The (electric) lamp… Is it time for juries to be replaced by AI?

Juries decide a fraction of all criminal cases. Despite their rarity, almost any proposal to restrict the right to trial by jury has been met with impassioned public and professional opposition, and the jury is often seen as one of the most important facets of our criminal justice system. Sir William Blackstone described it as ‘the palladium’ or ‘grand bulwark’ of our liberties, whilst Sir Patrick Devlin famously spoke of it as a ‘little parliament’.

In the near thousand years since its introduction, and despite its veneration as an institution, the jury system is vulnerable to attack, with some suggesting that we should do away with the system entirely and replace it with judge-only trials in the Crown Court. Despite this, and in light of the fervent support from legal practitioners and most of the general public, the jury system has remained functionally similar to that which was likely introduced with the Norman invasion of 1066. That said, technology has come on somewhat since the Battle of Hastings and, as the criminal justice system creaks and organisations rapidly rely upon generative artificial intelligence (AI) tools, is there an argument to suggest that technology has improved to such an extent that criminal defendants could be tried by machines?

What are the problems with juries and how could AI help?

The backlog

With over 65,000 cases waiting to be heard, Crown Courts in England and Wales are, and have been for some time, struggling to deal with an enormous backlog. Defendants and victims of crime are having to wait months or even years for their cases to reach court. The average hearing length of trial cases disposed of in the Crown Court by a not guilty plea in Q3 2023 was 19.4 hours. The length of trials can be drawn out further by the fact that sometimes jurors are late to court or fail to turn up. Occasionally, jurors need to take days off from sitting for planned appointments that cannot be changed. Usually, they can only hear evidence and deliberate during court sitting hours.

Philosopher Nick Bostrum wrote that ‘biological neurons operate at a peak speed of about 200 Hz, a full seven orders of magnitude slower than a modern microprocessor (∼2 GHz)’. An AI jury, then, could reach a decision around seven times faster than a human one, and drastically reduce the average hearing length as set out above. An AI jury, which would never be late, nor need to take days off, and could process information for 24 hours a day, with the need for little more than a reliable power supply and internet connection, could be the solution to the Crown Courts’ backlog.

Consistency of verdicts and hung juries

Channel 4’s recent TV show, The Jury: Murder Trial, took place in a reconstructed Crown Court and was billed as a ‘landmark experiment’ in which a ‘real-life murder trial’ was ‘restaged in front of two juries of ordinary people’. The question Channel 4 asked was ‘will they both reach the same verdict?’ The answer was that they did not. One jury returned a verdict of murder, the other manslaughter, despite hearing exactly the same evidence.

Obviously, Channel 4’s ‘experiment’ was created for the purposes of entertainment, which should be borne in mind when weighing up its probative value. That said, the programme’s conclusion does raise an issue with our method of trying those accused of crime: the fact that a jury can reach one verdict and another jury (considering exactly the same evidence) can reach an entirely different one. We see this problem in practice when a jury fails (following a majority direction) to agree on a verdict. Some members of the jury feel able to convict, and others do not, based on entirely the same evidence. Hung juries represent a societal and financial cost to the court system and the affected parties. A complainant may feel that justice has been denied. A defendant may have languished on remand for a considerable time awaiting their trial. They must now wait for the Crown to decide whether to seek a retrial and, if it does, the emotional turmoil of court proceedings begins afresh.

Jurors may reach inconsistent verdicts because humans make decisions by filtering objective facts and evidence through their personal values and beliefs, regardless of judicial direction. AI, as a non-sentient sequence of code – albeit a complex one – has no (known) values or beliefs. It relies upon a set of rules to be followed when reaching a conclusion as to a set of facts (i.e., its algorithm). It follows that were the same trial to be played out in front of an AI jury over and over again, the AI would always follow the same set of rules and, thereby, reach entirely the same verdict. In this way, AI juries (using the same algorithm) could contribute to more objective and consistent outcomes.

Unreasoned verdicts and witness evaluation

English juries do not provide reasons for their verdicts. This longstanding principle of our legal system means that those wishing to appeal against a conviction (on the basis that a jury’s decision may have been incorrect) are limited to focusing their grounds of appeal upon the fairness and adequacy of the trial judge’s directions to the jury, rather than arguing that their conviction is unsafe due to any deficiency in the quality of jurors’ reasoning. Indeed, a jury’s reasoning (unlike in the US), is never revealed. Cheryl Thomas, in her 2010 study, ‘Are Juries Fair?’, found that 51% of jurors in a case simulation felt that the judicial directions on the law were ‘difficult to understand’ and, when jurors’ actual comprehension was examined, only 31% of them ‘actually understood the directions fully in the legal terms used by the judge’.

The practical effect of this point was illustrated in R v Jogee [2016] UKSC 8, in which the Supreme Court ruled that the courts had been wrongly applying the law of ‘joint enterprise’ for three decades. The Court of Appeal subsequently ruled that it would only quash convictions on the basis of joint enterprise where to do otherwise would amount to a ‘substantial injustice’ because ‘the change in the law would, in fact, have made a difference’. In doing so, the Court would examine ‘the matters before the jury and the jury’s verdict (including the findings of fact which would have been essential to reach such a verdict)’. In most of the relevant cases, judges’ directions outlined several ‘routes’ by which a jury could convict a defendant and only one of these was found to be wrong in Jogee. Given that the juries in these cases had not given the reasons behind their decisions to convict, it is impossible to ascertain whether a particular jury reached its decision based on the old (and incorrectly applied) joint enterprise law. This precludes any appeal (or, at least, makes such an appeal very difficult) on the basis that the correct legal direction ‘would, in fact, have made a difference’.

What a witness says in oral evidence is often a decisive factor in the outcome of a trial. What is also crucial, albeit less obviously, is how a witness says what they say. Jurors evaluate non-verbal cues, such as body language, in reaching a conclusion as to the reliability of a witness’s evidence. Humans are remarkably poor at spotting a liar. In one study, people were only able to detect accurately whether someone was lying 54% of the time (i.e., just over the level of chance).

Perhaps AI could be used to detect specific behavioural variables such as near imperceptible facial micro-expressions, eye movements, and the time taken to respond to a question associated with lying, or analyse the stress patterns in someone’s voice when undergoing cross-examination. In this way, AI could establish the reliability (or lack thereof) of a witness giving live evidence with a far higher success rate than that offered by human jurors.

An AI jury could, at the conclusion of proceedings, be programmed to prepare a comprehensive ‘route to verdict’. This could set out the software’s reasoned decisions on all documentary and live evidence presented in a case, thus making an appeal in circumstances such as those similar to the Jogee ‘wrong turn’ much easier.

Humans are flawed, but so are computers

Jury tampering

Human juries are vulnerable to tampering, as most famously exemplified by the use of ‘Diplock Courts’ in Northern Ireland at the height of the Troubles. Northern Ireland still has a provision for non-jury trials in exceptional cases (see Section 1 of the Justice and Security (Northern Ireland) Act 2007). Similarly, Part 7 of the Criminal Justice Act 2003 (CJA 2003) provides for judge-only trials in cases where there is a danger of jury tampering or where jury tampering has taken place.

AI, despite being impervious to physical threats from those who might wish to sway the outcome of a trial, will always be vulnerable to hackers whatever the security measures built into the software. It may even be simpler to sway the decision of an AI jury than a human one. Tampering with a human jury would necessarily require the application of force or threats. Tampering with an AI jury, however, would require only the services of a hacker with a laptop whose cyberattack may even go unnoticed, if sophisticated enough. Therefore, replacing human juries with AI would not eliminate the risk of jury tampering and may even make it more prevalent.

Bias

Sajid Qureshi was convicted in 2000 of arson and sentenced to four years’ imprisonment. Subsequent to his trial, a woman claiming to have served on the jury wrote to the court to state that racist remarks were made throughout his trial. She claimed that some jurors seemed to have reached a decision as to his guilt before the conclusion of proceedings.

A jury, as a randomly selected representative cross-section of society, inevitably suffers from the biases and prejudices that afflict the wider society that it represents. Humans make decisions by filtering facts through their interpretation of events and people’s beliefs colour this interpretation. Some of these beliefs may be, whether consciously or unconsciously, prejudicial, which leads to biased decision-making. These prejudices can, as they did in Mr Qureshi’s case, result in jurors placing less weight on the evidence presented to them and reaching a verdict based, at least in part, upon one or more biases not related to the evidence. This, in turn, can lead to miscarriages of justice.

Perhaps AI could be taught to filter out irrelevancies during the decision-making process. It might be possible to programme the algorithm driving it to focus solely on the evidence and reach a purely objective verdict. However, given its nature, AI suffers from algorithmic bias, which refers to systematic errors in the system upon which AI relies that could create prejudicial outcomes. Generative AI must learn its task. The data set used to ‘train’ an AI system may not represent the entire population. As a result, the algorithm’s decisions may display negative bias towards demographics that did not feature in its training. Similarly, the architects of a particular algorithm will necessarily be human and will suffer from biases as a result. An AI designer might unknowingly assign their prejudices to the algorithmic software they create, which then automates and perpetuates them. An AI jury, then, is only as objective as its creator or its training. If either is biased, any verdict of the AI could be wrongfully reached.

Part of the rationale behind a jury consisting of 12 people is that a larger number of jurors should give the broadest range of views. If one juror suffers from a particular bias, the input of those on the panel that don’t should prevent the jury as a whole reaching a decision that is tainted by that particular bias. In the case of an AI jury, if the algorithm is infected by bias, there are no other jurors on the panel to cancel it out. The prejudice cannot be rectified before the verdict.

Lack of conscience

A plaque next to Court 1 at the Old Bailey commemorates Bushell’s Case (of 1670) which, as the plaque says ‘established the right of juries to give their verdict according to their convictions’.

In April, the High Court ruled that the Solicitor General had no ‘reasonable basis in fact and law’ to seek proceedings for criminal contempt of court against Trudi Warner, a climate activist (HM Solicitor General v Warner [2024] EWHC 918 (KB)). Ms Warner had attended the trial of several persons affiliated with Insulate Britain in respect of acts arising out of a protest. She raised a placard outside an entrance (used by jurors) to Inner London Crown Court which said ‘Jurors, you have an absolute right to acquit a defendant according to your conscience’.

Mr Justice Saini ruled, in denying the Solicitor General permission to make a contempt application, that Ms Warner’s placard ‘reflect[ed] essentially what is regularly read on the Old Bailey plaque by jurors, and what our highest courts recognise as part of our constitutional landscape’.

‘Jury nullification’ is what prompted Sir Devlin’s famous appraisal of the jury as the ‘lamp that shows that freedom lives’. It is a critical safety valve, representing the right of a jury to correct perceived unfairness and tackle any perceived unjust application of criminal law, which conflicts with societal values and moral conscience.

Perverse verdicts also highlight the key feature of the jury and one that AI lacks. Human jurors apply the moral conscience of society in coming to their verdict. A jury might return a perverse verdict in the face of police or prosecutorial corruption, or in the light of what it sees as state tyranny (as was the case in the acquittal of Clive Ponting) or, as what may have been the case in the ‘Colston Four’ case, in protest at an unjust law. AI lacks a conscience. It possesses no code of ethics nor moral compass. AI cannot apply the moral sentiment of the community in reaching a verdict. In this way, replacing human jurors with AI would result in the loss of a crucial safeguard against oppression.

A proposal

During a speech in the House of Commons in November 1947, Winston Churchill said that ‘democracy is the worst form of Government, except all those other forms that have been tried from time to time’. The jury system, by analogy to Churchill’s views on democracy, is undeniably flawed, but its proponents argue that it is the best method we have of determining a defendant’s guilt, or lack thereof. Replacing human juries with AI would solve some of the problems posed by the former, but might simultaneously exacerbate others, and could well create new obstacles to the administration of justice. In this sense, a direct substitution could be a flawed elixir.

My proposal is that we give defendants the choice. Perhaps the question of how one wants their guilt or innocence decided should be for the accused, as it already is in respect of either-way offences and the choice between a Magistrates’ or Crown Court trial. Do they wish to be tried in the Crown Court by their peers or by an algorithm? Lawyers already advise as to the selection of summary trial or trial on indictment. Why couldn’t we counsel clients on the respective advantages and pitfalls of trial by humans and by an algorithm?

It is becoming increasingly difficult in the age of information to prevent jurors from seeking details about their case extraneous to that presented in the courtroom. A curious juror (in defiance of judicial direction and despite the risk of criminal prosecution) could easily search the internet for what Fleet Street, or their friends on social media, have to say about the high-profile defendant in their case. An algorithm (unlike the inquisitive juror) could be programmed to reach a decision based only upon the evidence presented at trial. We could create a sequence of code to restrict its ability to search for wider information about a case, rather than relying on trust (and the threat of prosecution) as we do in the case of human jurors. Those in the public eye facing prosecution might seek to avoid the prospect of prejudice caused by such juror misconduct and have their fate decided by a machine.

Section 43 of the CJA 2003, which was repealed in 2012, would have allowed judge-alone trials in certain serious and complex fraud cases. The provision was never brought into force and the last attempt to do so proved so controversial that it was blocked by the House of Lords. Despite this, in civil litigation, judges regularly decide whether a defendant acted fraudulently. We do not assert that we should transplant juries into the civil courts to try a defendant’s character or assess the evidence against them, or that civil judges are incapable of doing so. Rather, our Commercial Court is viewed as the globe’s elite centre for international commercial litigation. The long-running argument as to whether juries are best placed to decide on serious and complex fraud cases is perhaps ill-suited to this general and serious crime edition of The Knowledge. That said, those in favour of juryless trials tend to argue that lengthy fraud trials cases represent a significant intrusion upon jurors’ lives and pose a significant risk of delays should a juror fall ill, and that judges sitting alone would more readily understand the financial and commercial context of the evidence presented. A corporate defendant facing allegations purportedly supported by volumes of abstruse evidence and seeking to advance a complex defence involving highly technical expert testimony might well elect trial by AI.

As to whether the human juror might be replaced, in a recent speech concerning the likely impact of AI upon the legal profession, Sir Geoffrey Vos, Master of the Rolls, refused to be drawn on the question of whether AI is likely to be used for any kind of judicial decision-making. His pithy conclusion: ‘We shall see’.