When the evidence thinks for itself: how will the courts keep AI-generated evidence in check?

Many aspects of AI are now commonplace in UK Courts, from computer-assisted disclosure to assisting judges in the preparation of their judgments. The rapid development of AI poses a new challenge, however: how will the courts deal with the inevitable volume of AI-generated material that will be placed before them and how will they distinguish between genuine and synthetic evidence? 

As noted in the recently updated Guidance for Judicial Office Holders, the judiciary is aware that "AI tools are now being used to produce fake material, including text, images and video" but that "Courts and tribunals have always had to handle forgeries […] involving varying levels of sophistication". On any view, however, courts have not faced forgeries of the nature, sophistication and scale now potentially posed. The rapid advancement in AI tools has arguably transformed the scope for forgery from a relatively rare and detectable anomaly into a widely available, existential threat to evidential integrity.

What issues are posed by AI-generated evidence? 

Some of the primary challenges posed by AI-generated evidence are that: 

• Unscrupulous parties may seek to advance AI-generated evidence in support of their claims, which may be, or become, indistinguishable from authentic evidence.
• Parties may increasingly cast doubt on (potentially) genuine evidence by claiming that the evidence was AI-generated, per a tactic coined the "liar's dividend" in political disinformation campaigns, whereby authentic material is dismissed as being fake. 
• Satellite disputes as to authenticity may divert significant costs and court resource despite AI being hailed as a tool for promoting efficiencies and cost savings.
• While technology exists which purports to identify AI-generated evidence, it is not a silver bullet. Not only will such tools be liable to challenges and likely require investigation as to their underlying methodology, they also provide AI with a continuous feedback loop that will allow system-generated evidence to iteratively improve to evade future detection. The resulting arms race between systems to generate versus detect AI-generated evidence will become central to the future of AI-generated evidence before the courts.  

What parallels can be drawn with what the courts see already?

The courts' existing approach to allegations of metadata interference may be instructive for how allegations of inauthentic AI-generated evidence will be addressed by the judiciary. Both issues share common challenges around authenticity and provenance, technical complexity often requiring forensic expert evidence, burden of proof considerations (a party calling such evidence into question would likely be required to provide credible evidence to support such allegations) and the amount of weight that may be placed on evidence where integrity has been questioned (if even it is admitted).

Courts are likely to adopt a cautious approach to suggestions of AI forgeries, heavily informed by expert evidence, just as they have done when addressing metadata tampering allegations. In reality, however, even if the approach to assessing metadata integrity will provide a basic framework for assessing the authenticity of AI-generated material, the development and implementation of a specific framework for AI evidence is likely to be necessary, not least because AI evidence will leave a far less concrete digital trace than metadata. 

When might AI-generated material be used legitimately and when will it amount to evidence? 

AI as a tool

AI-generated material may be legitimately provided to the court (AI-produced summaries of voluminous data that might otherwise require expensive expert analysis, for example). Such outputs will undoubtedly not only aid the judiciary but also potentially assist in democratising litigation, allowing claimants to bring claims and strategically litigate against bigger players with deeper pockets when they might otherwise be restricted in bringing a claim by the costs of expert input or processing vast quantities of highly technical information. 

The output from AI in this way would not, however, amount to evidence itself – instead, in this context, AI would be an evidence and data-processing tool and a human would need to take responsibility for its interpretation. In this way, courts might theoretically treat the output in a similar way to how they have now widely accepted technology-assisted review (TAR) and predictive coding in disclosure exercises (using machine learning algorithms), through for example: 

• welcoming the improved efficiency (provided the process is transparent and proportionate and that human oversight is demonstrated)
• not requiring detailed algorithmic explanations (but rather a demonstration that the workflow is reasonable and defensible)
• expecting human responsibility for decisions (for example in training the system, setting relevance criteria, validating results, confirming accuracy, making evidential decisions and ultimately explaining the process to the court)
• requiring transparency as to its use (potentially including the quality checks and risk mitigations in place). 

In theory, therefore, proportionality, rather than burdensome technical explanations, might sit at the heart of the courts' approach to AI evidence-processing. The reality is likely to look very different, however – in particular, being able to demonstrate human oversight, training and decision-making may be impossible when it comes to using AI as a processing tool. If the courts do in due course expect to interrogate the AI's training, accuracy, and algorithmic biases, it will undoubtedly add complexity to proceedings that AI may have been intended to simplify. Moreover, and as the Guidance for Judicial Office Holders notes, malicious actors can embed invisible 'white text' that is imperceptible to humans but detectable by AI to manipulate outputs through hidden instructions. This means that what may appear at face value to be neutral AI-generated outputs may instead have been covertly skewed, creating a new form of fraud that may be extremely difficult to detect.

AI as the 'expert'

If, on the other hand, an AI system generated opinions from data, it would effectively be akin to an expert rather than merely a processing tool (where a human would ultimately remain responsible for its interpretations) and its output (eg predictions, diagnoses, technical judgements or causation analysis) would resemble expert opinion evidence. 

It is difficult to see how a court would reconcile the requirements of CPR 35 (which covers expert evidence) in such instance – who, for example, is the expert? Is it the AI system itself, the developer, the human who 'ran' it, or someone else? If it was the system, how would that 'expert' satisfy the court that it was qualified, objective and impartial? Absent a human explanation of whether methods were reliable and why conclusions were reached, it is hard to envisage AI "evidence" being admitted or given much weight. 

The courts do of course already deal with machine-generated expertise (in some highly technical and scientific areas), but although the underlying expertise emanates from a system, the output only becomes admissible as 'expert evidence' when a human expert understands it, assumes responsibility for it, and can be cross-examined in relation to it. Without this, the AI output is little more than hearsay from a machine. For now, we seem very far indeed from AI systems becoming experts whose opinions are admissible in their own right. 

How will AI-generated evidence impact the future of disputes? 

Left with the more immediate possibility and challenge of AI-generated fake evidence, we foresee that the following potential impacts may materialise in the near future:  

• Verification statements and expanded disclosure obligations: The updated Guidance for Judicial Office Holders notes that all legal representatives have a professional obligation to ensure that material put before the court is accurate and appropriate and that, provided AI is used responsibly, and depending on context, there is no reason why a representative ought to refer to its use. That said, much as we have seen the introduction of certificates of compliance for lawyers under PD57AC and PD57AD, commentary in the market has mooted the possibility of parties and their legal teams being required to provide statements verifying whether evidence has, or has not, been AI-generated or AI-manipulated, through confirmation wording. 
• Satellite disputes over authenticity: Scepticism as to the authenticity of documents may be tactically weaponised and we expect to see increasing challenges to the authenticity of documents on the basis that they may have been AI-generated. A well-funded defendant, for example, may threaten protracted and expensive satellite AI-authenticity disputes, pressuring an under-resourced claimant. AI-generated evidence thus becomes not merely an evidential issue but risks becoming a tool that undermines both access to justice and cost efficiencies. Rather than democratising and streamlining certain phases of litigation, AI could theoretically risk pricing litigants in person out of disputes entirely and increase the burden on already stretched court resources.
• Additional layers of expert evidence may be required: Parties may require expert evidence to: authenticate their own AI-generated evidence, challenge the opponent's evidence as being fabricated, or even assess the reliability of the opponent's AI-detection methodology. Each layer of expert witness evidence will inevitably increase the time and costs of proceedings. Technology which could theoretically in future reduce reliance on expensive experts may therefore instead create entirely new categories of expert work to factor into litigation as a matter of course. Even where expert evidence is technically possible, and affordable for a party seeking to validate the output of an AI system, experts may struggle to obtain the evidence they require to properly analyse the AI system given that the data, such as its technical training information, is often protected by trade secrets. We expect this to become progressively challenging for the judiciary to navigate and harder for them to find in favour of a party, or place weight on the evidence they are advancing, on the basis on what experts are able to put forward. In turn, this risks undermining justice. 
• Judicial training and the knowledge gap: Judges might be required to assess highly technical evidence about AI generation, detection methodologies, algorithmic processes, and digital forensics, and will undoubtedly require specialist training. Specialist judges from the Patents Court or Technology and Construction Court may perhaps be required to hear satellite AI evidence disputes. The judiciary will clearly require substantial ongoing training to keep pace with rapidly evolving AI capabilities and new techniques relating to AI-generated evidence. The Guidance for Judicial Office Holders' suggestion that judges should be 'aware of' the new possibility of, and challenges potentially posed by, deepfakes is unlikely to even scratch the surface of what's around the corner. 
• Changes to evidence hierarchies: Video and audio recordings, once considered among the most probative forms of evidence, may become perceived as less reliable given the rise in highly sophisticated deepfake evidence. AI may even be used to create contemporaneous-appearing evidence in real time (such as the generation of fake emails with metadata and server logs which appear accurate). Conversely, less technologically-sophisticated evidence such as handwritten notes, physical documents, or contemporaneous photographs may acquire renewed value precisely because they are harder to fabricate convincingly at scale. 

As cited in the Law Commission's 2025 Discussion Paper, even where the technical details of AI systems are available, AI systems are challenging even for AI experts to explain as they are based on mathematical models of enormous scale and, as leading AI developer Anthropic has stated, "a surprising fact about modern large language models is that nobody really knows how they work internally". This fundamental opacity suggests that the key challenge that judges, lawyers and parties will grapple with, and which will shape the future of disputes, is how courts and tribunals, predicated on establishing facts and evidence, will identify or even accommodate evidence from processes that are, by their creators' own admission, inherently inexplicable.