86% fake - 100% admissible? Rethinking evidence in the AI era

This article was originally published on The Legal Intelligencer, Law.com, January 2026. 

From deepfake detection to content authentication, the admissibility of AI-generated and AI-manipulated forensic evidence is no longer a future problem - it’s a present challenge: How do we ensure that forensic methods for AI-detection are legally admissible? Courts are increasingly confronting AI-generated and AI-manipulated evidence land on their dockets. But with innovation comes scrutiny - and for good reason.

In this article, Katarina Zotovic, from the Digital Forensics team at S-RM, and Ashley Pusey and Brian Ramkissoon, specialising in cyber and technology risks at Kennedys Law, delve into the intersection of AI forensics and the legal landscape, exploring both the technical hurdles and legal considerations involved. They discuss not only the challenges of achieving validation and standardisation, but also the ongoing tension between transparency and explainability for court admissibility.

The legal foundation: FRE 41, 401, 403, and 702 

Admissibility is the floor, not the ceiling. The admissibility of relevant evidence is a pretty low bar under the Federal Rules of Evidence (FRE). Under Rule 401, relevance has a low threshold - any evidence is relevant if it makes a fact more or less probable and the fact is of consequence to the action. Rule 402 allows such evidence unless excluded by law. Rule 403 acts as a safeguard, allowing exclusion of relevant evidence if its probative value is substantially outweighed by risks like unfair prejudice, confusion, or undue delay.

Relevant evidence and expert witness testimony are governed by different standards but are both central to admissibility at trial. Expert witness testimony, governed primarily by Rule 702, requires a more rigorous foundation. A qualified expert may testify if their specialized knowledge will help the trier of fact understand the evidence or determine a fact in issue. The testimony must be based on sufficient facts or data, must be the product of reliable principles and methods, and must reflect a reliable application of those methods to the facts of the case. This reliability standard, shaped by the Supreme Court’s Daubert decision, requires the court to act as a gatekeeper and evaluate factors such as whether the expert’s methodology is testable, peer-reviewed, has a known error rate, and is generally accepted in the relevant field. That’s a high bar - and AI tools aren’t always clearing it. For example, a proprietary deepfake detector that outputs a confidence score but offers no audit trail may be challenged under Rule 702. 

The synthetic media paradox

The biggest issue? Synthetic media’s admissibility trap. A convincingly fabricated video may pass the relevance test while being entirely false. This presents a paradox: the most emotionally persuasive evidence may also be the most misleading. Without robust evidentiary checks, AI-manipulated content could mislead judges or juries.

Technical vs. legal reliability: explaining the ‘Black Box’ problem

Many AI forensic detection tools operate like black boxes - technically sophisticated, but legally opaque. If an expert can’t explain how an AI tool reached its conclusion in plain language (or to a judge), it probably shouldn’t be in a courtroom. For example, tools programmed by humans can be explained by them, however this isn’t regularly the case for AI. Experts testifying based on AI tools lack the visibility necessary into the steps the tool is taking, raising questions around the reliability of the results, its conclusions, and repeatability. 

Raising the bar: authentication under Rules 901 and 104

This raises important questions about whether Rule 402’s broad admissibility standard is sufficient in an era of synthetic media, or whether stronger gatekeeping is needed. Rather than amending the rule itself, a more pragmatic solution may lie in reinforcing the court’s role under Rules 901 and 104 to rigorously authenticate digital evidence, especially in contexts where AI manipulation is plausible. As the technology becomes more sophisticated - and indistinguishable from reality - the burden on courts to separate signal from synthetic noise will grow. Without updated evidentiary safeguards, there is a real risk that AI-generated content will be weaponized to mislead, manipulate, and undermine the fairness of judicial proceedings.

Standardising AI forensic methods

Unlike traditional forensics, which are backed by decades of peer-reviewed validation, AI-detection tools lack universally accepted standards. Many methods remain proprietary, increasing the risk of inconsistent results and reduced courtroom defensibility.

Traditional forensic disciplines have been subjected to decades of refinement, ensuring their scientific reliability and legal admissibility. In contrast, AI-detection tools remain an emerging field, where tools and methodologies are often proprietary and can introduce uncertainties in their results. This makes manual validation essential.

Extensive research has been conducted in visual (video and image) and audio forensics, applying forensic methodologies to examine lighting and shadows, facial movements, pixel intensity and spectrogram analysis. While general principles exist for standardising visual and audio forensics, such as identifying common deepfake markers and anomalies in spectrographic tools, increasingly sophisticated AI-generated content makes physical inconsistencies harder to detect. As deepfakes improve, forensic analysts require more robust methods beyond visual/auditory cues alone.

To complement visual and audio forensic techniques, forensic analysis relies on metadata and file structure analysis to verify the authenticity of digital media. Key methods include metadata examination, as well as hexadecimal (hex) and binary analysis. File metadata can serve as a treasure trove in forensics, especially as media files are shared and modified. Analysing timestamps, geolocation information, software used, and device identifiers can help verify the legitimacy of an image, audio or video file. Similarly, hex analysis can uncover subtle alterations that may indicate tampering.

Achieving accepted benchmarks for AI forensic methods requires addressing both the dynamic nature of AI-derived content and the current limitations of existing forensic tools. Integrating detection models with both audio-visual cues and manual analysis may provide the most effective solution for ensuring the reliability necessary for legal validation.

The three pillars for trusting AI evidence

1. Reproducibility

Reproducibility is a fundamental requirement for forensic evidence to be admissible in court. Forensic findings must be replicable, meaning another independent expert should be able to apply the same methodology and arrive at the same results. However, AI-focussed forensic tools introduce challenges. Many deepfake detection tools are still relatively untried and untested, particularly in the courtroom, making it difficult to assess their reliability in regards to legal evidence. Since the detection tools often rely on machine learning models, which are evolving rapidly due to the novelty of the progressing research, variations in results over time can pose serious concerns for reproducibility. For forensic findings to withstand legal scrutiny, experts should prioritise peer-reviewed, independently testable forensic software that allows for transparent validation.

In contrast, the manual analysis of metadata, hex, and binary structure discussed above is part of a well-established, widely accepted forensic framework. Unlike AI-driven detection, these techniques operate at the raw data level, where file structures and forensic markers tend to remain consistent across media formats and decades of digital data. This stability makes traditional forensic methods more reproducible and defensible.

2. Verification and authentication

Courts must also understand and distinguish the purpose for which AI forensics is being used: forensic verification v. forensic authentication.

Forensic verification is the process of confirming AI-generated and AI-manipulated content, allowing experts to determine whether content was generated, altered, or influenced by AI. We will likely see an uptick in the use of such experts in cases such as defamation, fraud, and family law. For example,

Methods used in forensic verification include:

• Analysis of spectrographic features for synthetic voice cloning detection by identifying patterns which are hard for AI to mimic. This encompasses examining pitch, intonation, or speaking rates and using machine learning models to compare against real human speech.
• Stylistic, linguistic, and cohesive checks for AI-generated text. Since AI-generated content frequently lacks the narrative-flow or nuanced context that aligns with human linguistics and etiquette, models are trained on a large volume of both AI-generated texts and human-written texts to determine differences in patterns of structure, perplexity, and readability.
• Scrutinising media, such as videos and images, in deepfake analysis for inconsistencies in physiological markers (e.g., blinking) and audio-visual markers (e.g., shadows). This is most commonly done by tools which detect oddities in textures, colour patterns, uniformity, and spatial anomalies.
• Forensic authentication is used to prove the contents’ origin and integrity. Authentication answers the following key questions which are essential for ensuring that digital evidence can be trusted and is admissible in legal contexts:
• Who created it? Understand the creator of digital content is necessary to establish authorship and accountability. Forensics experts examine artifacts such as digital signatures and embedded metadata to trace the content to its original source, additionally leveraging open-source intelligence and investigative techniques into the creator.
• When was it created? Confirm original creation dates, which serve as digital timestamps in legal proceedings.
• Is it what it claims to be? Ensure that the digital media has not been manipulated in order to misrepresent its nature is necessary to determine and maintain the integrity of the evidential source.
• Has it been altered in any way? If any discrepancies are identified in a file, it is important to understand if this is due to intentional tampering, or legitimate reasons such as file system transfers or formatting issues.

Could there be another explanation?

In terms of legal frameworks for AI forensics, existing computer forensics principles, such as those outlined in NIST SP 800-86, are being adapted to fit the emerging field of AI forensics. However, a universal legal standard has yet to be established. The lack of such standard presents a risk: AI forensic methods could face challenges in court due to the absence of a standardised benchmark.

Recent developments in the field highlight ongoing efforts to address these challenges. Internationally, research into AI forensics is gaining momentum. Initiatives like the European AI Act and various U.S. federal efforts aim to establish guidelines for the admissibility of AI-generated and AI-manipulated evidence in courtrooms. Meanwhile, the training of automated deepfake detection models is advancing, with AI-driven forensic tools continually honing their detection capabilities through extensive datasets.

Moreover, collaborations between technology developers and legal professionals are becoming increasingly important. These partnerships are essential for creating forensic reports that can withstand judicial scrutiny, ensuring that the findings are both robust and reliable in legal contexts. This synergy between tech and legal fields is necessary when seeking a standardised approach to AI forensics and the admissibility of AI-generated evidence in court.

From a technological perspective, many deepfake detection tools output a probability score (e.g., “86% likelihood that this video is a deepfake”). But these scores are often presented without context or alternative explanations. This contrasts with traditional digital forensics, where terms like "likely" are explained with supporting evidence. For example, internet browsing activity may show that an illicit website was accessed immediately after a user opened their email account. This could suggest a high likelihood that the user clicked on a link in an email that redirected them to the site. However, alternative explanations exist, such as a malicious pop-up, an auto-loaded advertisement, or an unintended background process.

In contrast, many AI-based deepfake detection tools fail to account for alternative explanations when assigning probability scores. They may classify a video as 86% likely to be fake without considering whether compression artifacts, post-processing effects, or natural inconsistencies in lighting and motion contributed to the classification. This lack of contextualization makes it more difficult to explain findings and assess the true evidentiary weight of AI-generated forensic conclusions in court

In the age of AI, is there a Call for Evidence and policy reform to ensure digital trust?

Rebuilding trust in digital media is a must if the courts are to effectively rely on and evaluate digital evidence with confidence, ensuring fair and just outcomes in legal proceedings. One thing is clear: Neither lawyers nor technologists can solve this alone. AI and detection tool developers must learn the language of evidentiary standards, while lawyers must understand what makes an algorithm tick. And both need to center ethical values like fairness, transparency, and accountability. So how can we spear this issue: through cross-disciplinary collaboration between AI engineers, forensic analysts, and legal experts to align technical and legal expectations.

Promising collaborations are underway. For example, The Coalition for Content Provenance and Authenticity (C2PA) is building on collaborations between major tech companies to develop technical standards and frameworks for content provenance and authenticity, such as requiring media files to be embedded with provenance data confirming their legitimacy. This cryptographic watermarking embeds information (a "mark") into cryptographic functions or digital content so it can be verified later without revealing the underlying information. Blockchain verification technology is also being used to record the origin and subsequent alteration of media so a provenance trail is in place for each file. This ensures a real-time verification of content authenticity and provides transparency to the changes of media files.

But realizing a trustworthy evidentiary future also requires judicial leadership. Judges may need to embrace a more proactive gatekeeping role early in the litigation process, particularly in cases involving AI-generated or synthetic media. This includes applying Rules 104 and 901 at the outset to evaluate whether forensic conclusions meet basic thresholds for authentication and explanation, not merely relevance under Rule 401. Courts must also consider Rule 902 in determining whether a piece of digital evidence is self-authenticating or demands additional expert support.

By intervening early to scrutinize digital exhibits, judges can prevent unreliable or misleading content from skewing legal outcomes. This proactive stance is crucial for safeguarding fairness and maintaining public confidence in judicial proceedings amid rapidly evolving AI capabilities. Only then can synthetic media be properly managed and the courtroom remain a place for truth - not just technological illusion.