A critical eye on AI evidence

Insight

08.04.2025

A recent case, Noel Anthony Clarke v Guardian News & Media Ltd, has raised the issue of the reliability of AI-generated transcripts. In this article, we consider how useful AI transcription tools can be used in a way which makes their outputs more likely to be relied on by the courts.

AI tools

Gone are the days of painstakingly scribbling down verbatim notes of meetings or calls. There is now an abundance of automated, AI-powered transcription services on offer, ranging from the built-in features on Zoom and Microsoft Teams to paid subscription models such as Otter.ai and Rev. These services use AI to convert spoken words into text in real time. This not only saves considerable time and energy for note-takers but may also be useful in the context of litigation. When trying to establish who said what, and when, it is much easier for a party to run searches across full written records of meetings, rather than listening back to recordings.

Increasingly, the courts are placing greater evidential weight on written material which was made contemporaneously rather than on the recollection of individuals. For example, a WhatsApp message which was sent on the day an incident occurred is likely to be considered as more credible than what a witness recalls about that incident. This written material could now include contemporaneous AI-generated transcripts of conversations or meetings as the use of AI tools becomes more commonplace.

What’s the problem?

Although AI transcription tools are getting better all the time, they are unlikely to be 100% accurate. For example, in the Clarke case, an AI transcription tool had repeatedly mis-transcribed the claimant’s name, "Noel", as "no". Generally, more mistakes are likely to be made when dealing with audio involving loud background noise or strong accents. In 2024, it was reported that Whisper – an AI-powered transcription tool provided by OpenAI – had even been "hallucinating" (ie completely making up chunks of text or even entire sentences).

The occasionally dubious accuracy of AI-generated transcripts creates an evidential difficulty for the courts – to what extent can contemporaneous AI-generated transcripts be assumed to be accurate?

In the Clarke case, the defendant provided the claimant with a number of audio files and also disclosed contemporaneous transcripts of some of those files which had been produced by Otter.ai. The transcripts were of variable quality (some were unintelligible, and there was the issue of the mis-transcription of the claimant’s name mentioned above). So, the defendant decided during the course of the litigation to commission professional certified transcripts of the audio files. The claimant sought disclosure of those transcripts. This was refused by the court because the request had been made too late (in doing so, the court rejected the defendant’s suggestion that the transcripts should not be disclosed because they were privileged – they could not be so because the underlying conversations had not themselves been privileged).

What’s the answer?

Obviously, if a separate recording of a conversation or meeting has been made and retained, then the AI-generated transcript can be checked for accuracy against that. Having said this, it was made clear in the Clarke case that there is no obligation on a party to proofread or correct errors in contemporaneous transcripts before disclosing them to the other side – these “fell to be disclosed as they were”. Alternatively, a fresh – perhaps human-generated – transcript could be made from the original recording itself, as happened in the Clarke case.

However, some parties may have data retention policies or processes of automatic deletion which mean that recordings may not be retained even if they have been made. That might particularly be the case if a party faces any data storage limitations, as audio files take up more storage space than written documents. Of course, once litigation is contemplated, a party should suspend the destruction of any relevant evidence such as recordings. However, disputes often arise a long time after evidence underpinning them has been deleted under business-as-usual processes. If there is no separate recording to check a contemporaneous AI-generated transcript against, then there is a risk that it will not be given due evidential weight by the court as a result of doubts about its accuracy.

The best way to counter that risk is for parties to check and correct AI-generated transcripts in real time, and to make a record showing that this has been done. Not only is that good practice which ensures that written records are as accurate as possible, but it could also help convince a court later down the line that the transcript is accurate and should be relied upon as good evidence. These checks are particularly important where a transcript records the more important conversations that have taken place.

This publication is a general summary of the law. It should not replace legal advice tailored to your specific circumstances.