A budding trend amongst institutional fact-finders – encompassing national and supranational institutions, agencies and bodies, to the exclusion of private parties – is the increased involvement of Artificial Intelligence (AI) technologies (Article 3(1) Reg. No 2024/1689, AI Act – AIA) in discovery and evidence appraisal procedures in areas like public procurement and cartel control (Imhof, 2023), criminal investigation (Case C-670/22, EncroChat), border control (Ozkul, 2026) and fraud detection (Hague District Court, NJCM et al. v. State of the Netherlands).
While the ultimate decision-makers in those procedural settings remain human, AI-mediated fact-finding raises compelling questions about the epistemic qualities AI systems and their output should possess to be considered as evidence, particularly expert evidence. A cursory overview of the relevant caselaw of the Court of Justice of the EU (CJEU), reveals that those requirements are cumulative and of two kinds: 1. substantive, relating to the characteristics expertise must showcase to qualify as evidence and 2. personal, relating to the qualities expected from the experts called to submit scientific opinions in support of legislation or adjudication.
While AI-generated output may, in principle, satisfy the substantive requirements for qualifying as expertise (1.), AI technologies are not yet sufficiently developed to exhibit the epistemic features expected of experts. In that connection, a significant concern emerges: the neutrality of experts who rely on AI may depend on the neutrality (essentially, freedom from bias) of the AI systems they use – a condition that may be practically difficult to demonstrate (2.).
Before I proceed, a word on the legal systems considered in this study and the analogies that can (and will) be drawn between them. The available and most telling caselaw on AI-assisted or AI-generated expertise is North American and involves criminal proceedings. One could, of course, argue that US legal provisions on expertise in criminal trials hardly have a counterpart in EU law and cannot, therefore, serve as an example for the Union’s and the Member States’ courts. Additionally, the design of criminal procedures usually includes specific procedural components (e.g. jury involvement) and requirements (e.g. proof beyond reasonable doubt) that do not apply to fact discovery and appraisal in administrative procedures.
While these are valid points, they fall outside of the scope of this paper’s discussion which focuses on the possibility that AI-assisted/-generated output present the epistemic qualities (accuracy, impartiality, trustworthiness…) associated with expertise. This epistemic issue has, to date, been primarily addressed by North American jurisdictions; the Union and Member State courts are likely to be called upon to do so in the future, making the existing American caselaw a relevant point of comparison.
1. The prospect of AI output qualifying as expertise
In Union law, institutional factfinders enjoy significant discretion in the discovery and appraisal of highly complex scientific and technical facts (Case C-71/23 P, France v. CWS Powder Coatings, para. 106; Case C‑14/10 Nickel Institute, para. 60; Case C-293/22 P, Chemours Netherlands v ECHA, para. 134, Case C‑119/21 P, PlasticsEurope v ECHA, para. 46) their main limit being the requirement to refrain from manifestly erroneous appraisals (Case C‑14/10, Nickel Institute; para. 60; Case C-293/22 P, Chemours Netherlands v ECHA, para. 134).
That level of discretion notwithstanding, the CJEU has gradually defined three epistemic features that can reinforce the evidence quality of expert opinions. Those features are relevance (for the adoption of a measure or fair adjudication); accuracy (essentially, objective, complete and reality-correspondent fact assessments) and reliability (essentially, trustworthiness of both the experts and their opinions) (Case C-71/23 P, France v. CWS Powder Coatings, para. 73). These can be interpreted as facets of what evidence scholars call the best evidence rule (Nance, 1987) when exploring a technical and scientific issue, an institutional factfinder is required to rely on expert studies that offer the ‘best possible scientific advice’ (Joined Cases C‑6/21 P and C‑16/21 P, Germany v. Pharma Mar, para. 45) in terms of the relevance/accuracy/reliability triad. In this context, one might argue that AI output can offer the ‘best possible scientific advice,’ provided it is seen as sufficiently relevant, accurate and reliable.
The Union and Member States’ courts have not yet addressed this issue, but the American (New York) courts have, in the People v. Wakefield case (NY Court of App., 38 N.Y.3d 367, 2022). During a criminal investigation, the New York police authorities had run DNA samples collected from a crime scene through TrueAllele, a probabilistic genotyping system. Its output was crucial in allowing the police to identify and arrest a suspect. The compelling question during the trial was whether that output met the epistemic requirements to be seen as expertise, which US law articulates in a more elaborate and demanding way than EU law. Indeed, Rule 702 of the Federal Rules of Evidence – which codifies seminal American caselaw on expertise – qualifies expert opinions as testimonies that (should) offer scientific, technical or other specialized knowledge relevant for a trier (like a court) to understand the evidence or determine a fact. It is also based on sufficient facts or data and is a product of reliable principles and methods which ought to be reliably applied by the experts. Guided by these principles – designed to weed out ‘junk science’ (Green & Sanders, 2015, at 210) – the judges in People v. Wakefield concluded that TrueAllele’s output could, indeed, qualify as expert evidence.
This case can be understood as foreshadowing issues that European courts are likely to face, as AI-mediated methods of fact-finding and evidence appraisal become more prevalent. The judges’ reading of TrueAllele’s output through the lens of Rule 702 of the Federal Rules of Evidence may point to a possible need for European courts to further develop, their own epistemic requirements of relevance, accuracy and reliability to adapt them to technology-mediated forms of expertise offered as evidence in judicial proceedings. This epistemic finetuning might be especially needed to improve the effectiveness of judicial scrutiny and protection which traditionally requires not only epistemic soundness of the expertise (AI-mediated or not), but also epistemic integrity of the AI-deploying expert – a point discussed in the following Section.
2. The challenge of AI bias (demonstrably) restricting expert impartiality
A key personal requirement from experts is impartiality – an epistemic predisposition thought to foster the objectivity of the scientific opinions expressed (Joined Cases C‑341/06 P and C‑342/06 P, Chronopost, paras 36 ff). Those should, indeed, offer sufficient guarantees to exclude any legitimate doubt of bias (Case C-291/22 P, D&A Pharma, para. 73, Case C‑680/16 P, August Wolff, para. 27; Case C‑883/19 P, HSBC Holdings, para. 77). If such doubts arise, they should be established, typically through statements and positions taken by the experts concerned (Case C-291/22 P, D&A Pharma, para. 80).
When an expert opinion has been mediated or produced by an AI system – that is, the experts concerned adhere to its output – the impartiality of those experts depends on the absence of bias in the system’s functioning and output. The issue of proving impartiality in such a setting has – again – not been directly addressed by European courts, but by American ones. Two cases are emblematic in that connection: the Loomis ruling (881 N.W. 2d 749 (2016) 2016 WI 68) and the Pickett ruling (Docket N° A-4207-19T4).
In Loomis, the Minnesota lower courts had used a recidivism-predicting AI (COMPAS) while trying a suspect in a drive-by shooting. The defendant contended that COMPAS’s prediction was inaccurate because it was gender biased, flagging male offenders as presenting a higher risk of reoffence than female ones. The debate conducted before the Minnesota Supreme Court on the relevance/accuracy/reliability triptych is noteworthy. First, it was argued that gender is a valid criminogenic factor given that, according to the available statistical data, men are more frequent reoffenders than women. As statistically relevant information, it was not seen as revealing gender discrimination but as faithfully portraying a state of fact. Second and importantly, the courts were asked to explain if and how they verified COMPAS’s accuracy or, at the very least, justify their reliance on its output. Referring to general studies on COMPAS’s performance, the Minnesota Supreme Court confirmed COMPAS’s general aptitude for accuracy, adding that the lower courts’ sentence would have been the same, even if they had not involved COMPAS in the decision-making process (881 N.W. 2d 749 (2016) 2016 WI 68, at 92). The Minnesota Supreme Court’s justification is, indeed, rather ‘thin’, as it rests on ex hypothesi reasoning that amounts to conjecture, rather than a genuine reassurance as to what the outcome of the ruling would have been, had AI not been involved.
A more elaborate debate on the bias neutrality (ergo accuracy) of AI output was in Pickett. In this case, the police authorities had used TrueAllele – the same system as in Whitefield – to perform DNA testing to identify and arrest a suspect. During the trial, the latter asked that the accuracy of the system’s output be confirmed, forcing the New Jersey court to order an independent expertise. The experts arrived at the defeating conclusion that the manual reverse-engineering of TrueAllele’s output would take 8,5 years (Docket N° A-4207-19T4, at 17) thereby defeating the purpose of the trial’s reasonable duration. Faced with the unfeasible case-specific expertise, the New Jersey judges had the same reflex as the Minnesota Supreme Court in Loomis, reverting to already existing, general expertise on TrueAllele’s tendency to give bias-free output. Though the studies were not unanimous, most of them agreed that, indeed, the system was generally accurate. Based on these findings, the New Jersey court considered there was sufficient reason to presume that, considering the circumstances in Pickett, TrueAllele had given a correct DNA match.
Like Loomis, Pickett is open to criticism: the defendants in both cases asked for reassurance from the fact appraisers (i.e. police experts) that they were justified in relying on their AI systems’ output and conducted individualized fact assessments by determining how those technologies affected the defendants’ individual situations; not how they performed generally. This plea for an accurate and individualized fact appraisal is understandable. Both rulings were given following criminal proceedings which, as a rule, require that the facts be established beyond reasonable doubt. Any failure to satisfy that standard of proof (by, say, recklessly relying on biased AI output) increases the risk of miscarriages of justice (i.e. the condemning of innocent persons).
In EU law, the requirement of individual assessment is crucial under Article 47 of the EU Charter of Fundamental Rights, as demonstrated by cases concerning, say, asylum seekers in which the Court has frequently stressed that, for the purpose of individual claim assessments, the involvement of (impartial) experts might be warranted (Case C‑756/21, X). The European Court of Human Rights (ECtHR) follows a similar approach, viewing individual fact assessments as integral to the equality of arms principle. The ECtHR has held that the litigants’ confidence in the ‘workings of justice’ (Pellegrini v. Italy, App. no. 30882/96, para. 45; Zagrebačka banka d.d. v. Croatia, App. no. 39544/05, para. 203; Juričić v. Croatia, App. no. 58222/09, para. 75) is based on the assumption that ‘they are afforded the opportunity to express their views’ and have a right to have ‘knowledge of and comment on all evidence adduced’ (Cangi et al. v. Turkey, App. no. 48173/18, para. 54) including (ideally objective) expert evidence.
There is not yet any European case law clarifying how longstanding (CJEU) requirements on expertise could capture the proof of impartiality in instances where AI technologies are integral to the involvement of expertise in fact-finding and fact-assessment procedures. Perhaps the right to an explanation, enshrined in Article 86 of the AI Act, will allow individuals affected by high-risk AI systems to acquire preliminary knowledge about the role those systems played in shaping their deployers’ decisions. These explanations may, in turn, offer the basis to raise questions about the correctness or the neutrality of specific (expert) fact appraisals before courts. This, however, is a point that the Union courts have yet to clarify in their future caselaw.
The few but significant non-European judicial instances on the role as evidence of AI-assisted (or AI-produced) expertise can, perhaps, be interpreted as foreshadowing court trends in assessing whether that expertise is trustworthy. A point of convergence between the Loomis and Pickett cases is that the opacity of the AI technologies concerned made them inscrutable, hindering the courts from verifying how those technologies affected the trial actors. To overcome the difficulty of performing such verification, the strategy seems to be to ‘go general’ that is, determine if independent experts agree on the aptitude for accuracy of the technologies concerned. I have argued in a previous study (Grozdanovski, 2024) that cases like Loomis, Pickett and even Wakefield, mark a departure from the traditional requirement of probatio (proof of an AI’s specific impact on person’s individual situation) to fama (proof of the reputation of accuracy experts associate with an AI system). While this shift is open to criticism – since the principles of equality of arms and adversarial proceedings require individualized fact assessments – relying on general expert opinions as evidence is undoubtedly the most feasible evidentiary strategy available to courts for overcoming the verification barrier created by AI opacity.
Posted by Lj. Grozdanovski (Assoc. Research Professor FNRS, coordinator of the JUST-AI Jean Monnet Centre of Excellence – University of Liège)
This post is part of a thematic section on REALaw blog on the epistemic boundaries of EU courts, which aims to bring together contributions that analyse how EU courts access, interpret, and evaluate technical and factual information. Together, these contributions assess whether the resulting epistemic boundaries remain adequate in light of contemporary expectations of effective judicial review. This section is coordinated by Dr. Pietro Mattioli, postdoctoral researcher at the University of Maastricht. His research deals with the growing technical specialisation of administrative decision-making and its implications for judicial review.
Cited sources
Green Michael D., Sanders Joseph, ‘Admissibility versus sufficiency. Controlling the quality of expert witness testimony in the United States’ in Miguel Martin-Casales, Diego M. Papayanis (eds.), Uncertain Causation in Tort Law (Cambridge Univ. Press, 2015), 203.
Grozdanovski Ljupcho, “The Explanations One Needs for the Explanations One Gives. The Necessity of Explainable AI (XAI) for Causal Explanations of AI-related harm – Deconstructing the ‘Refuge of Ignorance’ in the EU’s AI Liability Regulation” (2024) 2 International Journal of Law, Ethics & Technology, 155.
Huber Martin; Imhof David, ‘Flagging cartel participants with deep learning based on convolutional neural networks’ (2023) 89 International Journal of Industrial Organization, 102946.
Nance Dale A., ‘The Best Evidence Principle’ (1987) 73 Iowa Law Review, 227.
Ozkul Derya, ‘Automating Immigration and Asylum : The Uses of New Technologies in Migration and Asylum Governance in Europe,’ Refugee Studies Centre (University of Oxford), available on : https://www.rsc.ox.ac.uk/publications/automating-immigration-and-asylum-the-uses-of-new-technologies-in-migration-and-asylum-governance-in-europe (last accessed on 1 September 2026).