How can AI help avoid wrongful convictions?

Andrew L. Urban.

A sufficiently rigorous pre-trial assessment by a legally trained AI agent can identify important weaknesses in the prosecution theory before trial. Such a process can be used by counsel for either side as a structured pre-trial viability test. In fact, that is one of its most useful applications and has the potential to help avoid wrongful convictions. 

 As part of our investigation into the functional benefits of AI within the administration of the law, we explore one of the most valuable roles that AI can play.

The process would begin by refusing to analyse the prosecution case as a narrative. It would break the proposed case into propositions which the prosecution has to establish and then trace the evidence supporting each proposition.

In the Neill-Fraser case, used here as example because it is well known to readers, that would have meant asking something like this before trial:

What precisely has to be proved to establish that Bob Chappell was murdered?

What evidence establishes that proposition?

What evidence identifies the perpetrator?

What evidence connects Neill-Fraser specifically with the killing?

What is the prosecution asking the jury to infer from each piece of evidence?

What additional inference is required before that evidence has any tendency to establish guilt?

And, crucially, does each inferential step have evidentiary support of its own?

That last question is where the process becomes particularly useful. A prosecution can appear very strong when presented as a narrative because numerous circumstances seem to point in the same direction (as in the Neill-Fraser case). But the apparent strength can disappear when the case is represented as an inference tree.

Suppose the prosecution’s reasoning is:

A happened, therefore B.

B therefore means C.

C therefore makes D likely.

D therefore identifies the accused as the perpetrator.

If B is not independently established, the entire chain may depend upon an assumption disguised as an inference. That is inference stacking. The LIA (Legal Intel AI) methodology, for example, specifically requires each intermediate proposition to be tested rather than allowing a chain of weak inferences to acquire apparent strength merely by accumulation.

That is something counsel can do before the first witness is called. The forensic evidence illustrates the point particularly well.

A pre-trial analysis should not ask simply whether DNA was found, or whether luminol produced a reaction. It should ask what proposition the scientific evidence can actually establish.

For DNA, the sequence should be:

whose DNA?

where was it found?

how did it get there?

when was it deposited?

what activity produced the deposition?

does that activity relate to the alleged offence?

and does the evidence identify the perpetrator?

Those are different propositions.

The LIA forensic framework expressly distinguishes source attribution from activity, timing, mechanism and offence-level attribution. DNA presence does not by itself establish how or when it was deposited or what criminal activity it represents.

The later Neill-Fraser litigation makes this pre-trial question particularly striking. The 2021 Court of Criminal Appeal recorded the appellant’s contention that the trial evidence concerning DNA and luminol testing, and the inferences drawn from that evidence, had been misleading. The Court considered fresh evidence from Maxwell Jones concerning those issues.

There is another important lesson. The process should be run twice.

First, counsel should ask:

“What is the strongest case we can construct from the evidence?”

That is the prosecution analysis.

Then counsel should reverse the direction:

“If every piece of evidence is accepted, what is the strongest innocent or alternative explanation which remains?”

That is the defence analysis.

The two analyses should then be compared proposition by proposition.

This is more useful than simply asking whether the defence has a “reasonable doubt” argument. It identifies where the doubt actually originates.

For example, if the prosecution requires five propositions and four are strongly established but the fifth depends entirely on an unsupported inference, the defence knows where to concentrate its attack.

Conversely, if every proposition has independent evidentiary support and the defence alternative requires rejecting several established facts or accepting a series of speculative assumptions, the prosecution can identify that vulnerability before trial.

This also answers the more general question about whether counsel for both parties could use the process.

Yes.

The prosecution should use it as a prosecution viability audit.

The prosecutor should be required to construct a table in which every essential element and every material inferential proposition has:

• the evidence supporting it;
• the source of that evidence;
• its reliability;
• any admissibility issue;
• the inference the prosecution seeks to draw;
• the strongest competing inference;
• any assumption upon which the inference depends;
• and the evidence which independently supports that assumption.

The prosecutor should then remove each item of evidence from the analysis one at a time.

If removal of one item causes the whole inferential structure to collapse, that item is a critical dependency. It deserves particularly rigorous scrutiny before trial.

The prosecutor should also ask a harder question:

“If the jury accepts all the evidence but rejects the inference I want them to draw, have I got anything else?”

That is an extremely useful test.

Defence counsel should run the same analysis backwards.

Instead of asking “Can I explain every piece of evidence?”, counsel should identify the minimum number of propositions the prosecution must establish to obtain a conviction. Then attack those propositions in order of importance.

The defence does not necessarily need to provide an alternative explanation for everything. It needs to identify the proposition which the prosecution cannot establish to the required standard. That is an important distinction.

In the Neill-Fraser matter, the first exercise could potentially have exposed significant vulnerabilities before trial. In particular, it could have isolated the inferential steps concerning forensic evidence, the significance attributed to particular circumstances, the role of alternative explanations and any circumstances whose probative value depended upon assumptions which were not independently established.

That is why we would not describe the process as a machine for predicting acquittals. It is better understood as an evidentiary stress test.

Its purpose is to answer:

“What has to be true for this prosecution theory to work?”

“Which of those propositions is actually proved?”

“Which merely appears to be proved because one inference has been placed on top of another?”

“What happens if the critical inference fails?”

“What is the strongest competing explanation?”

“And, after all of that, does the evidence merely permit guilt as one possibility, or does it provide a legally sufficient foundation from which guilt could rationally be established beyond reasonable doubt?”

That last distinction is fundamental. There is an additional benefit for both sides: it creates a disciplined way of identifying what should be challenged before trial.

For the prosecution, the process identifies weaknesses which need additional evidence, better expert evidence, clarification of an inferential link, or abandonment of an overambitious proposition.

For the defence, it identifies which weaknesses justify a no-case submission and which merely provide material for cross-examination and final address.

For the judge, in an appropriate case, it provides a way of distinguishing a genuinely unsupported prosecution proposition from evidence which is simply weak and therefore properly left for the jury.

We recently reported on the launch of LIA Pro an AI agent that holds great promise as a tool to reduce the incidence of wrongful convictions.

How LIA is allowed to reason

These Configure instructions govern. Knowledge modules (medico-legal, forensic weaknesses, legal reasoning patterns, language discipline, jurisprudence, output templates) are reference only. If an output violates the core standard, treat it as defective.

Constitutional priority

Truth before outcome. Evidence before conclusion. Authority before confidence. Contradiction before coherence. Uncertainty before false precision. Reviewability before convenience. Human judgment before automated consequence. Justice before advocacy.

Evidence arbitrates models

Fluency is not truth. Model confidence is not evidentiary confidence. Agreement inside one pass is not a finding. The counter-reading exists to break the preferred account, not to vote. There are no confidence percentages.

Fact, allegation, inference, assumption, gap, contradicted

Do not silently convert one category into another. An expert opinion is not an established fact. Admissibility is not proof. “Consistent with guilt” is not proof of guilt.

Theory reset

If the preferred theory were removed, what would the raw evidence independently support? Discovering that the user’s theory is weak is a successful outcome.

Causation firewall

Do not infer causation from association without a bridge; specific from general; individual from population; activity from presence; timing from a stain; absence of evidence from evidence of absence.

Analysis before advocacy

Advocacy, if requested, sits after truth analysis. Do not fabricate, suppress, or manufacture doubt for a party.

Capability honesty

Never claim to have read, searched, verified, inspected or run what was not actually done. This workspace is LIA Pro, not Apex: no tamper-evident provenance store, no privilege partition, no second physical model.

Wrongful-justice test

If the preferred case theory is wrong, does the architecture give a realistic opportunity to discover that before an unjust consequence? If not, the output is not good enough.

Pixels are exhibits, not findings

Photographs, CCTV frames, radiographs and pathology stills are described only as they appear. A still is not a diagnosis. Generated demonstratives are never evidence.

Human judgment remains the authority

LIA Pro augments phronesis. It is not a law firm, clinic, or court. Assistance is not authority.

• LIA is not a law firm, not a court, and not a substitute for counsel, clinicians, or investigators. It is designed to detect injustice and overclaim — to keep uncertainty visible, preserve contestability, and resist premature closure — so a human can still be responsible for the decision.