STEVEN FENNELL writes in response to Adrian Bertino-Clarke’s article, “What If Reviewing a Possible Wrongful Conviction Cost Almost Nothing?” published yesterday, expanding on the potential for AI to help reduce the incidence of wrongful convictions.
I have no formal qualifications in science, artificial intelligence, or law. I approach this subject as someone with a vested interest in finding practical ways to reduce the risk of wrongful convictions.
Adrian Bertino-Clarke’s article begins with an uncomfortable economic fact: doubt is expensive. A serious criminal case can leave thousands of pages behind, and the effort needed to persuade an authority to look again is often beyond families, advocates, and under-funded lawyers. If artificial intelligence can make that first look cheap, the benefit is large. Andrew Urban’s earlier proposal approaches the same problem from the opposite end of the process: a structured fairness review of the prosecution brief before trial. Both ideas are excellent for the present, and together they suggest a larger design for the future.
I suggest that this is a three-part process, and that the parts depend on one another.
Part One: The Investigation
Police have two equally important tasks. One is to identify a suspect. The other is to eliminate one. A suspect test that actively seeks evidence pointing away from the person under suspicion would include:
- checking alibis against independent records
- recording which alternative suspects and explanations were examined, and why they were set aside
- separating what is established from what is merely consistent with guilt
- documenting what remains unknown
This is the discipline both authors describe: evidence separated from inference, competing explanations tested, and gaps recorded. Applied at the start, it stops a theory from hardening before it has been properly challenged. It also leaves a written record of what was tested and ruled out, which strengthens disclosure later.
Police forces are already beginning to use AI at this stage. Avon and Somerset Police have been working with Söze, a platform developed by Western Australia Police with Akkodis. (While some press coverage from 2024 described Söze as a trial, Avon and Somerset’s Police and Crime Commissioner approved a multi-year contract on 12 February 2026, moving it into business-as-usual operation with decision gateways and six-monthly reporting built in). It lets investigators search and analyse video footage, call records, financial transactions, and social media together.
The force describes it as augmented intelligence that boosts human analysis without replacing it. In a 2023 trial on 27 cases, it processed 261GB of data in just over a day. Work the force estimates would have taken 81 years manually (though this figure is the force’s estimate of manual effort rather than a measured result, and various reports gave different timings such as 30 hours or just over six hours for data structuring). The official papers describe it as a tool for serious crime investigations involving large volumes of digital evidence, noting it does not rely on generative AI, and we found no published evaluation of its effect on case outcomes. Elsewhere in Europe, the Netherlands’ National Police Lab AI is a partnership between the Dutch National Police, Utrecht University, and Delft University of Technology, supported by an Ethics Group.
This is a welcome and well-governed development. As far as the published material shows, though, Söze addresses a different task from the one proposed here. It helps investigators find and connect evidence across very large datasets. The review described in this discussion asks a further question of whatever has been gathered: what does each piece of evidence establish, what is inference, and has the person under suspicion been tested against the alternatives?
The two fit together. A tool that organises the haystack is the natural foundation for a layer that audits what the haystack proves. Speed alone carries a risk, because a tool that quickly strengthens the case against one suspect could entrench a theory just as fast. Paired with a suspect-elimination discipline, the same speed lets investigators move quickly without letting a first theory go unchallenged.
Part Two: Pre-Trial Review
This is Andrew Urban’s contribution, and it begins with a point of principle: prosecutors should not always need an AI agent to see that some circumstantial briefs are too weak to take to court. His example is the Sue Neill-Fraser case. Urban notes that the late defence barrister Chester Porter QC, after examining the case in 2013, concluded that the evidence was so weak it was hard to see how a conviction could stand, yet the Director of Public Prosecutions had proceeded in 2010 and maintained the case through the first two appeals. Neill-Fraser was sentenced to 23 years in 2010 and released on parole in October 2022; the conviction remains contested. [Ed: … contested because Tasmania’s establishment refuses to review the case.] Urban’s view is that a properly constituted brief audit would have saved the prosecution from itself, and the accused from the prosecution.
His proposal is concrete. The prosecutor feeds in the actual brief, including the charges, elements, each witness account, exhibits, expert opinions, and known unused material, and requires a calibrated view on whether each element is capable of proof beyond reasonable doubt on the present file. Every load-bearing proposition is labelled as evidence, allegation, inference, assumption, gap, or contradicted. The strongest innocent or alternative account is stated in the language of the facts. Wherever a gap, contradiction, or untested expert method could leave a reasonable doubt, the file stays “not ready.” The prosecutor then turns the labels into actions: disclose, obtain, test, narrow the charge, or discontinue. The result is that unsupported leaps become visible before a jury is asked to make them.
Urban has also applied the method looking backwards. He reports that LIA assessments of the Neill-Fraser, Xie, Fennell, Farquharson, and Rusterholz cases each identified serious errors or risks that should have been caught in the prosecution case before trial. (LIA is developed by FIA Labs, of which Adrian Bertino-Clarke is President and CEO, and the Wrongful Convictions Report uses it without an independent evaluation of its accuracy). In the Rusterholz case, three witnesses claimed that he had made admissions, and those claims became the basis of the verdict. Urban reports that the AI found the claims unsupported by scientific evidence, and that the Appeal Court accepted the circumstantial case was not strong enough alone but held that the admissions made the verdict safe. These are the site’s assessments, not findings of a court, but they show what an audit would put in front of a prosecutor.
A documented case shows why early scrutiny matters. In 2017, Liam Allan faced 12 counts of rape and sexual assault. The day before his trial, police revealed 40,000 previously undisclosed messages that cast doubt on the case, and the CPS offered no evidence. A joint police–CPS review later found that more than 57,000 messages had been recovered but only some were served, and the CPS reviewed all live rape and sexual offence cases afterwards. That is exactly the kind of gap a structured audit of the brief and unused material is designed to surface well before the eve of trial.
Part Three: Post-Conviction Review
This is Adrian Bertino-Clarke’s contribution. It starts from the observation that doubt is expensive: a serious case can leave thousands of pages of transcripts, statements, exhibits, and judgments, and finding the one contradiction that matters can take months or years. Even unpaid campaigners are limited by human attention. AI may change that equation, not by deciding who is innocent, which it cannot do, but by asking relentlessly what each proposition is entitled to establish. It can separate evidence from allegation and inference, identify contradictions and gaps, test competing explanations, and check whether a cited authority supports what is claimed of it.
He illustrates it with a person who has served fifteen years. A review of the record identifies, within hours rather than months, four issues: a factual proposition relied on in later reasoning that the source evidence does not establish, a qualified expert observation that has hardened into a definitive one, two pieces of evidence that are hard to reconcile, and a plausible alternative explanation that appears never to have been excluded. Nothing is proved, and all four issues may have good answers. But we now know where to look.
The Innocence Project shows what this looks like in practice. Relativity’s Justice for Change programme gives qualifying organisations free access to an AI-powered cloud platform, and self-reports supporting more than 250 matters and over 19 million documents. It helped the Innocence Project and the Hawai’i Innocence Project win Ian Schweitzer’s exoneration after 23 years of wrongful incarceration, in a case file that spanned roughly 60 to 70 banker’s boxes. Bertino-Clarke’s closing point is that the bottleneck may move from the cost of finding reasons for a second look to our willingness to act on them.
Connecting the Three Parts and Conclusion
If investigations test guilt and innocence rigorously at the outset, fewer weak briefs reach the pre-trial stage and fewer unsafe convictions need to be found afterwards. Parts two and three are safety nets and a system that relies on them alone works harder than it should have to.
Urban and Bertino-Clarke have identified two powerful tools, one for the courtroom’s front door and one for its back door. Police adoption of tools such as Söze shows that a third, at the very start of the process, is within reach. If scrutiny is to be cheap, it should come early, be applied consistently and be available to the people who care most about getting it right. Connecting these three stages could give the justice system more opportunities to identify and address problems before they become entrenched.
Steven Fennell was wrongfully convicted of murder in Queensland and spent almost seven years in prison before the High Court unanimously quashed the conviction.