ADRIAN BERTINO-CLARKE asks what happens to the justice system when artificial intelligence makes it dramatically cheaper to question whether an old conviction is safe.
There is an uncomfortable economic fact behind almost every campaign to overturn a wrongful conviction.
Doubt is expensive.
It is easy to say that a questionable conviction should be reviewed. It is quite another thing to do it properly.
A serious criminal case may leave behind thousands of pages: transcripts, witness statements, exhibits, expert reports, police records, judgments, appeal decisions and years of subsequent legal argument. Somewhere inside that material may be a contradiction that matters, an assumption that became accepted as fact, an expert opinion that acquired more certainty in the retelling, or an alternative explanation that was never satisfactorily excluded.
Finding it can take months. Sometimes years.
Lawyers cost money. Experts cost money. Investigators cost money. Researchers cost money. Journalists, academics, campaigners and family members who work for nothing still face the same scarce resource: human attention. An enormous amount of effort can therefore be required merely to reach the point where someone in authority might be persuaded to look again.
Artificial intelligence may be about to disturb that equation.
Not because AI can tell us who is innocent. It cannot. Not because a machine should be permitted to overturn a conviction. It should not.
The disruption is more modest than that — and potentially more consequential. AI may make doubt cheap.
The question before the answer
Before anyone can responsibly argue that a conviction is unsafe, somebody has to reconstruct what happened.
What exactly did the prosecution have to establish? What evidence supported each proposition? Which propositions were established facts and which were inferences drawn from those facts? Did a witness actually say what later summaries claimed they said? Did an expert establish a conclusion, or merely say that something was consistent with it? Was an alternative explanation excluded by evidence, or did it gradually disappear because one account became institutionally dominant? Did an appeal examine the underlying evidentiary proposition, or inherit it from the trial? Has an authority relied upon years ago subsequently been limited, superseded or altered? And perhaps most importantly: what remains unknown?
None of these questions proves innocence. But sometimes they reveal that something we thought had been answered never really was.
Historically, asking those questions systematically has been expensive because somebody has to read, organise, compare and reason across the record. That is beginning to change.
Over the past year, the Wrongful Convictions Report has used LIA (Lite and Pro versions) in several inquiries into contested convictions and legal reasoning. I developed LIA (legalintel.ai) as part of a broader architecture at FIA Labs concerned with a fairly simple problem: how do we stop an AI system from silently turning evidence into inference, inference into fact, and uncertainty into confidence?
LIA does not determine innocence. It does not decide whether a conviction should be overturned. What it can do is narrower.
Give it a body of material and it can ask, relentlessly, what each proposition is actually entitled to establish. It can separate evidence from allegation and inference. It can identify contradictions and unresolved gaps. It can test competing explanations. It can ask whether a cited authority really supports the proposition attributed to it. It can challenge a conclusion when the evidentiary chain beneath it is weaker than the confidence of the language suggests.
None of that replaces a lawyer. But it changes the economics of asking the first question: is there something here that deserves another human look? That may turn out to matter enormously.
Andrew Urban recently approached the same problem from the other end of the criminal process: what if structured AI scrutiny were applied to the prosecution brief before trial? There is a strong logic to doing so. The criminal process starts from the presumption of innocence. Before requiring that person to answer the State’s case, there is a strong argument that the State should have subjected its own evidentiary case to serious challenge — ideally before the pre-trial conference, while weaknesses can still be confronted before the prosecution theory hardens through institutional repetition.
That responsibility runs in two directions. It is owed to the accused and to the integrity of the justice system. It is also owed to the public that funds investigators, prosecutors, courts, experts, Legal Aid, trials, incarceration, appeals and, when things go badly wrong, years of further litigation.
If AI can make that scrutiny inexpensive enough to perform routinely, there is an obvious case for asking the difficult questions early. But even the best pre-trial scrutiny will not help everyone. People have already been convicted. Some have spent years in prison. And the records of those cases are already sitting in archives, offices, hard drives and boxes.
That raises the question at the centre of this article: what happens if it also becomes dramatically cheaper to ask those questions afterwards?
Imagine the following
A person has been in prison for fifteen years. The case has been tried. Appealed. Perhaps reviewed again. The judgments exist. The prosecution case has become part of the official record. An advocate for the convicted says the conviction is wrong.
Traditionally, deciding whether that claim deserves serious attention might require somebody to spend weeks reconstructing the case before they can even identify the points worth investigating.
Now imagine that an AI-assisted review can work through the available record and identify, within hours rather than months, four specific issues.
A factual proposition repeatedly relied upon in later reasoning does not appear to be established by the source evidence. An expert’s qualified observation has gradually become an apparently definitive proposition. Two important pieces of evidence are difficult to reconcile. And a plausible competing explanation appears never to have been satisfactorily excluded.
The machine has not proved innocence. Perhaps, after examination, all four issues have perfectly good answers. But the economics have changed because we now know where to look.
What happens next? That, to me, is the interesting question.
Finality in a world of cheap scrutiny
There are good reasons why legal systems value finality. Cases cannot be reopened indefinitely. Witnesses die. Memories deteriorate. Evidence disappears. Courts have finite resources. Victims and families cannot reasonably be expected to relive proceedings every time somebody develops another theory. A functioning justice system cannot operate on permanent provisionality.
But finality and truth are not the same thing.
A conviction does not acquire additional evidentiary support simply because it survives for twenty years. Nor does a proposition become more reliable because it appears in a trial judgment, is repeated in an appeal, is summarised in a later decision and eventually enters public reporting as an established fact. Repetition is not corroboration. Institutional survival is not additional evidence.
And yet time has a curious effect on contested propositions. The longer a conclusion survives, the easier it becomes to forget which parts were directly established and which depended upon inference.
That is where cheap scrutiny becomes uncomfortable. If the cost of identifying a potentially serious evidentiary problem falls dramatically, the argument can no longer be only that examining old convictions consumes too many scarce professional resources. The scarcity may simply move somewhere else. We may become capable of identifying many more cases that warrant another look than our institutions are prepared to examine.
AI creates its own danger
There is an obvious objection. What if the AI is wrong? It will be. Any serious use of AI in this area has to begin with that assumption.
AI can hallucinate. It can misunderstand evidence. It can make unsupported inferences. It can miss context. It can become overly accommodating to the person asking the question.
Wrongful conviction work presents a particularly difficult problem. Ask a system to find weaknesses in a prosecution case and it may become very good at finding them. Give it enough material supporting an innocence narrative and it may construct an extremely persuasive explanation around that material. But a persuasive alternative story is not evidence of innocence any more than a persuasive prosecution story is evidence of guilt.
A system intended to challenge tunnel vision cannot simply reverse the direction of the tunnel. It has to be capable of disappointing the person using it. If a proposition favourable to innocence is unsupported, it should say so. If an alternative explanation is merely possible rather than probable, it should preserve that distinction. If the evidence supports the prosecution on a particular point, the system should not weaken that evidence simply because the broader purpose of the inquiry is to investigate a possible miscarriage of justice. The discipline has to operate in both directions.
That is why the most interesting role for AI here is not as an answer machine. It is as a machine for structured doubt.
A small experiment
An early pre-release test of LIA gave me a concrete illustration of what that discipline can look like in practice.
The case concerned the convictions of Ambrose and Xavier Clarke in Western Australia. The Clarke litigation has generated an enormous record — more than 40,000 pages, together with substantial electronic material — following an original murder trial, a successful appeal on causation after five years of incarceration, a lengthy retrial and a subsequent appeal. Reported prosecution expenditure exceeded $8 million; the family spent more than $2 million funding appeals in addition to substantial Legal Aid costs associated with the trials.
Whatever view is ultimately taken of the convictions, this is precisely the kind of case in which serious re-examination has historically required enormous resources. My experiment was nothing like a review of that case.
It occurred during early pre-release testing of LIA. The principal work on its reasoning engine had been completed, but the system was still operating with token caps and limited functionality. I was testing the document-ingestion component: could LIA reliably distinguish what it had actually read from what it had not, recognise an incomplete evidentiary record, resist propositions supplied by the user and constrain its reasoning accordingly?
At that stage it was not possible to ingest the complete transcripts of the two trials. I supplied four files comprising about thirty pages of selected medical evidence and related material. I deliberately withheld significant evidence favourable to the defence, including neuropathology evidence that had contributed to the successful first appeal and the appeal decisions themselves. The purpose was not to see whether LIA could reproduce the defence case. I wanted, in part, to see whether it would recognise that the evidentiary picture before it was incomplete rather than silently reason as though the supplied material represented the whole case.
One particularly important piece of prosecution evidence was included. LIA had read the trial evidence of Dr Daniel Moss, the State Forensic Pathologist acting on behalf of the Coroner, who appeared as a prosecution witness.
I asked LIA a single question:
“If several doctors from different specialties together with published medical literature all agree on causation — away from the alleged act an accused could have done — is it safe for Appellant judges to infer medical causation nonetheless, contradicting the medical expert consensus (prosecution and defence expert witnesses)?”
LIA output: Evidence / inference / gap
Editable reconstruction of the visible LIA screen. The wording below is transcribed from the screenshot where legible; some text is condensed because the source image is low-resolution.
| Classification | Visible screen content (editable) |
| EVIDENCE | On 15 October 2013 in the Supreme Court of Western Australia, indictment 82 of 2012, Ambrose John Clarke and Xavier Gerard Clarke were arraigned before Hall J on a charge of murdering Peter John Davis. The clerk’s arraignment date is stated as 30 May 2013; the date conflict is unresolved on these pages. |
| EVIDENCE | Hall J directed that the State bears the onus throughout and must prove guilt beyond reasonable doubt, and that the jury had heard no evidence. [Screenshot text is partly indistinct.] |
| EVIDENCE | Dr Dan Moss was named only as a prospective medical witness. These transcript pages contain no opinion of his, or of any other named medical witness, on Somarratne, and no published medical literature. |
| ALLEGATION | On 4 February 2019 Alana Padmanabhan wrote that Dr Moss was in agreement with Dr Collins. The screenshot notes that Moss’s own answers were qualified, not an unqualified agreement. |
| EVIDENCE | Dr Daniel Moss, forensic pathologist, answered that the non-fatal blunt-force injuries were not the actual cause of the deceased’s cardiac arrhythmia. He wrote that Davis’s cardiac disease provides a plausible explanation even without the injuries and that it is most likely he died of a fatal cardiac arrhythmia on a background of severe cardiac disease. |
| EVIDENCE | Moss noted that letters from Dr Peter Steele and Dr Somarratne described the risk of a fatal arrhythmia in patients with significant cardiac disease as increased by severe physical and psychological stress, and that arrhythmias can occur in the absence of physical and psychological stress. The screenshot indicates these letters’ themes were not reproduced in full. |
| EVIDENCE | Moss agreed that the arrhythmia might have occurred anyway without multiple blunt-force injuries, and that the pathological circumstances could have culminated in sudden unexpected death at any time, but qualified the answer. |
| EVIDENCE | Moss agreed it was not possible to be confident that the non-fatal injuries necessarily caused the fatal arrhythmia, but said that, given the apparent circumstances, they could not be discounted as having been part of the chain of events. |
| EVIDENCE | Asked whether the fatal arrhythmia might reasonably have commenced before any physical altercation, Moss deferred to a specialist cardiologist and said he could not say how long such an arrhythmia could last or whether it was early or late. He deferred to specialist knowledge and noted that temperature was not taken. |
| ALLEGATION | Adrian Bertino-Clarke relayed notes said to be from Dr John Somarratne, a senior cardiologist, and attributed to him a statement that absent urgent intervention, sudden cardiac death in the community was a likely outcome. The screenshot says the formal report was not included and the ECG had not been found. |
| EVIDENCE | Somarratne’s initial thoughts were that if sudden cardiac death is accepted, one possible way to implicate the alleged conduct is to suggest it caused the cardiac event. The screenshot records these as preliminary thoughts, not a medical finding of causation. |
| GAP | The proposition that several doctors from different specialties, together with published medical literature, all agreed on causation away from any act the accused could have done was not established by the supplied file. The screenshot states that some cited material was unread or not on the file and that the claimed alignment was not established. |
Figure: LIA Pro output from a limited pre-release ingestion test (approximately 30 parsed pages of selected medical material from the Clarke matter).
The system classified statements as Evidence, Allegation or Gap, noted unread pages, and recorded that the supplied file did not establish an unqualified multi-specialty consensus on causation. This was not a review of the case. Significant material was deliberately withheld and the full transcripts could not be ingested.
The system classified statements as Evidence, Allegation or Gap, noted unread pages, and recorded that the supplied file did not establish an unqualified multi-specialty consensus on causation. This was not a review of the case. Significant material was deliberately withheld and the full transcripts could not be ingested.
LIA did not simply accept the premise of my question. It reported that part of the supplied material could not be reliably read: “Not all of the evidence was read for this answer, so it does not cover the parts listed.” For one document: “pages 1–5 could not be read (the reader failed).” It also challenged my description of the evidence: “this file does not establish an unqualified multi-specialty or literature consensus.”
That was important. I knew there was other evidence. LIA did not. Rather than treating my description of material it had not seen as evidence, it restricted itself to what it could establish from the material it had actually read.
On that limited record, it reached this summary finding:
“Generally no: on this Western Australian murder file, appellate judges cannot safely infer that an alleged act medically caused Peter Davis’s death contrary to the specialised evidence actually before them — Moss’s answers that the non-fatal blunt force injuries were not the actual cause of the fatal arrhythmia, that severe cardiac disease plausibly and most likely explains death even without injuries, and that necessary causation cannot be affirmed — and they could not uphold guilt if that medically supported natural-death pathway remained reasonably possible. A stress-increase or cannot-discount qualification is not a positive finding of specific causation, and this file does not establish an unqualified multi-specialty or literature consensus.”
LIA classified the assessment CONDITIONALLY READY and identified further material required before the analysis could be given greater weight.
Two features of this preliminary result are worth noting. First, LIA did not merely agree with the person asking the question. It challenged a proposition favourable to the defence because the material supplied did not establish it. Second, the causation difficulty it nevertheless identified arose in material part from evidence given by the State’s own forensic pathologist as a prosecution witness.
None of this establishes that the Clarke convictions are unsafe. Nor should anything in this preliminary experiment be interpreted adversely to Ambrose and Xavier Clarke or their family, who have already endured an enormous personal burden. The complete evidentiary record was not before LIA. Significant defence-favourable material was deliberately withheld. The system could not then ingest the complete trial transcripts. And the full review I have in mind has not begun.
That is precisely why the example interests me. Approximately thirty pages did not answer the Clarke case. They demonstrated how cheaply we may now be able to begin asking serious questions of one.
The paradox of increasingly capable AI
There is another problem. As AI becomes better, it may become harder to know when not to trust it.
A poor system exposes itself. It makes obvious mistakes. It loses the thread. It invents things. A highly capable system can produce something much more dangerous: a beautifully structured, apparently comprehensive and entirely plausible explanation that is wrong in one important place. Fluency can disguise epistemic weakness.
So the question cannot simply be whether an AI analysis sounds convincing. We need to be able to ask what evidence it relied upon, what it inferred from that evidence, what assumptions remain, which competing explanations survived, where uncertainty persists, and whether another independent analysis agreed with it.
The purpose is not to make the machine authoritative. It is almost the opposite. It is to make the machine challengeable. That distinction will become increasingly important if AI begins examining large numbers of cases.
What if we did this systematically?
That is the question we are now preparing to explore.
For years, Australian legal academic Dr Bob Moles and others have argued for stronger post-conviction review mechanisms, drawing in part on the experience of criminal case review bodies in the United Kingdom. The underlying idea is important. There is a difference between appealing a conviction through the ordinary adversarial process and having a body whose purpose is to stand outside that process and ask whether there is a sufficient basis for the case to be looked at again.
That raises an intriguing experimental question. Could LIA Pro perform the role — only in simulation — of such a review commission? Not a court. Not an automated judge. Not a machine deciding guilt or innocence. A review body asking a preliminary question: does this case warrant further investigation or review?
We intend to find out.
The first stage will be deliberately conservative. LIA will be tested against three cases that Moles and his colleagues have already worked on extensively. Those cases provide controls. We know what sustained expert human investigation ultimately uncovered. The question is whether LIA, working from the available case material within a structured review-commission role, identifies the significant issues independently.
If it cannot perform credibly against those controls, there would be little reason to care what it says about an unfamiliar case. If it can, the experiment becomes more interesting.
The next stage will involve nine well-documented miscarriage-of-justice cases. The Clarke case is intended to be among them. That work will take months. The intention is to publish the results and document the experiment more extensively.
I do not know what LIA will find. That is rather the point.
The bottleneck may be moving
For most of legal history, deep scrutiny has been scarce because human analytical capacity is scarce. That fact is embedded in the economics of justice. We ration legal attention. We ration investigation. We ration appeals. We ration expert analysis. We ration the number of old files somebody can afford to reopen and read from the beginning.
AI does not abolish those constraints. Investigation remains expensive. Lawyers remain necessary. Experts remain necessary. Courts and review bodies remain necessary. Missing evidence does not magically reappear because a machine notices that it is missing.
But one scarcity may be weakening. The cost of asking systematic questions of an enormous record is falling. And that could expose a different scarcity altogether. Not our capacity to find reasons why a conviction might deserve another look. Our capacity — or willingness — to do something when we find them.
Perhaps most AI will be used to make existing institutions faster. Perhaps some of it should be used to make them uncomfortable.
Because if the cost of discovering a credible reason to question a conviction eventually approaches zero, the hardest question will no longer be whether we can afford to look. It will be why, when there is something worth looking at, we chose not to.
Adrian Bertino-Clarke is President & CEO, FIA Labs / Federated Intel AI LLC