CLA News / Deepfakes and Epistemic Authority in Adjudication: Who May Certify Reality? By Ana Zdravković and Bojan Spaić

For decades, audiovisual evidence occupied a privileged place in legal fact-finding. Even when courts formally required authentication, photographs, videos, and audio recordings carried a particular evidential force because they appeared to bring the fact-finder to the event itself. They appear to show, rather than merely describe, what had happened. Deepfakes disturb this confidence at its foundation. Their danger is not limited to the possibility that a fabricated video, image, or voice recording may be accepted as genuine. They also unsettle a deeper institutional question on which adjudication depends: who is entitled to certify reality when reality itself can be convincingly simulated?

This question is particularly important for courts. They do not merely apply the law to facts that already exist in an uncontested form. Adjudication depends not only on rules, but also on the court’s capacity to establish facts in a publicly credible way. Courts establish legally relevant facts by using evidence through procedures of proof, challenge, evaluation, and justification. This is especially so in the “common law jurisdictions”, where the reliability of evidence is often tested through adversarial procedures, expert testimony, disclosure, cross-examination and judicial evaluation of admissibility and weight. Deepfakes therefore directly threaten the institutional authority of courts to determine, in a reasoned and contestable way, what counts as a reliable representation of reality.

A deepfake can be broadly understood as AI-generated or AI-manipulated media that realistically depicts a person as saying or doing something they did not say or do.[1] It should, however, be distinguished from other forms of misleading media. For instance, “AI-altered content” is often described as a broader category that covers content that is partly or wholly generated or modified by AI.[2] There are also “cheap fakes” or “shallow fakes”, which are misleading media produced without advanced AI, for example through cropping, slowing down, splicing, or ordinary editing.[3] One could also encounter the notion of “out-of-context media”, referring to technically authentic but paired with a false date, caption, location, or factual claim. Such a conceptual distinction matters legally, because, for example, a forensic detector designed to identify generative artefacts will not solve a case in which genuine footage has simply been miscaptioned. Legal analysis must identify the mechanism of deception before assigning burdens, remedies, or institutional responsibilities.[4]

However, audiovisual evidence has always had a special persuasive force, with its authority being built on at least four pillars. First, even though judges are aware that images and recordings may be staged, selectively edited, wrongly attributed, or taken out of context, recordings still create an impression of immediacy because they seem to collapse the distance between the fact-finder and the event. Then, there is the idea of mechanical objectivity, meaning that the camera as an object is “objective” compared to human witnesses, having no memory problems, emotions, or biases in the same way a person might. Moreover, audiovisual evidence can preserve details that testimony may omit or distort, such as tone of voice, timing, or facial expressions. Finally, they are also replay able, as the same exhibit can be inspected repeatedly, collectively, or in slow motion.

Deepfakes exploit precisely this inherited authority of audiovisual evidence. They parasitise the credibility of the recording form and use it against the legal process itself, thereby creating a two-fold threat to adjudication.

The first is false acceptance, in which a fabricated exhibit may be treated as authentic and may support wrongful factual findings. The second danger is false rejection, meaning that genuine evidence may be dismissed as fake because the mere existence of deepfakes makes denial more plausible. Chesney and Citron describe this second risk as the “liar’s dividend”, whereby growing awareness of audiovisual manipulation makes it easier for wrongdoers to dismiss authentic evidence as fabricated.[5] In other words, the mere existence of deepfakes gives a new evidential argument, since the parties to the case can theoretically claim that the video is not real, even when it is. For courts, this means the challenge is not only to exclude fakes but also to prevent authenticity from becoming so expensive or uncertain that genuine evidence loses its practical value and access to justice is jeopardised.

There are some recent examples which illustrate this double threat. In one of the proceedings in Serbia concerning the alleged criminal group led by Veljko Belivuk and Marko Miljković, one photograph became contested, as the defence argued that the photograph had been manipulated and pointed to metadata allegedly showing different creation and modification dates; hence, they asked for forensic examination.[6] As the outcome of this examination is still unknown, the point here is not whether the photograph was authentic, but the credibility contagion. In this case, which is still ongoing, a dispute over one photo was used to cast doubt on a much larger evidential record, including SKY communications, which are the prosecution’s key evidence.[7] In other words, an authenticity challenge can spread from one item to an entire evidential ecosystem, making the judicial task significantly more difficult. Therefore, an epistemic lesson from this case is that the court needs item-specific analysis because a single weak exhibit does not automatically contaminate everything else.

The so-called “deepfake defence” in the United States provides a second example. In January 6 litigation, Joshua Christopher Doolin challenged open-source videos that placed him at the Capitol and argued that their authenticity had not been adequately verified.[8] The government did offer circumstantial evidence that provided a prima facie basis to believe that the open-source video was authentic, such as by comparing several distinctive points among surveillance footage from the US Capitol Police, body camera footage from the Metropolitan Police Department, and the open-source video.[9] They also responded that this argument affected the jury’s assessment of the videos, not their admissibility, especially because Doolin offered no evidence that the videos were fake.[10] The court admitted the videos, and the authenticity challenge was treated as going primarily to weight rather than admissibility, especially because the defendant did not offer concrete evidence that the videos were fake.[11] This is also an important example because the court did not rely on the visual appearance of a single video, nor did it require impossible certainty. Instead, authority arose through convergence between independent sources.

A third example is a wrongful death case, where recordings of Elon Musk speaking publicly about Tesla’s self-driving capabilities were challenged on the ground that they might be deepfakes.[12] The judge ordered a limited deposition so that Musk could answer under oath whether he made the statements, thereby refusing to let technological uncertainty become a privilege of the powerful.[13] A generalised allegation that public recordings might be deepfakes was not enough to avoid ordinary evidential accountability; hence, the court returned the authority to a familiar legal mechanism: sworn, contestable testimony.

These examples suggest that deepfakes do not abolish the epistemic authority of audio-visual evidence. They redistribute it. In traditional authentication, authority may appear to rest mainly with a witness confirming that a recording accurately depicts what it claims to show, but in deepfake disputes, this is often inadequate. A witness might recognise the person or scene without being able to detect synthetic manipulation. Yet epistemic authority should not shift entirely to experts or detection systems, since a detector’s confidence score does not establish legal authenticity. Moreover, deepfake detection tools have serious limits, as their performance may depend on the generation method, training data, compression, platform processing, and the type of media being examined. They are also vulnerable to adversarial adaptation because once detection methods become known, creators of synthetic media may optimise fakes to evade them. For that reason, detection should be treated as evidence about authenticity, not as authenticity itself.[14]

Therefore, it seems that a more appropriate model is layered authentication, where courts should ask a series of connected questions. For example, relevant questions would concern provenance, or in other words, what is the origin of the file. Then, integrity also matters, as it reveals whether it has been altered after capture or collection. Third, corroboration would show whether other independent evidence converges with it. Fourth, forensic examination is indispensable for a qualified analysis, including the level of certainty of the result. Finally, contestability should not be abandoned, as it allows the opposing party to inspect, challenge, and, where appropriate, reproduce the analysis.

This model is consistent international standards in the field, primarily with the Council of Europe’s approach to electronic evidence, which emphasises authenticity, integrity, metadata, expert assessment, and judicial evaluation,[15] and Article 50 of the EU AI Act, which also insists on transparency,[16] as well as the standards stemming from Article 6 of the European Convention on Human Rights, which require that a fair trial must include the effective opportunity for parties to challenge decisive evidence and the procedures used to validate it.[17]

Such a model would also preserve the proper role of the judge by framing deepfake authentication not as a transfer of authority from judges to computer scientists, but as a structured network of distributed epistemic authority. Within that network, witnesses would contribute recognition and first-hand knowledge; custodians would contribute information about origin, handling, and chain of custody; experts would contribute forensic analysis and explain methodological limits; and technical systems may contribute metadata, signatures, watermarking, or provenance records. Now, judges would remain responsible for admissibility, probative value, procedural fairness, and the final justification of reliance on evidence.[18]

A final aspect of this model is its preference towards a calibrated burden-shifting approach. The danger of allowing a party to defeat evidence by merely saying “deepfake” has already been highlighted. Such an empty objection would reward strategic denial and amplify the liar’s dividend. Needless to say, it would be rather problematic to require full expert verification for every digital exhibit, as that would make authentication prohibitively expensive and could systematically exclude genuine evidence. A balanced approach would require the challenger to identify specific grounds for suspecting fabrication or material alteration. Once such grounds are shown, the proponent should provide an appropriately stronger foundation for authenticity and reliability.[19]

The practical response should therefore be procedural rather than purely technological. Courts should require concrete allegations, not generic technological doubt. They should preserve original files, metadata, and chain-of-custody information, triangulating disputed media with independent recordings, testimony, timestamps, geolocation, contextual facts, and platform records where available. They should also use qualified experts where the dispute is material and technically complex and requires disclosure of methods, assumptions, limitations, and materials needed for meaningful challenge. Finally, courts should explain authenticity, admissibility, probative weight, and residual uncertainty separately.

This last point is crucial. The court should not hide uncertainty behind the single word “authentic.” In many cases, a better conclusion will be more nuanced, emphasising that evidence has sufficient foundation for admission, that some uncertainties remain and that those uncertainties have been considered in the final evaluation. This kind of transparent reasoning is essential for preserving public confidence in adjudication.

To conclude, deepfakes do not create a wholly new “post-truth courtroom.” Courts have always dealt with forged documents, unreliable witnesses, manipulated photographs, broken chains of custody, and disputed expert evidence. What deepfakes do is intensify these older problems and make the institutional architecture of proof more visible, showing that evidential authority cannot safely be located in a single object, witness, expert, or technology.

The answer, therefore, should not be blind trust in audiovisual evidence, nor generalised suspicion toward all digital media. The answer should be calibrated uncertainty. Courts must resist both fabricated evidence and the liar’s dividend, moving from object-based trust to process-based authentication, from expert supremacy to distributed competence, and from absolute confidence to reasoned, contestable justification. In the age of deepfakes, the court remains a legitimate authority on matters of fact only by making its dependence on other sources of knowledge visible, structured, and open to challenge.

Ana Zdravković , Institute of Comparative Law, ORCID ID: https://orcid.org/0000-0001-7514-8892,  a.zdravkovic@iup.ac.rs

Bojan Spaić, University of Belgrade Faculty of Law, ORCID ID: https://orcid.org/0000-0002-8887-9683, bojan.spaic@ius.bg.ac.rs

[1] R. Chesney, D. K. Citron, “Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security”, 107 California Law Review, 2019, 1753, 1757; E. Tuysuz, C. Kılıç, “Analyzing the Legal and Ethical Considerations of Deepfake Technology”, Interdisciplinary Studies in Society, Law, and Politics 2(2), 2023, 4; J. Kazaz, Regulating Deepfakes: Global Approaches to Combatting AI-Driven Manipulation – Policy Paper, Centre for Democracy & Resilience, GLOBSEC, 2024, 2.

[2] B. Paris, J. Donovan, “Deepfakes and Cheap Fakes: The Manipulation of Audio and Visual Evidence”, Data & Society, 2019; Kazaz, 2024, 6.

[3] Ibid.

[4] Kazas, 2024, 6.

[5] Chesney & Citron, 2019, 1785. They noticed the same thing with regard to fake news phenomenon.

[6] Available via https://n1info.rs/vesti/aleksandar-vucic-majica-belivuk-navijaci-partizan/, accessed 10.7.2026.

[7] Ibid.

[8] M. Ferraro, B. Gurney, “The Other Side Says Your Evidence Is A Deepfake. Now What?”, Law360, 2022.

[9] Ibid.

[10] Ibid.

[11] Ibid.

[12] Huang v Tesla, Santa Clara County Superior Court, order of Judge Evette Pennypacker, 27 April 2023; S. Bond, “People Are Trying to Claim Real Videos Are Deepfakes. The Courts Are Not Amused.” All Things Considered, NPR, 2023.

[13] Ibid.

[14] Louise Matsakis, 2018; Claire Leibowicz, Sean McGregor and Aviv Ovadya, 2021; Ba et al, 2024.

[15] Council of Europe, Committee of Ministers, Guidelines of the Committee of Ministers of the Council of Europe on Electronic Evidence in Civil and Administrative Proceedings, adopted January 30, 2019, CM(2018)169-add1final, paras. 13–18, 22–25, 33–35.

[16] European Parliament and Council of the European Union, Regulation (EU) 2024/1689 of June 13, 2024, Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act), Article 50(2), (4), Official Journal of the European Union, July 12, 2024.

[17] European Court of Human Rights (ECtHR), Yüksel Yalçınkaya v. Türkiye, App. No. 15669/20, Judgment of 26 September 2023, para. 303; ECtHR, Bykov v. Russia, App. No. 4378/02, Judgment of 10 March 2009, para. 90.

[18] This model would also support and include what Citron has coined as “technological due process”, insisting that companies must be truly transparent about content policies and revising accordingly Terms of Service of digital platforms, see D. K. Citron, “Technological Due Process”, Washington University Law Review 85, no. 6, 2008, 1249–1313.

[19] J. P. LaMonaga,  “A Break From Reality: Modernizing Authentication Standards for Digital Video Evidence in the Era of Deepfakes”, American University Law Review 69, no. 6, 2020, 1945–1988. See also Advisory Committee on Evidence Rules, Agenda Book for the Meeting of the Advisory Committee on Evidence Rules, Administrative Office of the United States Courts, April 19, 2024.