AI-Generated Evidence in the Judicial Proceedings

The concept of generative AI systems

To fully comprehend the nature of AI-generated outputs and effectively integrate them in judicial proceedings, a thorough understanding of generative AI is required. Since generative AI systems produce data that parties may submit to the courts or that courts can use in judicial proceedings, it is also important to understand the risks associated with their use.

In 1978, Richard Bellman defined AI as the automation of activities associated with human thinking, including decision-making, problem-solving and learning (Bellman, 1978). AI was later defined as the hypothetical ability of a computer to match or exceed a human’s performance in tasks requiring cognitive abilities, such as perception, language understanding and synthesis, reasoning, creativity and emotion (Grimm et al., 2021, p. 14). The European Commission’s High-Level Expert Group on Artificial Intelligence defined AI as systems designed by humans that, given a complex goal, act in the physical or digital world by perceiving their environment, interpreting the collected structured or unstructured data, reasoning on the knowledge derived from this data and deciding the best actions to take (according to pre-defined parameters) to achieve the given goal (European Commission, 2018a, p. 7).

An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions that may influence physical or virtual environments (AI Convention, 2024, art. 2). Different AI systems vary in their levels of autonomy and adaptiveness after deployment and can be defined differently. AI systems can be purely software-based, acting in the virtual world as voice assistants, image analysis software or search engines, or embedded in hardware devices such as advanced robots, autonomous cars and drones (European Commission, 2018a). Generative AI encompasses a wide range of AI systems dedicated to creating or generating content or data that often resembles human-generated content (Feuerriegel et al., 2023). It also refers to deep-learning models that can generate high-quality text, images and other content based on the data they were trained on (What Is Generative AI?, 2021).

Generative AI is a technology that can provide conclusions and analyses that previously required human intelligence. Such systems analyse data and produce opinions and conclusions based on the information they receive and the methods programmed into them. It may generate information, such as text, video or images, by identifying patterns and creating new variations based on them. The complexity of using AI-generated outputs in judicial proceedings is heightened by their distinct characteristics that may affect the fairness of the proceedings.

First, AI systems may generate false information, which raises concerns about the accuracy and reliability of their results (Explanatory Report to the AI Convention, para 44). AI systems produce evidence with much lower accuracy than humans, missing 96% of relevant studies and information and making up to 70% screening errors (Clark et al., 2025). Second, algorithmic prejudice and bias are major concerns regarding AI-generated results (Cuellar et al., 2025). AI-generated outputs are compromised by errors resulting from data-quality issues, such as incomplete or incorrectly labelled training sets (Neil and Zanger-Tishler, 2025). Finally, there are explainability issues with the decision-making processes and outcomes of AI systems, as even developers frequently struggle to fully understand AI systems (Singh and Devi, 2025, p. 47).

Hence, generative AI systems are trained on data sets to produce conclusions, just as humans do. Because of their complexity, AI-generated results may be biased, difficult to explain and prone to accuracy issues. These issues raise concerns about whether AI systems can produce data that may be used as evidence in judicial proceedings, while complying with the right to a fair trial.

The treatment of the AI-generated outputs as evidence

Traditionally, courts rely on human experts, witness testimonies whose credentials, methods and findings are subject to cross-examination, and documentary records and physical objects (Singh and Devi, 2025, p. 46). In the majority of civil law systems, evidence is classified into five main categories: the testimony of the parties, witnesses and experts; documents; and physical objects (Melillo, 2021, para. 6). Evidence is created by identifiable individuals who can be questioned and held responsible when their reasoning can be explained and challenged. AI evidence is distinguished from these traditional types of evidence by its generative and complex nature, as it involves substantial analysis and judgement by machines, not humans (Li, 2025, p. 62). The question is whether AI-generated outputs can be regarded as evidence.

Evidence may be defined as any matter of fact that can be admitted by a trial submitted by a party to a case to prove or disprove an issue in the case (‘Evidence’, n.d.). It is also regarded as information which may be used to prove the existence of a fact in issue or a collateral fact or to disprove a fact in issue or a collateral fact (Monaghan, 2015, p. 2). The term may encompass a proposition of fact that is established by evidence in the first sense, ‘evidential fact’ (Ho, 2021). A functional understanding of evidence allows AI-generated outputs to be qualified as evidence, as they constitute information that may assist in establishing the facts of the dispute. As long as AI-generated outcomes are used to establish a conviction or resolve a civil dispute, they should be considered as evidence. The question remains how AI-generated evidence should be treated to ensure fairness in the trial.

AI-generated evidence and expert testimonies

AI-generated evidence that provides analyses and summaries of complex data sets may resemble expert testimonies. Experts convey scientific and/or technical knowledge to the judge, enabling the latter to conduct an objective and clear investigation and evaluation of the facts. Expert reports are prepared by analysing the given data and making a decision drawing on their specialised knowledge. Similarly, AI systems trained on previous data sets can analyse new data and draw conclusions using algorithms. The main difference is the ability to assess the trustworthiness of expert reports and AI-generated outputs.

The U.S. Judicial Conference’s Advisory Committee on Evidence Rules in Washington, D.C., proposed a rule that would require courts to use existing criteria for assessing the reliability of expert testimony to assess the reliability of AI and other machine-generated evidence (Nojeim and Bowman, 2026). Under this proposal, the reliability of AI-generated evidence shall be subject to the same requirements as those in Rule 702 of the Federal Rules of Evidence (Reuters, 2025). The proponent shall demonstrate that the AI system will help the trier of fact to understand the evidence or determine a fact in issue that the AI-generated results are based on sufficient facts or data, are the product of reliable principles and methods, and reflect a reliable application of the principles and methods to the facts of the case (Legal Information Institute, n.d.).

The reliability of a human expert’s testimony depends on their education, experience and adherence to professional norms. The testimony shall be based on sufficient data analysed by reliable and relevant principles and methods. Likewise, to understand the reliability of AI-generated outputs, the training data, the learning algorithms and the source code, which encompasses the full model parameters of an AI system, shall be disclosed (Mitchell et al., 2023, pp. 291–293).

Several problems stem from this approach. First, the AI system may not be cross-examined as thoroughly as human experts. Cross-examination is deemed the surest test of truth and a better security than the oath (E and Wellman, 1904, p. 3). It enables verification of the expert’s actual technical knowledge, understanding of the facts of the case and the methods applied. In addition, it allows confronting the expert’s conclusion with competing authorities, challenging underlying assumptions, questioning why one variable was given more weight than another and asking for reconciliation of the differences. Second, an AI system’s expertise depends on its training data sets and learning method, which are often withheld from examination by the defence (Mitchell et al., 2023, p. 331). Even if the AI system’s internal materials are disclosed, they shall be analysed by experts with sufficient technical knowledge to assess the system’s inputs and outputs. Finally, when the expert can be held liable for being impartial or for fabricating the facts in the report, who will be held liable for an AI system’s mistake? Hence, the AI-generated evidence shall be treated differently from expert testimonies, as its trustworthiness cannot be verified or guaranteed in the same way.

AI-generated evidence and electronic evidence

One may argue that AI-generated evidence shall be treated as electronic evidence. The latter refers to any evidence derived from data contained in or produced by any device, the functioning of which depends on a software program or data stored on or transmitted over a computer system or network (Conseil de l’Europe, 2019, p. 6). The ECtHR recognised that it differs from traditional forms of evidence by its nature and the specialised technologies required for its collection, storage, processing and analysis. Most significantly, it raises serious reliability issues, as it is inherently more prone to destruction, damage, alteration or manipulation (Yüksel Yalçinkaya v Türkiye, 2023, para. 312). However, electronic evidence is treated no differently from other types of evidence in judicial proceedings. Its different nature is taken into account, but the submission of electronic evidence does not require additional safeguards.

Information generated by AI systems is electronic in nature but exhibits additional characteristics. The development of an AI system centres on building an accurate statistical model of the training data, from which it can predict, make decisions and generate outputs (Flach, 2012, p. 3). The quality of the AI-generated evidence is determined by the quality of its training data and the choices made by the selected algorithms. The difference is that, unlike electronic evidence, which records and traces real events, AI-generated evidence is a probabilistic inference based on statistical patterns in training data. Hence, AI-generated evidence shall be treated differently from electronic evidence.

AI-generated evidence and hearsay

AI-generated evidence may be treated similarly to hearsay evidence. The latter is evidence of a fact not actually perceived by a witness with his or her own senses but of what someone else has been heard to say (Mallow, 2020, pp. 453–454). Since the statement is made outside the courtroom, it cannot be subjected to cross-examination to verify the witness’s credibility and the evidence’s reliability. Just as hearsay evidence raises concerns about the reliability and testability of second-hand assertions, AI-generated outputs raise similar concerns. The importance of a careful assessment of any hearsay evidence relied upon has been emphasised (Dzelili v Germany (Déc.), 2009), and thus, it is given weight only to the extent it is found to be reliable. Per se, hearsay evidence may be admitted in civil law countries, but it requires a case-by-case assessment and additional safeguards to ensure the fairness of the proceedings. Similarly, AI-generated evidence requires careful evaluation of data quality, the algorithms used and safeguards, given its inherent potential to undermine the fairness of the trial upon submission.

Hence, the accuracy and reliability of AI-generated evidence can be influenced by many factors, including data, algorithmic and process objectivity. It differs from traditional types of evidence, such as expert reports, because its reliability may not be verified. There is no human intervention during the decision-making process, and it is not subject to cross-examination. AI-generated evidence also differs from electronic evidence, as the latter is simply the form that confirms the real evidence that occurred. The untested nature of AI-generated evidence makes it akin to hearsay. Similar to the latter, AI-generated evidence requires careful assessment of data sources, algorithm design, processing reliability and additional safeguards to ensure the fairness of the trial when it is submitted.

AI-generated evidence and the right to a fair trial under Article 6 of the ECHR

Article 6 of the ECHR does not prescribe any rules on the admissibility of evidence (Schenk v Switzerland, 1988, para. 46). It is considered that the rules of evidence are, in principle, a matter for each state to determine (García Ruiz v Spain, 1999, para. 28). It is not the role of the ECtHR to pronounce on in what circumstances and format intelligence information may be admitted in proceedings as evidence (Yüksel Yalçinkaya v Türkiye, 2023, para. 316).

When the applicant alleges the violation of Article 6 of the ECHR, the ECtHR assesses whether the proceedings as a whole were fair. It considers whether the rights of the defence were respected and the strength of the evidence, especially where there are doubts as to its authenticity (Al-Khawaja and Tahery v the United Kingdom, 2011, paras. 131–134; Yüksel Yalçinkaya v Türkiye, 2023, para. 303).

In assessing the overall fairness of the proceedings, particular regard must be given to whether the rights of the defence are respected in a manner consistent with Article 6 of the ECHR (Bykov v Russia, 2009, para. 90). The ECtHR assesses whether the applicant was given the opportunity of challenging the evidence and of opposing its use in circumstances where the principles of adversarial proceedings and equality of arms between the parties were respected (Kobiashvili v Georgia, 2019, para. 56).

In terms of the authenticity of evidence, the quality of the evidence shall be taken into consideration by the ECtHR when assessing the fairness of the proceedings, as well as the circumstances in which it was obtained. The ECtHR assesses whether these circumstances cast doubt on the reliability or accuracy of the evidence (Bykov v Russia, 2009, para. 89; Erkapić v Croatia, 2013, paras. 72–73). In cases where the collection or processing of intelligent information is not subject to prior independent authorisation or supervision, post facto judicial review or other procedural safeguards, or where it is not corroborated by other evidence, its reliability is more likely to be called into question (Yüksel Yalçinkaya v Türkiye, 2023, para. 316). Hence, when AI-generated evidence is used in the judicial process, the rights of the defence shall be respected, especially the right to challenge the evidence on equal footing with the other party, and the authenticity and reliability of the evidence shall be ensured to guarantee the fairness of the proceedings under Article 6 of the ECHR.

The ECtHR stresses the reliability of AI-generated evidence when it constitutes decisive evidence in criminal cases.1 The truth can only be established based on reliable information (Evidence and Proofs from the Perspective of the European Court of Human Rights, n.d., p. 1). Domestic courts must base their decisions to convict the person only on evidence that is not subject to any doubt as to its authenticity or accuracy. The guilt of the accused shall be proven beyond a reasonable doubt on evidence that is concrete and reliable (Çakici v Turkey, 1999, para. 85). The domestic courts should approach untested evidence, the reliability of which cannot be verified, with caution and provide detailed reasoning for why they considered it reliable (Al-Khawaja and Tahery v the United Kingdom, 2011 para. 161; Brzuszczyński v Poland, 2013 paras. 85, 86, 89). This becomes more significant when an accused is unable to assess its reliability directly through cross-examination or is denied access to its internal materials (Yüksel Yalçinkaya v Türkiye, 2023 para. 334).

Where a conviction is based decisively on evidence, the reliability of which cannot be verified and guaranteed, the Court must subject the proceedings to the most searching scrutiny and take into consideration its effects on the rights of the defence. For example, the admission of statements of absent witnesses as evidence results in a potential disadvantage for the defendant, who, in principle, should have an effective opportunity to challenge the evidence against him or her in a criminal trial (Jasper v the United Kingdom, 2000 para. 51). In particular, the party should be able to test the truthfulness and reliability of the evidence given by the witnesses, by having them orally examined in his or her presence, either at the time the witness was making the statement or at some later stage of the proceedings.

Hence, Article 6 of the ECHR does not establish rules for the admissibility of evidence. The quality, reliability and the rights of the party to challenge and examine the evidence are important to ensure the fairness of the proceedings, since the right to a fair trial is an obligation measured by the consequence. The party is put at a potential disadvantage when evidence is submitted without an adequate opportunity to challenge it, thereby violating the fairness of the trial.

Safeguards when AI-generated evidence is submitted to the court

AI-generated evidence undermines the fairness of proceedings if used without caution and submitted without additional safeguards.

First, most generative AI systems have restricted access to their internal materials for national security or trade-secret reasons, which affects the parties’ rights to access the evidence needed to prepare the defence and to challenge the results under Article 6 of the ECHR. When access to internal materials, such as source code or training data, is denied, the party is deprived of the opportunity to verify the reliability and accuracy of the AI system that generated the evidence. This became evident in the Marengo case, where the defence emphasised that it had no adequate opportunities to review and contest evidence obtained through a digital forensic platform due to a lack of access to its source data and software (Skaramuca, 2025, part 4.3.3). Moreover, the AI system cannot be effectively challenged by the defence, as the accused cannot question it or verify its results. The party is then put at a disadvantage, unable to interrogate the AI system, verify how it produced the outcomes and thus challenge the evidence submitted. This tension was already evident in the case State v Loomis, where the accused was unable to examine a risk-assessment COMPAS system because the vendor claimed it as a trade secret (State v Loomis, 881 N.W.2d 749, 2016). The Supreme Court acknowledged the risk of such systems being used and prescribed how these assessments must be presented to trial courts and the extent to which judges may use them (hlr, 2017).

Second, the decision-making process of generative AI systems cannot be fully understood or explained even by their creators, and their outputs may be affected by biased training data, flawed input information or algorithmic errors that are neither visible nor correctable from the outside. Even if such an AI system is deployed in full technical compliance with the standards prescribed under instruments such as the AI Convention, this cannot guarantee the accuracy of any generated output, nor can it render the transparent decision-making procedure in a manner compatible with the right of the defence under Article 6 of the ECHR. Given the central role that the establishment of truth plays in justice, any doubt that the circumstances in which evidence was obtained have impacted its reliability and accuracy is sufficient to deem the trial unfair.

Third, issues arise when AI-generated evidence is used as the decisive basis for establishing the guilt of the accused. Given the importance of such evidence, it shall be reliable and its quality shall be subjected to the utmost scrutiny. Reliability of evidence means that it can provide accurate information about facts that can prove or disprove any of the relevant elements of the case. However, the inherent nature of generative AI systems means that the accuracy of AI-generated results cannot be guaranteed.

Because of these risks, when submitting AI-generated evidence, it would be crucial to require sufficient counterbalancing measures, including strong procedural safeguards. It is appropriate to treat AI-generated evidence similarly to untested evidence, such as hearsay, since the reliability of both cannot be guaranteed or verified by cross-examination. The ECtHR would similarly examine whether there are sufficient counterbalancing factors in place, including measures that permit a fair and proper assessment of the reliability of AI-generated evidence when assessing the violation of Article 6 of the ECHR (Al-Khawaja and Tahery v The United Kingdom, 2011 para. 147).

Justice Bradley in the State v Loomis case proposed that AI-generated outputs may be used if judges who must rule on the admissibility of AI-generated evidence have a working knowledge of what AI is and how it works (hlr, 2017). However, this alone will not suffice to ensure the fairness of the proceedings. The domestic courts should indeed approach AI-generated evidence with caution and provide detailed reasoning as to why they considered such evidence to be reliable (Al-Khawaja and Tahery v The United Kingdom, 2011 para. 161; Brzuszczyński v Poland, 2013 paras. 85, 86). The courts should approach AI-generated content with heightened scepticism, given its potential for manipulation (National Center for State Courts, 2025). However, the nature of AI-generated evidence and the technology used to generate it are complex, which may diminish national judges’ ability to establish its authenticity and accuracy and explain why they consider such evidence reliable. The courts will require the expert to rule on the accuracy of the results, which may lead to longer, more costly proceedings (National Center for State Courts, 2025).

The Federal Court of Australia issued a practice note on the submission of evidence that was enhanced or created using AI (Federal Court of Australia, 2026). According to this note, any person who uses Generative AI should have a basic understanding of its capabilities, limitations and risks and any use of AI must not adversely affect the administration of justice. In addition, the Federal Court expects that the responsible person will have confirmed that facts stated in pleadings are based on what the party reasonably considers can be proved and claims for relief are based on proper legal principles, legal authorities cited in submissions exist and support the proposition stated, evidence cited in submissions exists and is reasonably likely to be admissible, statements about what the evidence proves are findings reasonably open for the Court to make, and chronologies are accurate (Federal Court of Australia, 2026). These safeguards promote procedural integrity but are limited to regulating human use of AI systems and do not address their internal reliability.

A significant safeguard is the corroboration of AI-generated outputs (Schatschaschwili v Germany, 2015). This refers to the support for a proposition already provided by AI-generated evidence, by another piece of evidence (Gardiner, 2023, p. 1). Corroborative evidence must be reliable and independently indicate the guilt of an accused (Bykov v Russia, 2009 para. 90). It shall be noted that this evidence shall not point to the technical characteristics of the AI system that generated the evidence or the reliability and accuracy of its results. Such evidence should prove the same fact as AI-generated evidence. Corroborative evidence may comprise further factual evidence secured in respect of the fact already established in the case, including forensic evidence or expert opinions (Gani v Spain, 2013, para. 48). Judges using AI outputs must explain the factors other than the AI assessment that support the sentence imposed (hlr, 2017).

Another important safeguard will be an independent examination of the AI system that generated the evidence. However, the expert shall have full access to the system to examine its reliability to produce a reasoned and accurate conclusion. The possibility of submitting questions concerning the functioning or outputs of the AI system during the proceedings to developers and the users of the system should also constitute a safeguard (Yevgeniy Ivanov v Russia, 2013 para. 49).

Hence, when AI-generated evidence is submitted, the defence faces a disadvantage against the other party because of the inability to cross-examine, verify or even understand the mechanism by which the evidence was generated. Transposing the ECtHR’s approach to untested evidence, the domestic courts shall consider whether sufficient safeguards are in place when AI-generated evidence is submitted, including measures that permit a fair and proper assessment of its reliability. When AI-generated evidence is submitted, a judgement may be based on it only if it is sufficiently reliable, given its importance in the case and if accompanied by other safeguards. The safeguards developed for untested evidence in existing case law, such as the submission of corroborating evidence, an independent examination of an AI system’s reliability and a reasoned explanation why the courts consider AI-generated evidence reliable, could be applicable. The safeguards will be sufficient if they ensure a fair and proper assessment of the reliability of AI-generated evidence. The reliance on AI-generated evidence in the absence of additional safeguards renders judicial proceedings unfair, not only in criminal cases but also in civil cases.