88,000 complaints re-examined in 6 weeks after a tragedy. 15,000 ever recorded. 7 minutes per file. The AI could have read them continuously. Nobody deployed it.
88,000.
This is the number of complaints for sexual violence against minors that the French justice system has just re-examined in six weeks, after the tragedy in Gers. An 11 year old girl. A suspect targeted by several complaints. Classified. Untreated. Never crossed.
The Minister of Justice ordered everything to be taken back. “No one will go on vacation.” Result presented this week: among these 88,000 files, more than 15,000 complaints were discovered by investigators — they had never been recorded by the public prosecutor’s office. 7,452 concern crimes where the alleged perpetrator is identified. 970 priority cases involve victims who are still minors facing perpetrators who have already been convicted.
These files existed. This evidence existed. These children were waiting.
Nobody had read them.
60 files per day, 2 magistrates, 6 weeks
Here is how this operation actually took place.
In Nantes: two magistrates and two civil servants examined 60 files per day for weeks. Do the math: 7 hours of work, 60 files, that’s 7 minutes per file. Seven minutes to decide if a child is in danger.
An investigator specializing in violence against minors testifies: “We are not able to handle all the cases, there are too many.”
And while all staff are mobilized for this review, the general secretary of a police union warns: “There will be delays elsewhere, it’s mechanical.” Including — listen carefully — on other cases of violence against minors.
To empty a barrel, we fill another.
It is a zero-sum system. The available humans are finite in number. Their attention is over. Their energy is finite. Moving staff from one file stock to another does not address the problem. He moves it.
Unless we change tools.
What an AI would have done with these 88,000 files
Let’s talk concretely. No science fiction. Documented capabilities, available today.
A artificial intelligence reads a 200-page file in seconds. She reads 88,000 in a few hours. It extracts facts, dates, names, places. She crosses. It detects that the same individual appears in three complaints filed in three different jurisdictions, under three different procedure numbers, processed by three services which do not communicate.
This crossing – the one that was missing in the Gers – an AI does it continuously. Not in six weeks of general mobilization after a tragedy. Permanently. Each new complaint automatically compared to the existing one. Each potential recurrence reported. Each dormant file brought to the surface.
She doesn’t get tired by the fortieth file of the day. She doesn’t skim pages 87 to 134 because it’s a long day. She does not forget the complaint of 2022 when that of 2025 arrives.
And above all: she does not sort out of apparent urgency. She processes everything. At the same level of attention. From the first to the 88,000th.
The first file in the morning has a 65% chance. The last one, almost none.
This number is real. Documented. Published in the Proceedings of the National Academy of Sciences.
A study of more than 1,000 judicial decisions measured the impact of fatigue on verdicts. At the start of the hearing: 65% favorable decisions. At the end of the day, after around ten cases: almost zero. The exhausted brain no longer decides. He chooses by default. And the default choice is the status quo. The ranking. The pile.
Now apply that to a civil servant who has to review 60 files a day.
Do file number 3 and file number 58 really have the same chances of being analyzed with the same finesse?
The answer is no. Not through negligence. By biology. Human working memory manages seven elements simultaneously. Not sixty files. Not 88,000.
If your surgeon had 7 minutes to operate and 60 patients in a row — we would stop him. We would talk about endangerment. We would open an investigation.
To decide if a child is in danger, we find it normal.
The paradox that no one wants to see
We use AI to sort emails. Analyze insurance contracts. Detect bank fraud. Read thousands of invoices in seconds.
This data has value. So we equipped them.
But complaints of violence against minors? Sleeping reports? Repeat offenders who appear in several files without anyone making the connection? This data – the most sensitive there is – is still processed by human brains alone, exhausted, at 60 files per day.
This is the absolute paradox of our time: the more serious the subject, the less we have the tools to analyze it with precision.
An accountant has software so as not to miss a line of numbers. A doctor has a scanner to see what the eye cannot see. A child protection investigator has a stack of files and his seven working memory items.
France has 10.9 magistrates per 100,000 inhabitants. The European average is 21. We spend 100 euros per capita per year on justice. The European average is 200.
We have not suffered from this underfunding. We chose it.
And in the absence of doubling the number of magistrates – which would take fifteen years of recruitment and training – there is a solution that can be deployed in a few months.
What AI won’t do — and why it’s its strength
Let’s be specific, because this is where the debate systematically goes off the rails.
Nobody is talking about replacing the judge. Nobody talks about an algorithm that condemns. The decision must remain human — it is a fundamental principle, and it is non-negotiable.
But between “AI decides” and “AI does not exist”, there is an immense space that France refuses to occupy.
AI as a tireless instructor. The one who reads everything, crosses everything, reports everything. The one that prepares a factual summary – not a decision – so that the human can decide on the entire file. Not on what he had time to read.
The AI has no personal history that filters its reading. She does not project her experience onto the file she is handling. She is not affected by certain details rather than others depending on what they remind her of. It evaluates what is documented. Which is proven. What is there.
The Ministry of Justice has been experimenting timidly since 2025. The Albert tool is running in a few prosecutors’ offices. But the deployment is progressing at the speed of administration – while the stocks of files are moving at the speed of life.
The annoying question
The operation of 88,000 files was praised. Rightly so. Investigations relaunched. Suspects arrested. Victims who feel “finally believed.”
But let’s ask the question that no one asks: why did it take a tragedy to read these files?
They were there. For months. Years for some. They were sleeping in piles that no one had time to process. It took the death of a child and a national shock for us to mobilize – urgently, in six weeks, at the cost of a “mechanical” delay in everything else – what should have been treated continuously.
And in six months, when the media pressure has subsided, when the staff has returned to their normal assignments, the new files will start piling up again.
Unless we decide, this time, to equip the system.
When we know – and we will know – that an AI could have continuously crossed these files, reported these repeat offenders, brought up these dormant files, and we have not deployed it:
What will we say?
Was it complicated? That it took time? That the administration was not ready?
We said the same thing before each announced disaster.
Conclusion
I am someone who criticizes AI when it deserves to be criticized. I spend my professional life dismantling the hype, the empty promises, the botched deployments.
But there are areas where I no longer have doubts.
Justice is one of them.
Not because the magistrates, the police, the gendarmes are bad. Because what is being asked of them is beyond the biological capabilities of any human brain. 60 files per day. 7 minutes per file. A child’s life by file.
This is not a reform that is needed. He’s a co-pilot.
The one who reads what humans don’t have time to read. The one that crosses what services do not have the bandwidth to cross. The one who doesn’t tire. Who doesn’t forget. Who doesn’t let a complaint from 2022 lie dormant while that of 2025 arrives.
AI will not deliver justice.
But she could finally give him the means to return it.
Denis Atlan is Fractional Chief AI Officer, founder of ENDKOO (Qualiopi organization), European Commission Expert Evaluator (IA), DPO and author of “IA Without Bullshit 2026”. He has deployed more than 200 B2B AI projects in France with a documented median ROI of 159.8%.
Sources: Franceinfo, review of the re-examination of complaints (July 2026) — Ministry of Justice, circular of June 8, 2026 — Danziger S., Levav J., Avnaim-Pesso L., “Extraneous factors in judicial decisions”, PNAS (2011) — CEPEJ, evaluation report of European judicial systems — Ministry of Justice, “Developing AI for Justice” (2025)
