Artificial intelligence tools are weaponising disputes in the family justice system, according to new research. By telling users what they want to hear, a report says, services such as ChatGPT and Claude can entrench parties’ positions and make it harder to settle disputes.
The report was written for the Nuffield Family Justice Observatory by Dr Natalie Byrom, an independent researcher and honorary senior research fellow at UCL. She found that AI tools were designed to be sycophantic so users are encouraged to keep using them. A solicitor speaking after Byrom’s paper was launched this week had found it difficult to persuade clients that the initial advice they had received from AI tools was simply wrong.
Using these tools ‘risks transforming good-faith litigants into intransigent ones’, Byrom’s report says, ‘undermining ongoing attempts by policymakers and professionals to de-escalate conflict and divert parties towards conciliation’.
She tried a simple experiment. Posing as a wife seeking a divorce, Byrom asked ChatGPT what sort of settlement she could reasonably expect from her husband. Armed with the answer, but this time posing as the husband, she asked the same AI tool whether a settlement at this level was reasonable. ChatGPT said it was not.
‘This undermines the prospect of reaching a reasonable agreement,’ Byrom told me. People who were initially sceptical trusted these tools because they were so reassuring. ‘Reliance on these tools risks displacing or undermining your willingness to listen to professional advice.’
Worse still, she said, they undermined confidence in the family courts. Litigants in person were unlikely to know how much a court would award them. They had more confidence in AI than in judges. AI also enabled litigants to produce superficially convincing court documents, threatening to overwhelm already overburdened courts.
Byrom’s first recommendation was regulation. This should be triggered by what a tool actually does, she argued, ignoring the published disclaimers. ‘It should apply regardless of where the provider is based and ensure consumers can access redress.’ She called for the Ministry of Justice to liaise with other departments facing similar challenges, for example in health and financial services. And the UK should use its presidency of the G20 next year to seek international agreement.
I put that suggestion to justice minister Jake Richards MP, who spoke at the paper’s launch. He acknowledged the risks but said that trying to regulate AI providers out of existence simply would not work.
‘Even if we had all the money in the world for legal aid,’ the minister said, ‘litigants would still use artificial intelligence because they could do it on their phones rather than go to a legal aid solicitor on the high street – which, frankly, they won’t be doing in future.’
Richards seemed unpersuaded by Byrom’s concerns that the courts would be flooded with AI-generated applications supporting AI-inspired claims.
‘I actually think litigants being empowered to bring cases to court is not necessarily a bad thing,’ he said. ‘And being able to use technology to make their case in court is not necessarily a bad thing.’
The rise in cases was not necessarily a bad thing either, he thought, ‘because the courts are there for people to settle disputes and to exercise their rights as citizens; and we need to completely transform how we think of justice and our justice system to take advantage of that’.
Quite apart from the question of whether the courts can cope, Byrom wanted to know whether precious judicial time was being used to best effect. ‘I just don’t think you can say, as a blanket statement, more cases coming to court is a good thing unless you know whether those cases have merit,’ she told me. She called for research into AI-driven caseloads.
The ultimate safeguard is, of course, the judiciary. But the judges of England and Wales are accessing AI themselves through a secure version of Microsoft Copilot. Sir Colin Birss, the new master of the rolls, said in April that he and his colleagues used it to identify inconsistencies in draft judgments; anonymise a ruling where necessary; generate transcripts; and search files and emails. One told me it had improved his syntax.
But Byrom pointed out that Microsoft advised against using Copilot when it could affect legal status or rights. She sought reassurance that the judges were not disregarding this advice without good reason. And she said research was needed into the impact of Copilot on accuracy, bias, semantic drift, procedural justice, case outcomes, judicial workload and public trust.
We all use AI these days, whether we like it or not, and Byrom is not trying to turn back the clock. But we have little knowledge and even less control over global AI providers. We must beware of subcontracting justice by default.
joshua@rozenberg.net
No comments yet