What holds up in court? Alan Parfery on AI, evidence and the discipline of the courtroom

His answer was not anti-technology. If anything, it was more hopeful than some of the surrounding debate. The central point was that courts can and should make use of AI, but only where appropriate checks are in place and where human judgement remains central. That theme has only become more concrete since the conference: the Lady Chief Justice has since said that judges in England and Wales are being trained to recognise both the use of AI in courtrooms and the risks that new technologies may pose to the administration of justice.

Context and background

Parfery began by drawing a distinction between the growing use of AI outside court and its more limited role inside it. In legal practice generally, AI tools are already being used to draft, summarise, structure arguments and accelerate research. In court, though, things are less forgiving. If there is a weakness in the reasoning, the source material or the evidential chain, someone on the other side is likely to find it.

That point was tied to professional duties. As a prosecutor handling the most serious criminal matters, including homicide cases, Parfery said his worst nightmare would be prosecuting a case that resulted in a miscarriage of justice. That line was important because it framed the speech in practical rather than theoretical terms. AI is not just a tool for speed or efficiency. It has to be assessed against the duties lawyers owe to the court.

He also made a wider institutional point. The courtroom is not a place where AI will simply arrive and dictate its own terms. Courts already have established mechanisms for testing expert evidence, excluding material that goes too far and insisting on accountability for what is put before them. That meant his speech, while cautious, was not gloomy. The justice system is not helpless in the face of these technologies.

The courtroom as a stress test

One of the strongest features of the speech was the way Parfery treated the courtroom itself as the ultimate stress test. Outside court, a tool can look impressive because it is fast or superficially persuasive. Inside court, that is not enough, and, more importantly, the output has to stand up to scrutiny within the rules that govern the admissibility of evidence and procedural fairness.

He illustrated that with a Scottish homicide appeal involving an expert report that modelled the probability of a death having occurred by accident. The defence position was that the fatal injury had been caused accidentally during a struggle. The report concluded that there was only a 0.25% probability of that account being correct. Despite the apparent force of the report, the appeal court held that it was inadmissible because it usurped the function of the jury by assigning a probability to whether the accused was telling the truth.

That example raised interesting evidential questions in its own right. But Parfery’s real point was broader. The courts are capable of pushing back. The issue is not that the courts lack the tools to deal with AI; it is that they will need to use those tools carefully and consistently.

Augmentation, not substitution

Parfery was equally clear that AI can have a positive role. The most compelling example he gave was Sheriff Alastair Carmichael’s use of AI-generated voice technology following a diagnosis of motor neurone disease. That technology allowed the sheriff to continue giving directions to juries in a version of his own voice.

The significance of the example was that the technology did not replace judicial thought or decision-making. It simply enabled the judge to continue carrying out that role. That is a useful way of thinking about AI in justice more generally. It does not need to replace legal actors to be transformative. It can facilitate and assist human participation.

That point sat neatly with a wider conference theme. Some of the strongest use cases for AI are the more practical ones: transcription, accessibility, document management, summarisation and other tools that help people do their jobs better without displacing responsibility for judgement.

At the same time, Parfery’s examples showed why it is difficult to draw bright lines between safe and unsafe uses. It is tempting to say that administrative uses are acceptable while substantive uses are not. The line between helping with a task and influencing the outcome is not always obvious. That matters because debates about AI often try to fall back on neat labels. Parfery’s speech underlined the importance of supervision, transparency, testing and verification, rather than in assuming that one label makes a use case inherently safe.

Duties, hallucinations and authenticity

Professional responsibility therefore remained central to the speech. Parfery addressed hallucinated authorities directly, along with the reputational and ethical risks of relying on invented material. However useful AI may be in preparation, the duty to check remains with the lawyer.

He also spoke about deepfakes and synthetic media, using the example of a finance worker deceived into making a large transfer during a video call involving a fake chief financial officer. The point was that these issues are no longer hypothetical. If they are already occurring in the real world, they will inevitably enter legal proceedings too. The challenge is therefore bigger than just the prospect of false legal research. It is also a wider warning that lawyers and courts need to be ready for arguments about whether evidence is what it appears to be, and how far it can be trusted. That concern is no longer limited to deepfakes in the abstract: in March 2026, a High Court judge found that a witness had been assisted during cross-examination through smart glasses, a reminder that questions of authenticity, coaching and hidden technological interference are already arriving in live proceedings.

Why human judgement remains central

Parfery closed with a story from a murder trial he had prosecuted, where a witness unexpectedly repeated a devastating confession made by the accused shortly after the killing. Parfery used the story to make a broader point about how courts actually work. Cases are still decided by people listening to witnesses, testing evidence, making submissions and exercising judgement in real time.

That is also why the speech felt more positive than anxious. His point was that human judgement still matters and that the courts are already built to deal with this kind of problem. They are used to testing evidence, pushing back on material that looks too neat or too definitive, and controlling what is and is not admitted. On that view, caution around AI-generated evidence is not a sign that the system is falling behind. It is part of the discipline that should allow the system to adapt without losing its bearings.

Practical takeaways

• Courts are likely to welcome AI most readily where it assists rather than replaces human judgement.
• The line between “administrative” and “substantive” use is less clear than it first appears.
• Hallucinations and synthetic evidence are not just technical issues; they go to professional duties and evidential integrity.
• The courtroom’s habits of scrutiny and challenge are a strength, not an obstacle.
• AI will not displace the courtroom’s basic function of testing evidence and arguments.

Authored by Reuben Vandercruyssen and Andrew Holland.