to FOIA requests regarding the use of AI in federal rulemaking — a process that, under the Administrative Procedure Act, requires reasoned decision-making and public input. The lawsuit raised the question of whether delegating rulemaking to AI systems is constitutionally and statutorily permissible — a question that has not yet been definitively resolved by any court.
The Explainability Requirement
The right to an explanation for automated decisions affecting individuals is recognised in principle by the EU's GDPR (Article 22), the EU AI Act, and a growing body of national legislation. In practice, the explainability requirement is frequently honoured in form rather than substance. Explanations provided by AI systems are often post-hoc rationalisations — outputs generated to satisfy a legal requirement — rather than genuine accounts of the causal factors that produced a particular decision.
The distinction matters. A genuine explanation of why an AI system denied a benefit application would allow the applicant to identify and correct errors in the underlying data, challenge the validity of the model's assumptions, or demonstrate that the decision was inconsistent with the system's own stated criteria. A post-hoc rationalisation provides the appearance of transparency without the substance.
The OECD's 2025 report on governing with AI emphasises that transparency, fairness, and accountability must be operationalised through specific technical and procedural requirements — not merely stated as principles. The EU AI Act and GDPR are increasingly serving as benchmarks for global digital government policies, but their implementation requires institutional capacity that many governments currently lack.
The Fragmented Liability Landscape: Five Risk Indicators
Legal experts monitoring the post-AILD landscape have identified five key indicators that will determine how the fragmented liability framework affects accountability in practice:
When AI output is used as evidence in legal proceedings, it must be subject to standard rules of admissibility, disclosure, and cross-examination — a requirement that most current deployments cannot satisfy.
- National Divergence: Whether similar AI-related harm results in different legal outcomes depending on the member state. Early evidence suggests significant divergence, with some jurisdictions applying strict liability principles and others requiring proof of fault that is practically impossible to establish for opaque AI systems.
- First-Wave Litigation: How courts integrate EU AI Act compliance obligations — logging, transparency, documentation requirements — into existing private law categories of negligence or fault. The first wave of AI liability cases is expected to reach national courts in 2026–2027.
- CJEU Judicial Harmonisation: The role of the Court of Justice of the European Union in providing preliminary rulings to harmonise the interpretation of AI Act obligations within civil proceedings. Without CJEU guidance, national courts will develop divergent interpretations.
- Insurance Market Differentiation: Whether insurance markets begin pricing risk differently based on jurisdiction and regulatory environment. Early signals suggest that insurers are beginning to factor AI liability exposure into premiums for organisations deploying high-risk AI systems.
- Jurisdictional Arbitrage: Whether the fragmented liability landscape influences where companies choose to deploy AI or how they structure their legal liabilities. The risk of regulatory arbitrage — deploying AI systems in jurisdictions with weaker accountability frameworks — is real and growing.
What Accountability Requires
The data brief's findings point to a set of structural requirements for meaningful algorithmic accountability — requirements that current frameworks partially address but do not yet fully satisfy.
Mandatory disclosure: Individuals must be informed when AI has materially influenced a decision affecting them. This is a minimum condition for the right to contest. Current disclosure practices are inconsistent and frequently inadequate.
Rigorous bias testing: AI systems used in high-stakes public administration contexts must document error rates across all demographic groups before deployment. The EU AI Act requires this for high-risk applications; enforcement is deferred to 2027. In the interim, voluntary standards and procurement requirements are the primary mechanisms.
Human-in-the-loop requirements: AI outputs must not serve as the sole basis for consequential decisions. Traditional investigative work, human review, and professional judgement must remain primary. This principle is widely stated; its implementation is inconsistent.
Access to technical expertise: Individuals challenging algorithmic decisions must have access to technical experts capable of interrogating the system's architecture, training data, and decision logic. Without this, the right to contest is formal rather than substantive. Providing this access — particularly for indigent or disadvantaged parties — requires institutional investment that most jurisdictions have not made.
A horizontal liability instrument: The withdrawal of the AILD left a gap that the Product Liability Directive and national tort law cannot adequately fill. The case for a replacement instrument — one that addresses the specific evidentiary challenges of AI-related harm — remains compelling. The Commission's 2026 work programme does not include such an instrument. The gap will persist until political will catches up with the scale of the problem.
Conclusion: The Accountability Deficit Is Structural
The accountability deficit in AI justice is not primarily a technical problem. The technical tools for explainability, bias testing, and audit logging exist. The problem is institutional: the incentives, resources, and political will to deploy those tools consistently and to enforce accountability when they reveal failures are absent or inadequate.
The EU AI Act represents the most ambitious attempt yet to address this deficit through law. Its full enforcement — deferred in the most consequential domains until 2027 — will be a test of whether regulatory ambition can be translated into operational accountability. The fragmented national tort landscape that currently governs AI liability is not a stable equilibrium. It is a transitional state that will produce divergent outcomes, jurisdictional arbitrage, and — for the individuals harmed by AI systems in the interim — inadequate remedies.
The liability gap is not a gap in the law alone. It is a gap between the pace of AI deployment and the pace of institutional adaptation. Closing it requires not just legislation but the investment in technical capacity, judicial expertise, and enforcement infrastructure that makes legislation meaningful. That investment has not yet been made at the scale the problem requires.
