AI#2 The tsunami and circumstantial trials: the Garlasco case as an example of probabilistic analysis

In recent days, therefore, I have tried to carry out an experiment: using AI not to "decide" whether a person is guilty or innocent, but to analyze a complex circumstantial process, outlining the different theses of the prosecution and the defence. The case examined is that of the murder of Chiara Poggi. 

I would like to point out that I am an engineering professor and I have no specialist skills in legal disciplines; however, I am familiar with probability theory which, in my opinion, is the heart of circumstantial trials. Below I report the summary of the analysis carried out together with artificial intelligence which produced about ten pages of reports with high degrees of in-depth analysis, in terms of probability estimation, physical dynamics of some events, evaluation of the overlap of time intervals to evaluate alibis, etc. Analyzes carried out with ChatGPT Pro 5.5 and, essentially, also confirmed with Gemini 3.1 Pro (and, in fact, easily reproducible by each of you too).

The method and estimation of probabilities with AI

The AI, fed with the main sentences of the trial, was able to build, with user interaction and very rapid times, a comparative matrix between the different evidentiary blocks on approximately a thousand pages of sentences: shoes and route, computer alibi, timetables, bicycle, DNA, biological traces, motive, alternative hypotheses.

The most interesting point is not the “response” of the AI, but the method. For each element, in fact, the AI ​​can ask itself: how compatible is this data with the prosecution's theory? How compatible is it with the defense thesis? How technically robust is it? What are the alternative explanations? What elements attenuate or strengthen the value of the clue?

This approach is particularly useful in circumstantial trials, where there is often no single and definitive piece of evidence, but a network of elements that must be evaluated together.

In the Garlasco case, the analysis highlighted that not all evidence has the same weight. Some elements appear weaker or more ambiguous: the motive, the bicycle, the soap dispenser, some biological traces that can also be explained by frequenting the house. Obviously it was a predictable result, however what is striking is the ability to schematize, to grasp the essential aspects in hundreds of pages in just a few minutes (sometimes seconds). Also elaborate the critical issues, the logical and procedural ambiguities.

The strongest clue remains that relating to route declared by Stasi And to the absence of blood on the shoes. According to the probabilistic evaluation reconstructed by the AI, the possibility that Stasi completed the declared route - entering the house, reaching the cellar area, seeing the body and returning back - without intercepting blood, appears very low. Not impossible, but very low.

The AI ​​writes: "However, I would not take the number of CTUs "at face value". A probabilistic model can be very sensitive to:

* size assigned to the spots; * minimum threshold considered as "blood"; * size and orientation of the sole; * number of hypothesized supports; * initial position of the subject; * step width; * degree of freedom left to the rotation of the foot; * assumption of casual or oriented walking; * exclusion/inclusion of microtracks.

So the numbers “13 in a billion” or “1 in five million” should be read with caution. It may be overestimated or underestimated depending on the assumptions. But the question is: even if you correct it a lot, does it change the logical order? In my opinion no. Even if that number were exaggerated by 100 or 1000 times, it would still be a very rare event.”

The prudential estimate provided by the AI, without automatically assuming all the experts' conclusions to be true, places this probability approximately between 1 in 10,000 and 1 in 500,000, depending on the hypotheses considered: natural walking, voluntary avoidance of visible patches, dried blood, dispersion of microtraces, delay in investigations.

This does not automatically mean "guilty". It means saying that that evidential block - path, shoes, absence of contamination - is the one that most undermines the version of Stasi as a simple discoverer of the body.

At the same time, the AI ​​also highlighted the defense's serious arguments: the computer and telephone alibi, the narrow time window, the failure to find the attacker's shoes, weapon and clothes, the initial investigative uncertainties, the possible dispersion of microtraces.

And this is where the true usefulness of artificial intelligence emerges. Don't construct comfortable truths. Don't confirm prejudices. Don't replace the judge. But help order reasoning.

AI can be impartial precisely because the same method can be applied to any process: a murder case, an economic trial, a judicial error, a closed investigation, an acquittal or conviction. The machine should not sideline, if it is used with method, verifiable data and critical control. Analyze consistencies, contradictions, probabilities, alternatives, relative weights of the clues. It can support the preliminary analysis, the comparative reading of the documents and the logical understanding of the evidentiary elements.

Naturally, the AI ​​does not have a legal awareness, does not directly evaluate the immediacy of the evidence and cannot replace the trial. But it can become a very powerful tool to make complex decisions more understandable, both for professionals and citizens. 

In the following table you can read the summaries provided by the AI ​​of the different evidentiary blocks.

The AI ​​did not “decide” guilt or innocence. Instead, it estimated the relative weight of the evidentiary blocks. The strongest data was that of the route declared by Stasi and the absence of blood on the shoes: in his prudential assessment, the probability that that route was completed without intercepting blood is, as anticipated, in the order of 1 in 10,000 - 1 in 500,000. It's not impossible, but it's statistically very rare.

In the documents uploaded, the theme is actually central: the 2011 appeal reports the prosecutor's argument according to which there would be "absolutely zero statistical probability" of carrying out the route without intercepting bloodstains, but the acquittal court considers it not proven with certainty that Stasi had to step on the large puddle in front of the folding door.  However, the same sentence also recalls the defensive fact of the time that had elapsed: approximately 17 hours before the delivery of the shoes, an element that prevented automatic deduction from the failure to find blood as proof that the route had not been completed.  The 2013 Supreme Court, however, highlighted the criticality of the acquittal motivation precisely on the route, recalling that a 0.6% probability of avoiding the tracks in front of the cellar stairs had been indicated, and that the issue could not be liquidated without a more rigorous overall evaluation.  

The estimate of 1 in 10,000 - 1 in 500,000 is not a "mathematical procedural truth", but a prudential AI evaluation constructed taking into account the sentences, the expert reports referred to and the possible defense objections: dried blood, dispersion, delay in the seizure of shoes, voluntary avoidance of visible stains.

AI and public opinion: where the gap arises

A further development could be the analysis of the relationship between sentences and public opinion: comparing what emerges from the procedural documents with the dominant "mood" in the media, in social networks, in public comments. Here too, AI could help to distinguish between collective perception and procedural data, between media narration and evidentiary structure. This would perhaps be one of the more civil uses of artificial intelligence: not produce automatic sentences, but increase the quality of public debate. Because an informed citizen doesn't need slogans. He needs tools to understand.

I also tried to do this test on the Garlasco case which shows well how the collective "mood" can change over time. For years, public opinion has been divided between those who are guilty and those who are innocent. Today, with the new investigations and the attention on Andrea Sempio, the media climate appears to have once again shifted towards doubt about the consolidated procedural truth.

The AI ​​writes: "When evaluating public opinion, I used a prudent formula: today the case has become very divisive again because the new investigations into Andrea Sempio have reopened the debate and the international press expressly speaks of "fresh doubts" and a possible turning point in the case. There is also at least one recent journalistic survey according to which a large share of Italians consider the new Sempio lead credible. This signals not a procedural truth, but a change in the public climate."

This is therefore the second level of analysis carried out: not only what the procedural documents say but also how public opinion perceives that trial today. In the Stasi/Garlasco case, we saw that the AI ​​identifies the blockade as the strongest point against Stasi shoes/path/blood: the probability of a “clean path” appears very low.

However, public opinion today seems to contest above all the overall picture, not just the individual technical data. The public question seems to be: “Even if that lead is strong, is it still sufficient if new doubts, new leads or possible investigative errors emerge?”

The new investigations into Andrea Sempio have reopened the media debate and put issues such as DNA, alternative leads, investigative gaps and reliability of procedural truth back at the center. The international press has talked about reopening the case in light of new forensic elements.

AI schematizes the issue like this:

The interesting fact is that public opinion does not always think like a judge. The judge evaluates evidence, clues, procedural rules and reasonable doubt. The citizen, however, also perceives the context: investigative errors, delays, omissions, contradictions, television narratives, new leads, suspicions that remain open. For this reason, it may happen that a technically strong (and probabilistically very relevant) clue is no longer enough to reassure public opinion, if a framework of doubt reopens around that clue.

And it is precisely here that AI can be useful: not to replace the judge nor to chase the "sentiment" of social media, but to distinguish three different things: trial data, forensic probabilities and public perception.

Because a mature democracy should not choose between incomprehensible sentences and trials celebrated on television. It should have better tools to understand.

Below is the comparison table between AI and public opinion:

Naturally this reflection has no claim to replace the work of magistrates, lawyers, experts or criminal trial scholars. It is just an attempt to show how artificial intelligence can help to better read complex documents, to distinguish between data, interpretations and collective perceptions, and to make a debate often dominated by slogans or emotional contrasts more accessible. Every observation remains open to comparison, correction and technical analysis. 

The objective is not to "be right", but to contribute, with prudence and respect, to a more informed public discussion also thanks to the help of AI.

Naples, 05/13/2026