The Legal Implications of AI-Generated Evidence in Courtrooms.

Reliability, Admissibility, and the Risks of Bias and Manipulation

By: Stephanie Hepzibah

Abstract

The courtroom is among the most consequential forums in civil society, a space in which the quality and integrity of evidence can determine whether a person is freed or imprisoned, whether a corporation is dissolved or vindicated, whether a family receives justice or is denied it. The emergence of artificial intelligence as a tool for generating, processing, and presenting evidence has introduced capabilities that simultaneously expand the possibilities of justice and threaten its foundations. This paper examines the legal implications of AI-generated evidence across three interconnected dimensions: reliability, admissibility, and the risks of bias and manipulation. Drawing on peer-reviewed scholarship, case law, Federal Rules of Evidence, and recent empirical research, the paper traces the evolution of courtroom evidence from the pre-digital era to the present, examines how AI is being deployed by both prosecution and defence, interrogates the moral and legal arguments for and against its admission, and analyses how the same technology that could assist in convicting the guilty can equally be weaponised to destroy the innocent. The paper concludes that AI-generated evidence is neither inherently trustworthy nor inherently inadmissible. Still, it demands a rigorous and evolving regulatory and judicial framework that the current legal system has not yet adequately developed.

 1. Introduction

1.1 Courtroom Evidence Before the Age of Artificial Intelligence

For most of the history of formal adjudication, the evidence that determined the outcome of legal proceedings was human in origin and analogue in form. Witnesses gave oral testimony subject to cross-examination; physical objects were introduced as exhibits; written documents, letters, contracts, and records were authenticated by those who created or received them. The rules governing this evidence evolved over centuries to address a fundamental problem: human testimony is fallible, human memory is reconstructive rather than reproductive, and human beings are capable of deliberate deception (Mnookin, 1998). The common law rules of evidence and their codified successors in jurisdictions such as the United States Federal Rules of Evidence (FRE)  were constructed around this basic epistemic reality. The hearsay rule, the best evidence rule, the authentication requirements of Federal Rule of Evidence 901, and the standards for expert testimony codified in Federal Rule of Evidence 702 all reflect, at their core, a system designed to manage the unreliability of human observation and the temptation to fabricate.

The arrival of forensic science in the twentieth century introduced a new category of evidence that promised greater objectivity: fingerprint analysis, blood typing, ballistic examination, and eventually DNA profiling. Each of these technologies was initially contested in courtrooms before gaining acceptance through the twin processes of scientific validation and judicial gatekeeping (Gless, 2020). The landmark decision in Frye v. United States, 293 F. 1013 (D.C. Cir. 1923), established that scientific evidence must be "generally accepted" within the relevant expert community to be admissible, a standard that shaped American evidence law for decades. The subsequent replacement of Frye with the Daubert framework established in Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993), extended judicial gatekeeping to require active assessment of scientific reliability, testability, peer review, and known error rates (Federal Rules of Evidence, 2023 Advisory Committee Notes). These foundational debates about scientific evidence foreshadowed, with remarkable precision, the arguments that now surround AI-generated evidence.

1.2 The Arrival of Artificial Intelligence and Its Courtroom Footprint

Artificial intelligence entered legal proceedings gradually, initially through the use of statistical software programs such as SAS and Stata, whose outputs formed the basis of expert testimony in employment discrimination and antitrust cases (Secretariat International, 2025). Courts accepted these tools with relative speed once it was established that their underlying methods were transparent, peer-reviewed, and replicable. The transition from statistical software to machine learning represented a qualitative shift. Machine learning models, including the deep learning architectures that underlie today's most powerful AI systems, operate through processes that are not easily reducible to human-readable logic. They are, in the language of the courts and scholars, "black boxes": systems whose inputs and outputs can be observed but whose internal reasoning cannot be fully examined by the parties or the judge (National Violent Death Reporting System [NVDRS] Courts Materials, 2023).

The first AI tool to gain widespread use in American criminal courts was not a generative model but a predictive one: the COMPAS algorithm (Correctional Offender Management Profiling for Alternative Sanctions), developed by Northpointe, Inc., and used across 46 American states to generate recidivism risk scores used in pretrial detention and sentencing decisions (Angwin et al., 2016). The constitutional and ethical questions raised by COMPAS culminated in State v. Loomis, 881 N.W.2d 749 (Wis. 2016), where the Wisconsin Supreme Court upheld the use of the algorithm at sentencing while acknowledging serious due process concerns, signalling a new era in which machine-generated outputs directly shaped judicial outcomes. Since then, AI has entered courtrooms in multiple additional forms: facial recognition matching surveillance footage to suspects, AI-enhanced video analysis, DNA probability modelling, natural language processing of communications data, and, most recently, generative AI tools capable of producing entirely synthetic text, audio, image, and video evidence (Delfino, 2023; Linna et al., 2025).

1.3 The Courtroom Today

By 2025, the integration of AI into legal proceedings had progressed from a theoretical concern to a documented reality. In Mendones v. Cushman & Wakefield, Inc., California Judge Victoria Kolakowski identified a deepfake video, a fabricated audiovisual recording submitted by self-represented plaintiffs as authentic witness testimony and dismissed the case based on the fraudulent submission (National Centre for State Courts [NCSC], 2026). In a separate documented case in Florida, a woman spent two days in jail after her ex-boyfriend allegedly submitted AI-fabricated text messages as evidence of a protective order violation; prosecutors eventually dropped the charges after eight months of proceedings (NCSC, 2026). In Matter of Weber (2024), an expert witness used Microsoft's Copilot AI chatbot to generate opinions, was unable to explain the prompts used or the model's reasoning, and the court questioned the admissibility of the resulting testimony (Secretariat International, 2025). These cases establish that the legal implications of AI-generated evidence are no longer hypothetical; they are actively being litigated in courts today, in the absence of a coherent and uniform regulatory framework.

 

2. How Artificial Intelligence Can Generate Evidence. Prosecution, Defence, and Their Counsel

2.1 AI Evidence Generation by the Prosecution

For prosecution teams and law enforcement agencies, AI offers a range of evidence-generating capabilities that did not exist a decade ago. Facial recognition software can match a suspect's image against databases of millions of photographs drawn from surveillance footage, social media platforms, and identity documents, producing an identification that a human analyst might never have achieved (Gless, 2020). Predictive analytics can process mobile phone location data, financial transaction records, and communication metadata to construct detailed timelines and pattern-of-life analyses demonstrating where a defendant was, who they communicated with, and what they purchased in the days surrounding an alleged offence. AI voice analysis tools can claim to identify stress markers or deception indicators in recorded speech, though the scientific reliability of such tools is actively contested (Justice Speakers Institute, 2025).

Generative AI introduces a further, more legally fraught category of prosecutorial evidence. AI models can reconstruct degraded or incomplete recordings, enhancing audio to make inaudible words audible, sharpening blurred images, or extrapolating missing frames from video. While these enhancements may assist in uncovering what genuinely occurred, they also introduce a layer of algorithmic interpretation that may not accurately represent reality. As University of Colorado researcher Ristovska observed in 2025, "AI enhancement may exacerbate already existing inequalities in access to justice," since not all parties can afford the same enhancement technologies, and judges apply different admission standards to enhanced media (University of Colorado Boulder, 2025). At the furthest and most ethically alarming extreme, the same generative AI tools can be used to fabricate evidence, a capability whose emergence in real proceedings, as documented in the Mendones case and the Florida text message case, demands the most urgent regulatory attention.

2.2 AI Evidence Generation by the Defence

Defence counsel and defendants have equal access to the same AI capabilities, and the adversarial structure of common law proceedings means that AI evidence generation is not exclusively a prosecutorial tool. Defence teams can deploy AI to challenge prosecution evidence: feeding surveillance footage through competing facial recognition systems to demonstrate that a match was erroneous, using AI-powered audio analysis to argue that a recorded confession was obtained under coercion, or subjecting digital forensic evidence to automated integrity analysis. These are legitimate and potentially justice-enhancing uses of AI technology that reflect the adversarial system's reliance on both parties having the resources to contest evidence.

More controversially, defence counsel has begun invoking what practitioners and legal scholars have termed the "deepfake defence": arguing that genuine prosecution evidence, a video, an audio recording, or a digital photograph, is in fact an AI fabrication, even when it is authentic (Dixon, 2024; Delfino, 2023). As Judge Herbert Dixon of the Superior Court of the District of Columbia observed in an American Bar Association publication: "in the absence of a uniform approach in the courtroom for the admission or exclusion of audio or video evidence where there are credible arguments on both sides that the evidence is fake or authentic, the default position, unfortunately, may be to let the jury decide" (Dixon, 2024). This strategy, exploiting the public's uncertainty about what AI can and cannot fabricate to undermine otherwise reliable evidence, represents a structural threat to the fact-finding function of criminal proceedings. In a 2023 Tesla wrongful death lawsuit, defence counsel attempted, unsuccessfully, to dismiss authentic video evidence by claiming it was a deepfake, an episode that illustrated how quickly the deepfake defence was moving from theoretical concern to courtroom tactic (University of Colorado Boulder, 2025).

2.3 AI as a Strategic Tool for Lawyers in Civil Litigation

Beyond criminal proceedings, AI-generated evidence is increasingly deployed in civil litigation. Corporate plaintiffs in intellectual property cases have used AI pattern-matching to demonstrate systematic copying of copyrighted material across millions of documents. Personal injury plaintiffs have introduced AI-reconstructed accident simulations as evidence of negligence. Employers defending discrimination claims have submitted algorithmic workforce analyses as evidence of objective hiring practices. In each of these contexts, AI evidence offers the party with access to it a potentially decisive advantage and raises, with equal force, all of the reliability and authenticity questions discussed below. Grimm et al. (2023) identified in the Duke Law and Technology Review the convergence of AI as both an evidence-generating and an evidence-authenticating tool as creating a recursive credibility problem: courts cannot easily evaluate AI-generated evidence without AI tools, but the reliability of those evaluation tools is itself contested.

 

3. The Reliability of AI-Generated Evidence

3.1 The Reliability Problem. Black Boxes and the Daubert Framework

Reliability is the foundational evidentiary question for AI-generated evidence. Under the amended Federal Rule of Evidence 702, effective December 1, 2023, the proponent of expert testimony must demonstrate by a preponderance of the evidence that the testimony is "based on sufficient facts or data," is "the product of reliable principles and methods," and that the expert "has reliably applied the principles and methods to the facts of the case" (Federal Rules of Evidence, 2023 Advisory Committee Notes). The Daubert factors- stability, peer review, known error rates, and general acceptance in the relevant scientific community- provide the judicial framework within which this reliability requirement is assessed (Daubert v. Merrell Dow Pharmaceuticals, Inc., 1993).

AI models present a fundamental challenge to each of these factors. Many state-of-the-art machine learning models, particularly deep neural networks, are not fully testable in the Daubert sense: their internal logic cannot be independently verified or replicated by opposing experts, particularly when the model's weights and training data are protected as trade secrets (Justice Speakers Institute, 2025). The error rates of AI systems vary significantly across different populations and deployment contexts. A system that performs with 95% accuracy on its training data may perform considerably worse on populations underrepresented in that data (Angwin et al., 2016). General acceptance within the scientific community is contested for most forensic AI applications; the 2023 report from the National Institute of Standards and Technology (NIST) found significant variation in the accuracy and demographic performance of facial recognition systems across different vendors and deployment environments (NVDRS Courts Materials, 2023). These gaps mean that the proponent of AI evidence faces a genuine evidentiary burden that, in many cases, the current legal framework provides insufficient guidance on how to discharge.

3.2 Reliability in Perspective. The Jury and the Opposition

The reliability concerns of AI-generated evidence affect not only judicial gatekeeping but also the ability of opposing parties and juries to meaningfully evaluate the evidence presented. A jury presented with an AI-generated probability score, a facial recognition identification with a 94.7% confidence rating, or a predictive algorithm's assessment of a defendant's future dangerousness faces a fundamental asymmetry of comprehension: the technical complexity of the evidence is beyond the evaluative capacity of most lay jurors (Gless, 2020). This creates what scholars have called the "automation bias" problem, the tendency of human decision-makers to accord excessive weight to algorithmic outputs simply because they appear precise and objective (Delfino, 2023).

For the opposing party, the reliability challenge is compounded by opacity and resource asymmetry. When prosecution evidence is generated by an AI system whose source code and training data are proprietary, the defence may have no meaningful way to challenge the methodology, a situation that raises serious due process concerns under the Confrontation Clause of the Sixth Amendment (Journal of Law, Technology and the Internet, Case Western Reserve, 2024). In State v. Loomis (2016), the Wisconsin Supreme Court acknowledged that the defendant had no access to the proprietary details of the COMPAS algorithm used to assess his recidivism risk, yet upheld the sentence. The court's reasoning that the algorithm was one of multiple factors considered and that the defendant had the opportunity to challenge the score's weight has been widely criticised by legal scholars as inadequate to address the structural informational asymmetry between the parties (UCLA Law Review, 2019).

 

4. Admissibility. Law, Morality, and the Gatekeeping Function

4.1 The Legal Framework for Admissibility

The admissibility of AI-generated evidence in United States federal courts is governed primarily by the Federal Rules of Evidence, as interpreted through the Daubert line of cases. Under FRE 901, evidence must be authenticated; the proponent must produce sufficient evidence to support a finding that the item is what it purports to be. Under FRE 702, expert testimony must satisfy reliability and relevance requirements. Under FRE 403, even relevant evidence may be excluded if its probative value is substantially outweighed by the danger of unfair prejudice, confusing the issues, or misleading the jury. All three rules apply with particular force to AI-generated evidence, where authenticity, reliability, and the risk of jury confusion are simultaneously elevated concerns (Quinn Emanuel, 2025).

The legislative landscape is evolving in response. The 2023 amendment to FRE 702 clarified the proponent's burden of proof, requiring demonstration of reliability by a preponderance of the evidence, a previously ambiguous standard (Federal Rules of Evidence Advisory Committee, 2023). At the state level, California's Senate Bill SB 970 of 2024 established standards for identifying falsified AI evidence and directed the Judicial Council to review by 2026 how AI affects evidence admissibility and to develop applicable rules (Quinn Emanuel, 2025). Proposed Federal Rule of Evidence 707, currently under development, would condition the admission of AI-generated or AI-processed evidence on independent corroboration and a qualified expert's testimony that the system was rigorously tested and has not been exposed to variables likely to cause material inaccuracy (Quinn Emanuel, 2025). The European Union's AI Act, adopted in 2024, introduces mandatory transparency, human oversight, and data quality requirements for high-risk AI systems deployed in law enforcement and criminal justice contexts, a framework that is more prescriptive than current American law (Vision Factory, 2025).

4.2 The Moral Case For Admissibility. Justice, Greater Good, and the Problem of the Habitual Offender

The moral argument for admitting AI-generated evidence rests on a powerful and emotionally resonant premise: the criminal justice system currently suffers from a profound and unjust asymmetry between what investigators and prosecutors know and what they can prove in court. This asymmetry is experienced most acutely in cases involving powerful, resourceful offenders who have the money to retain skilled defence counsel, the influence to compromise witnesses, the resources to destroy physical evidence, and the social capital to benefit from investigative deference that less privileged defendants do not receive. For such defendants, the current evidentiary system does not deliver justice; it delivers impunity. From a utilitarian perspective grounded in Bentham's (1789/1970) harm principle and Mill's (1863/1991) calculus of the greatest good, evidence that can pierce the veil of wealth-enabled evasion serves the collective good, even if its collection methodology is imperfect.

Consider a hypothetical scenario that reflects a pattern documented across organised crime prosecutions, corruption cases, and white-collar criminal proceedings: an individual widely known within law enforcement, the legal community, and the broader public to be the architect of systematic criminal conduct, bribery, violence, drug trafficking, and financial fraud, who has been prosecuted multiple times without success. Each prosecution has failed: witnesses have recanted, physical evidence has disappeared, digital records have been destroyed, and expert testimony has been neutralised by well-resourced counter-experts. The community knows the truth. Law enforcement knows the truth. The jury, each time, returns a not-guilty verdict because the standard of proof requires certainty beyond a reasonable doubt, and the defendant's resources have ensured that reasonable doubt was always available for purchase. Into this scenario, the prosecution introduces AI-generated evidence: a reconstructed pattern of financial transactions demonstrating the flow of criminal proceeds, an AI-enhanced audio recording that brings previously inaudible incriminating words into clarity, or a predictive analysis of communication metadata that places the defendant at the centre of the criminal enterprise. The moral argument is that the introduction of this evidence corrects a systemic injustice. The law's prohibition on evidence tampering, this argument goes, should not be a shield behind which the powerful escape the consequences of crimes that the less powerful cannot.

This moral argument carries real weight, and it is important to acknowledge it with full seriousness before identifying its fatal flaw. The philosophical tradition of legal moralism, as articulated in the work of Devlin (1965) and updated by legal scholars including Dworkin (1977), holds that law serves a moral end, and that procedural rules must be evaluated not merely as abstract constraints but against the substantive justice they are designed to achieve. When a procedural rule consistently produces outcomes that the community recognises as unjust, pressure for reform, including pressure to admit evidence that would not previously have been admitted, is a legitimate response. This pressure is real and politically significant. It has driven the adoption of AI-assisted prosecutorial tools across the United States and internationally, and it will continue to drive further adoption regardless of academic objections, because it reflects a genuine experience of injustice.

4.3 The Legal Counterargument. Why the Law Cannot Follow the Moral Argument Here

The legal counterargument to the moral case for AI evidence is, however, decisive and cannot be dissolved by appeals to outcomes. The Fourth, Fifth, and Sixth Amendments to the United States Constitution and their equivalents in constitutional systems worldwide establish procedural rights not as technicalities but as foundational guarantees against the abuse of state power. The right to be free from unreasonable search and seizure, the right against self-incrimination, the right to confront adverse witnesses, and the right to due process are not contingent on the defendant's guilt. They apply with equal force to the manifestly guilty and the manifestly innocent because the alternative system, in which procedural rights are suspended when guilt is sufficiently obvious, is a system in which there are no rights at all, only prosecutorial discretion (Dworkin, 1977).

The practical legal principle at stake is simple but non-negotiable: a court that admits fabricated or manipulated AI evidence against a genuinely guilty defendant has not merely done something procedurally irregular; it has created a precedent under which the same tools can be used against innocent defendants. The legal machinery cannot distinguish, in advance of trial, between the guilty and the innocent. If the prosecution is permitted to use AI to construct or modify evidence when everyone "knows" the defendant is guilty, there is no principled basis for refusing the same permission when the prosecutor "knows" the defendant is guilty and is wrong. As Delfino (2023) argued in the Hastings Law Journal, the threat posed by deepfakes and AI-generated evidence to the justice system is not merely the risk that a guilty person goes free; it is the risk that an innocent person is convicted on fabricated evidence that cannot be meaningfully challenged, and that the institution of the court has lent its authority to that conviction.

5. Case Study. The Powerful Repeat Offender and the Temptation of AI-Fabricated Evidence

The scenario most commonly invoked in popular and legal discourse to justify tolerating AI evidence manipulation involves a figure that can be described as the "untouchable offender", a powerful criminal actor whose guilt is socially acknowledged but legally unprovable. The archetype is familiar from both real prosecutorial history and fiction: the organised crime boss whose subordinates take all legal risks, the corrupt politician whose transactions are laundered through legitimate structures, the white-collar fraudster whose counsel is more sophisticated than the investigators. For this figure, AI-fabricated prosecution evidence is tempting precisely because it appears to correct a genuine injustice at no visible cost to a sympathetic victim.

This scenario is not entirely hypothetical. United States federal prosecutorial history includes documented instances in which overzealous prosecutors submitted manipulated or unreliable forensic evidence against defendants whom they genuinely believed to be guilty. The FBI's forensic laboratory scandal of the 1990s,  in which microscopic hair analysis, serology evidence, and other forensic outputs were found to have been improperly analysed or presented, affected potentially thousands of criminal convictions, including many of individuals who may indeed have been guilty of the charged offences (National Academy of Sciences, 2009, as cited in Gless, 2020). The lesson drawn by the National Academy of Sciences and by subsequent forensic reform efforts was not that manipulated evidence should be tolerated when the defendant is genuinely guilty; it was that the systems for producing, certifying, and challenging forensic evidence must be made robust enough that manipulation is detected and punished, regardless of the defendant's underlying culpability. 

In the context of AI-generated evidence, the same logic applies with greater force. The Mendones case demonstrated that even unsophisticated AI fabrications can reach courts and be submitted as evidence. If prosecutorial resources and sophistication are applied to AI fabrication as they could be in a case against a powerful organised crime figure, the resulting fabrication would be considerably more difficult to detect. A jury confronted with a seamlessly fabricated audio recording of a defendant issuing criminal instructions, enhanced to the quality standard of authentic professional audio, possessing no obvious digital artefacts, and supported by expert testimony about its authenticity, may have no practical means of evaluating its truthfulness. The conviction that follows is legally and morally worthless, not because the defendant is innocent, but because the system has substituted manufactured certainty for proven guilt, and the same system will then use the precedent of that conviction to justify the next one.

 

6. Risks of Bias and Manipulation in AI-Generated Evidence

6.1 Structural Algorithmic Bias. The COMPAS Paradigm

The risk of bias in AI-generated evidence is not hypothetical;  it has been empirically documented in the most widely used AI tool in American criminal justice. The COMPAS recidivism algorithm, as analysed by the ProPublica investigative team in their landmark 2016 study of over 10,000 criminal defendants in Broward County, Florida, was found to produce systematically racially biased outputs: Black defendants were almost twice as likely as white defendants to be incorrectly labelled as high risk of recidivism, while white defendants were more likely to be incorrectly labelled low risk (Angwin et al., 2016). Northpointe, the algorithm's developer, disputed the characterisation of this disparity as bias, arguing that the algorithm's overall accuracy rate of approximately 61% was equal across racial groups, a defence that ProPublica rebutted by demonstrating that the distribution of errors was systematically discriminatory even when overall accuracy was equated (Angwin et al., 2016).

The UCLA Law Review's 2019 analysis of the COMPAS controversy identified what scholars have termed the "impossibility of fairness": it is mathematically impossible for a predictive algorithm to simultaneously satisfy all major definitions of statistical fairness when the base rates of the predicted outcome differ across groups (UCLA Law Review, 2019). This is not a software bug; it is a structural property of the mathematics of prediction under conditions of social inequality. Training AI systems on historical criminal justice data that reflects decades of racially disparate policing, prosecution, and sentencing produces models that encode those disparities as predictive signals, generating a feedback loop in which the AI's outputs reinforce the inequalities embedded in its training data (UCLA Law Review, 2019). The implications for AI-generated evidence in all forms, not merely recidivism scores, but facial recognition, behavioural analysis, and pattern-of-life modelling, are profound: a biased training dataset produces biased evidence, and the authority of the algorithmic output obscures the human bias at its origins. 

6.2 Facial Recognition and the Risk of Wrongful Identification

Facial recognition technology has become one of the most contentious AI tools in criminal proceedings. A 2019 study by the National Institute of Standards and Technology (NIST) evaluated 189 facial recognition algorithms and found significant variation in accuracy across demographic groups: error rates for Black and Asian faces were between 10 and 100 times higher than for white faces on one-to-one verification tasks for many of the algorithms tested (NIST, 2019, as cited in NVDRS Courts Materials, 2023). Despite this documented unreliability, facial recognition outputs have been used in police identifications and prosecutorial evidence submissions in numerous American jurisdictions, in several documented cases contributing to the wrongful identification of Black men for serious crimes (NCSC, 2026; Gless, 2020). The structural problem is identical to the COMPAS problem: AI systems trained on unrepresentative data produce outputs that are less reliable for the populations most frequently subjected to law enforcement scrutiny.

6.3 The "Deepfake Defence" and the Erosion of Trust in Authentic Evidence

The proliferation of AI-generated evidence has produced a secondary risk of bias that operates in the opposite direction from algorithmic discrimination: the systematic delegitimisation of authentic evidence. Delfino (2023) identified this as one of the most serious long-term risks posed by deepfakes in legal proceedings. As generative AI tools become more widely known and more technically sophisticated, defence counsel can invoke the possibility of AI fabrication to undermine any digital evidence, such as photographs, videos, audio recordings, and digital documents, regardless of their authenticity. The "deepfake defence," as Dixon (2024) characterised it, exploits the lay jury's inability to distinguish authentic from fabricated digital evidence to introduce reasonable doubt about materials that are in fact genuine.

This dynamic creates a paradoxical outcome in which AI undermines both the justice available to victims of genuinely guilty defendants and the protection available to defendants against fabricated evidence. A domestic violence survivor's authentic video of an assault becomes contestable as a possible deepfake. A whistleblower's genuine recording of a bribe becomes dismissible as an AI fabrication. The degradation of trust in digital evidence is not a neutral development; it systematically advantages defendants with the resources to hire experts who can plausibly challenge digital evidence's authenticity, and it disadvantages parties who lack those resources. As the University of Colorado Boulder's 2025 analysis found, "if this continues, jurors will accord little or no weight to authentic footage that they really should be paying attention to" (University of Colorado Boulder, 2025).

 7. AI as a Double-Edged Instrument: Justice and Its Perversion

7.1 The Case for AI Evidence as an Equaliser Against Wealth-Enabled Evasion

There is a genuine, defensible argument that AI-generated and AI-enhanced evidence can serve as a corrective against the systemic advantage that wealth and social capital confer on defendants in criminal proceedings. Evidence destruction is not a theoretical concern: in organised crime, corporate fraud, and political corruption prosecutions, the systematic elimination of physical evidence, the intimidation of witnesses, the corruption of investigators, and the retention of counsel capable of exploiting every procedural opportunity is a documented and recurring obstacle to justice. AI forensic tools that can reconstruct deleted communications, detect financial transaction patterns invisible to human analysis, or extrapolate missing audio from degraded recordings represent capabilities that prosecutors did not previously possess, and that may be necessary to prosecute actors who have specifically structured their conduct to evade conventional evidence standards.

The broader utilitarian argument that the deployment of AI evidence tools reduces impunity for powerful offenders and thereby advances deterrence and social justice is not without intellectual merit. Bentham's (1789/1970) consequentialist framework supports the view that if the introduction of AI evidence significantly increases the conviction rate of genuinely guilty powerful offenders, while the risk of wrongful conviction from that evidence can be adequately managed through judicial gatekeeping and adversarial challenge, the net social benefit may justify its admission. This argument is strengthened by empirical evidence that the current evidence framework systematically underperforms in prosecuting complex economic and organised crime: UNODC (2020) estimates that less than 1% of illicit financial flows are confiscated globally, in part because the evidentiary standards for financial crime prosecution are high and the financial sophistication of major offenders is matched by their legal resources.

The moral force of this argument is real, and it should not be dismissed simply because it sits in tension with procedural values. However, its implementation requires a framework capable of maintaining the distinction between AI-assisted evidence analysis, which extracts or clarifies information genuinely present in authentic data and AI-generated evidence fabrication, which invents or constructs information that did not previously exist. No satisfactory technical or legal mechanism for maintaining this distinction in all cases currently exists, which means that the utilitarian argument for AI evidence, though valid in principle, cannot yet be safely operationalised without risking the abuses examined in section 7.2.

7.2 AI as a Weapon for the Powerful Against the Innocent

The same AI capabilities that could be used to pierce the shield of wealth-enabled evasion can, with equal technical sophistication, be weaponised by the wealthy against competitors, political opponents, whistleblowers, and inconvenient individuals. This is not a speculative risk: it is the structural mirror image of every prosecutorial misuse of AI evidence. A corporation facing a whistleblower's civil claim can commission a deepfake of the whistleblower making fabricated statements that discredit her testimony. A political operative with access to sophisticated AI tools can generate fabricated financial records implicating a rival in corruption, submitting them through civil litigation as a mechanism for reputational destruction, even if the underlying criminal claim is subsequently dismissed. A business competitor can use AI-generated evidence to initiate regulatory proceedings against a rival, the reputational and financial cost of the proceedings themselves constituting the intended harm, regardless of the legal outcome.

Linna et al. (2025), in their analysis of how judges should manage AI-generated material in national security cases, acknowledged that the courts face two simultaneous threats: "first, that evidence presented is AI-generated and not real and, second, that other evidence is genuine but alleged to be fabricated." Both threats have already materialised in documented proceedings. The Florida text message case, in which fabricated AI-generated messages led to a woman's arrest and eight months of legal proceedings before charges were dropped (NCSC, 2026), illustrates how straightforwardly AI evidence fabrication can be weaponised by private individuals with modest technical resources to destroy another person's life. For individuals with access to state resources, corporate legal departments, or organised criminal infrastructure, the same capability, scaled up, represents a mechanism of targeted persecution against which the current legal framework offers insufficient protection.

The deepfake deployed in the 2023 Chicago mayoral election, in which an AI-generated video falsely depicted candidate Paul Vallas making statements he never made, illustrates the same dynamic in the political domain (Linna et al., 2025). If a deepfake can sway an election, it can equally corrupt a trial. The convergence of political, economic, and criminal incentives to fabricate AI evidence against specific individuals makes the development of robust detection and authentication requirements not merely a technical aspiration but a precondition of the continued legitimacy of adversarial legal proceedings.

 

8. Toward a Coherent Legal Framework: Recommendations

The legal system's response to AI-generated evidence must address three levels simultaneously. At the legislative level, the proposed Federal Rule of Evidence 707  requiring independent corroboration and qualified expert testimony establishing the reliability of AI systems used to produce evidence represents an important step, but its passage should be accompanied by mandatory disclosure obligations that prevent trade secret protection from shielding algorithmically generated prosecution evidence from adversarial challenge (Quinn Emanuel, 2025). At the judicial level, the Daubert gatekeeping function must be applied more rigorously and consistently to AI evidence, with courts requiring proponents to disclose training data demographics, known error rates across relevant population subgroups, and testing methodology before admission is granted (Justice Speakers Institute, 2025). At the institutional level, the American Bar Association (2024) and the National Centre for State Courts (2026) have both recommended the development of specialised judicial guidance bench cards, AI task forces, and mandatory judicial education to equip judges with the technical literacy necessary to evaluate AI evidence claims without depending exclusively on competing expert witnesses paid for by the parties.

Authentication standards for AI-generated or AI-enhanced digital media require urgent reform. The ABA Advisory Committee on Evidence Rules' proposed Rule 901(which would shift the burden of proof to the proponent of evidence challenged as a likely deepfake, requiring demonstration that probative value outweighs prejudicial effect) addresses the deepfake defence problem directly and should be adopted (Dixon, 2024). California's Senate Bill SB 970 (2024) directive requiring the Judicial Council to develop AI evidence rules by 2026 provides a model that other states and the federal system should follow. Internationally, the EU AI Act's (2024) requirements for transparency and human oversight of high-risk AI systems in criminal justice contexts establish a benchmark against which American law is currently deficient.

The ultimate challenge, however, is not legislative or technical; it is cultural. The legal system's integrity depends on the presumption that evidence represents reality, not an algorithmically constructed version of it. CNBC's 2024 reporting on AI deepfakes in legal proceedings cited a 2023 Pew Research Centre survey finding that public trust in the US justice system stood at a historic low of 44% (CNBC, 2024). Introducing AI-generated evidence into this environment of diminished public trust, without a framework robust enough to guarantee authenticity, risks accelerating the erosion of the institution's legitimacy. As the NCSC (2026) has argued, AI tools can offer genuine benefits to courts but only if the framework governing their use maintains the court's truth-seeking function as the non-negotiable priority around which all other considerations are organised.

 9. Conclusion

Artificial intelligence has arrived in the courtroom without a rulebook adequate to manage it. The legal system that confronted the admissibility of DNA evidence in the 1980s had the advantage of a clear scientific consensus about what DNA profiling could and could not establish; the scientific community, the legal academy, and ultimately the courts converged on a framework that has served justice reasonably well. AI-generated evidence does not offer the same clarity. It encompasses a range of technologies from recidivism algorithms to facial recognition to generative deepfake video. The reliability of these technologies varies enormously, and their biases are empirically documented; their potential for weaponisation by both state and private actors is structurally unconstrained by current law.

The legal implications examined in this paper converge on a single conclusion: the reliability, admissibility, and manipulation risks of AI-generated evidence are not merely technical problems to be solved by better algorithms. They are political and moral problems about who gets to construct legal reality, whose interests are served by the courts' fact-finding function, and whether the procedural protections that form the backbone of due process can survive the democratisation of sophisticated fabrication technology. The framework needed to address them must be developed urgently, comprehensively, and with full awareness that the stakes are not abstractions; they are the liberty and lives of the people who appear before courts in the most vulnerable circumstances of their lives.

 References

American Bar Association. (2024). Artificial intelligence and the courts. ABA Centre for Innovation. https://www.americanbar.org/groups/centers_commissions/center-for-innovation/artificial-intelligence/ai-courts/

Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016, May 23). Machine bias: Risk assessments in criminal sentencing. ProPublica. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing

Bentham, J. (1970). An introduction to the principles of morals and legislation (J. H. Burns & H. L. A. Hart, Eds.). Athlone Press. (Original work published 1789)

CNBC. (2024, March 6). AI deepfakes are poised to enter court proceedings at a time of low trust in the legal system. https://www.cnbc.com/2024/03/06/ai-deepfakes-are-headed-to-court-at-time-of-low-trust-in-legal-system.html

Dalal, A., Gao, C., Grimm, P. W., Grossman, M. R., Linna, D. W., Jr., Pulice, C., Subrahmanian, V. S., & Tunheim, J. (2025). Deepfakes in court: How judges can proactively manage alleged AI-generated material in national security cases. University of Chicago Legal Forum, 2024, Article 3. https://chicagounbound.uchicago.edu/uclf/vol2024/iss1/3

Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993).

Delfino, R. A. (2023). Deepfakes on trial: A call to expand the trial judge's gatekeeping role to protect legal proceedings from technological fakery. Hastings Law Journal, 74(2), 293–352.

Devlin, P. (1965). The enforcement of morals. Oxford University Press.

Dixon, H. B., Jr. (2024, June 11). The "deepfake defence": An evidentiary conundrum. Judges' Journal, American Bar Association. https://www.americanbar.org/groups/judicial/resources/judges-journal/2024-spring/deepfake-defense-evidentiary-conundrum/

Dworkin, R. (1977). Taking rights seriously. Harvard University Press.

European Union. (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council on artificial intelligence (EU AI Act). Official Journal of the European Union.

Federal Rules of Evidence. (2023). Advisory Committee Notes to the 2023 Amendment to Federal Rule of Evidence 702. United States Courts.

Frye v. United States, 293 F. 1013 (D.C. Cir. 1923).

Gless, S. (2020). AI in the courtroom: A comparative analysis of machine evidence in criminal trials. Georgetown Journal of International Law, 51(2), 195–253.

Grimm, P. W., Grossman, M. R., Brown, D. G., & Xu, M. (2023). The GPTJudge: Justice in a generative AI world. Duke Law and Technology Review, 23(1). https://scholarship.law.duke.edu/dltr/vol23/iss1/

Hamilton, M. (2019). The biased algorithm: Evidence of disparate impact on Hispanics. American Criminal Law Review, 56(4), 1553–1577.

Illinois State Bar Association. (2025, March). Deepfakes in the courtroom: Problems and solutions. AI Section Newsletter. https://www.isba.org/sections/ai/newsletter/2025/03/deepfakesinthecourtroomproblemsandsolutions

Justice Speakers Institute. (2025, November 22). AI on trial: Admissibility of AI-generated evidence. https://justicespeakersinstitute.com/ai-generated-evidence-admissibility-on-trial/

Journal of Law, Technology & the Internet, Case Western Reserve University. (2024). AI in the courtroom: Confrontation clause implications of AI-generated evidence. JOLTI, 15(2), 416–460.

Linna, D. W., Jr., Dalal, A., Gao, C., Grimm, P., Grossman, M. R., Pulice, C., Subrahmanian, V. S., & Tunheim, J. (2025). Deepfakes in court: How judges can proactively manage alleged AI-generated material in national security cases. Northwestern Law & Economics Research Paper No. 24-18. https://ssrn.com/abstract=4943841

Mendones v. Cushman & Wakefield, Inc., Alameda County Superior Court, California (2024).

Mill, J. S. (1991). Utilitarianism (J. Grey, Ed.). Oxford University Press. (Original work published 1863)

Mnookin, J. L. (1998). The image of truth: Photographic evidence and the power of analogy. Yale Journal of Law and the Humanities, 10(1), 1–74.

National Centre for State Courts (NCSC). (2026, February 24). AI-generated evidence is a threat to public trust in the courts. https://www.ncsc.org/resources-courts/ai-generated-evidence-threat-public-trust-courts

National Violent Death Reporting System (NVDRS) Courts Materials. (2023). Artificial intelligence courts: Materials for judges. United States District Court, District of Nevada. https://www.nvd.uscourts.gov/wp-content/uploads/2023/04/AI-and-Trustworthiness-NIST.pdf

Quinn Emanuel Urquhart & Sullivan. (2025, November 6). Adapting the rules of evidence for the age of AI. Quinn Emanuel Publications. https://www.quinnemanuel.com/the-firm/publications/adapting-the-rules-of-evidence-for-the-age-of-ai/

Secretariat International. (2025, June 12). Statistics tools chart a path for AI use in expert testimony. https://secretariat-intl.com/wp-content/uploads/2025/06/Statistics-Tools-Chart-A-Path-For-AI-Use-In-Expert-Testimony.pdf

State v. Loomis, 881 N.W.2d 749 (Wis. 2016).

UCLA Law Review. (2019). Injustice ex machina: Predictive algorithms in criminal sentencing. UCLA Law Review. https://www.uclalawreview.org/injustice-ex-machina-predictive-algorithms-in-criminal-sentencing/

United Nations Office on Drugs and Crime (UNODC). (2020). Money laundering and globalisation. UNODC. https://www.unodc.org/unodc/en/money-laundering/globalization.html

University of Colorado Boulder. (2025, November 17). Deepfakes and AI in the courtroom: Report calls for legal reforms to address a troubling trend. CU Boulder Today. https://www.colorado.edu/today/2025/11/17/deepfakes-and-ai-courtroom-report-calls-legal-reforms-address-troubling-trend

Vision Factory. (2025, June 11). AI-generated evidence in court: Admissibility, bias, and the future of legal standards (2025 update). https://www.visionfactory.org/post/ai-generated-evidence-in-court-admissibility-bias-and-the-future-of-legal-standards-2025-update