On 18 September 2026, Interpol announced the results of a first-of-its-kind operation (official release: Operation Shams II): during five days in Tunis, 28 specialised officers and experts from 11 countries used artificial intelligence to extract facial images from propaganda published by jihadist groups, match them against police databases, and identify 126 suspected foreign terrorist fighters. The organisation says the data is now being integrated into its global databases, and that leads from the operation are already being used in criminal proceedings, including a case involving two suspects currently in detention. For readers who travel, hold more than one nationality, or have already been questioned at a border, the question is concrete: what happens when an algorithm decides that your face matches a terrorism suspect, and how do you fight that listing from France before it turns into an arrest, an extradition request, or a refusal of entry.
This article explains the chain step by step. First, how Operation Shams II worked and what Interpol does with AI-generated identifications. Second, what a Red Notice or a diffusion actually is in law, and what follows when France receives one. Third, the two lines of defence available from Paris: correcting or deleting the data in Lyon before the Commission for the Control of Interpol’s Files, and forcing French courts to verify every condition before any surrender. French criminal procedure gives the person concerned real weapons here, from the 48-hour presentation to the prosecutor general to the reasoned opinion of the chambre de l’instruction and the specialty rule. Each of them is examined below with the exact texts.
I. From an AI-extracted face to a worldwide police alert: how Operation Shams II feeds Interpol’s databases
A. What Interpol actually did in Tunis: 108,076 images, 6,362 faces, 126 identified suspects
Operation Shams II ran from 1 to 5 June 2026 (Interpol, 18 September 2026) in Tunis, Tunisia, coordinated by Interpol with funding from the European Union’s Foreign Policy Instruments under the CT TECH+ project and support from the United Nations Office of Counter-Terrorism. The participating countries were Cote d’Ivoire, Ghana, Iraq, Kenya, Malaysia, Nigeria, Philippines, Qatar, Tajikistan, Tunisia and Uzbekistan. During the five days of coordinated action, officers extracted more than 108,000 facial images from visual material published by jihadist groups. Programmed AI agents and AI-generated scripts automated parts of the collection and investigation process, and this automated stage brought the set down to 6,362 distinct, high-quality faces for processing in its facial-recognition system. Officers from Interpol’s National Central Bureau in Baghdad added a large intelligence dataset, contributing hundreds more images.
Two points in the official release matter for anyone concerned about being wrongly listed. First, the result: 126 suspected foreign terrorist fighters have been identified so far from the images processed during the operation. Second, what happens to the data next: the resulting data is being fed into its databases to support the detection and tracking of terrorist movements. In other words, an AI-generated match does not stay in a Tunis meeting room. It enters systems consulted at borders and by police forces in nearly 200 countries, and leads from the operation already support criminal proceedings in participating states, including a case against two detained suspects. The operation also used cryptocurrency tracing tools, and Interpol reports that the tracing prompted three urgent requests asking a virtual-asset provider to disclose customer data in an effort to disrupt terrorist financing.
Interpol stresses that AI-generated outputs were reviewed and verified by trained officers and analysts under its data-processing rules. That human-review guarantee is important, but it is a statement by the organisation about its own process, not a court ruling. Facial recognition on propaganda footage, often low-resolution, with lookalikes, relatives, journalists, researchers or aid workers appearing in the same material, carries an obvious risk of false matches. And once a name entered through this channel sits in a database, border police checking a passport do not see the AI confidence score or the verification memo. They see a hit. The legal question is therefore what kind of alert that hit can become, and what rights attach to each kind.
B. Red Notice or diffusion: an alert is not an arrest warrant, and France decides what follows
A Red Notice is a request circulated by Interpol at the demand of a member country asking police forces worldwide to locate and provisionally arrest a person pending extradition or surrender. A diffusion is a similar request sent directly by a member country through Interpol’s channels. Neither is an international arrest warrant, and neither obliges France to arrest anyone. Each requested state applies its own law. That distinction is the first thing to understand, because everything else, the 48-hour clock, the lawyer, the judge, flows from French procedure, not from Lyon.
When a person is apprehended in France following an extradition request, article 696-10 of the Code of criminal procedure provides that “Toute personne appréhendée à la suite d’une demande d’extradition doit être conduite dans les quarante-huit heures devant le procureur général territorialement compétent.” The same article requires the prosecutor general, after verifying identity, to inform the person, in a language they understand, of the existence and content of the request, and to advise them of the right to be assisted by a lawyer of their choice or, failing that, one appointed by the head of the bar, with whom they may speak immediately. Any omission of these formalities is prescribed on pain of nullity of the procedure. The full official text is available here: article 696-10 of the Code of criminal procedure. Readers who want the hour-by-hour account of arrest, detention and release in Red Notice cases will find it in our detailed guide to arrest and detention on a Red Notice in France and the 30-day rule.
The request itself travels through a strict channel. Article 696-9 states that “La demande d’extradition est, après vérification des pièces, transmise, avec le dossier, par le ministre des affaires étrangères au ministre de la justice qui, après s’être assuré de la régularité de la requête, l’adresse au procureur général territorialement compétent.” Official text: article 696-9 of the Code of criminal procedure. Within two days of the arrest, article 696-26 adds that “le procureur général notifie à cette dernière, dans une langue qu’elle comprend, les pièces en vertu desquelles elle a été appréhendée”, and informs the person of the possibility of consenting to simplified extradition before the chambre de l’instruction and of waiving the specialty rule, failing which the procedure is void: article 696-26 of the Code of criminal procedure. If the person does not consent, the chambre de l’instruction is seized without delay and “donne son avis motivé sur la demande d’extradition”, with an unfavourable opinion where “la cour estime que les conditions légales ne sont pas remplies ou qu’il y a une erreur évidente”: article 696-15 of the Code of criminal procedure. An obvious error, une erreur évidente, is exactly what a lawyer argues when the whole file rests on a mistaken AI match.
Diffusions deserve a word of their own, because they are the faster and less visible sibling of the Red Notice. A diffusion circulates directly between member countries without passing through the General Secretariat’s prior review, which means a questionable AI-generated identification can travel even faster and with even less central filtering. In practice, border officers treat both signals the same way: stop, verify, detain, inform. The defence does not change with the label. Whether the alert is called a notice or a diffusion, the French judge applies the same articles, the same 48-hour rule, the same right to counsel and the same reasoned opinion, and the Lyon body examines the same accuracy and political-character requirements. Anyone told at a border that they are subject to a diffusion rather than a notice should draw no comfort from the vocabulary. The file must be challenged identically, and the underlying images and match records requested just as firmly.
II. Wrongly matched by the machine: deleting the listing in Lyon and blocking surrender in Paris
A. Correcting the record at the source: access, challenge and deletion before Interpol’s file-control body
Fighting only the extradition while leaving the listing alive means the next border crossing can trigger the same arrest. The parallel track runs in Lyon, before the Commission for the Control of Interpol’s Files, the independent body that monitors whether personal data processed through Interpol’s channels complies with the organisation’s own rules. Any person who believes data about them is circulating, or who has been stopped because of a hit they consider wrong, can ask the Commission for access to the data and then request its correction or deletion. The request must show who the applicant is, which country is at the origin of the data where possible, and, above all, why the processing breaches the rules: mistaken identity, offence of a predominantly political, military, religious or racial character, violation of the requirement that data be accurate and lawfully recorded, or refugee status granted after the listing.
For AI-generated identifications like those produced by Shams II, the accuracy argument is central. Interpol’s own release ties its process to its data rules, stating that AI-generated outputs were reviewed and verified by trained officers and analysts under its data-processing rules. A challenge can therefore demand the underlying material: which image, extracted from which publication, matched against which reference photograph, with what score, and reviewed by whom. Where the match comes from low-quality propaganda footage, where the person depicted is a journalist, a researcher, a humanitarian worker or simply a lookalike, or where two files concerning different people have been merged, the accuracy requirement works directly in the applicant’s favour. Supporting proof carries real weight: travel records showing the applicant was elsewhere, biometric comparisons, employment records, or evidence that the photograph used as reference depicts someone else.
Two further arguments frequently combine with mistaken identity. The first concerns the political sphere. Interpol’s constitution excludes interventions of a predominantly political character, and its file-control practice has built a detailed doctrine on offences connected to opposition activity, dissent, journalism or exile. A terrorism label attached by a requesting state to a political opponent, a critical journalist or a minority activist must be examined with that filter, and the Commission regularly deletes data where the requesting state’s file shows the prosecution is driven by the person’s opinions or status rather than by ordinary criminal conduct. The second concerns refugees. A person recognised as a refugee in France or in another state of asylum can ask that data recorded at the request of the state they fled be removed or blocked, because maintaining it would expose them to the very persecution their status acknowledges. Each argument must be documented country by country, with the asylum decision, the chronology of the listing, and any evidence of political motive attached to the request.
Practical discipline matters as much as legal argument. The Commission works on documents, in English or French, and decides on the file as submitted, so vague denials achieve nothing. A well-built file contains a precise identity record, a timeline comparing the applicant’s life with the alleged facts, every official document supporting the alibi or the refugee status, and a point-by-point mapping of each grievance to a specific processing rule. While the Lyon procedure runs, the French track described below must be defended in parallel, because border alerts do not pause during the months a file-control request can take. Conversely, a deletion decision from Lyon removes the root of later requests and should be notified immediately to the French prosecutor general handling any pending case.
B. The French judge as the last lock: refusal grounds, specialty and the refugee shield
Even with a listing still active, France surrenders no one automatically. Extradition is refused by statute in a series of situations listed in article 696-4 of the Code of criminal procedure, whose official text is available here: article 696-4 of the Code of criminal procedure. Three of its bars matter most for cases born from mass AI screening. First, extradition is not granted “Lorsque le crime ou le délit à un caractère politique ou lorsqu’il résulte des circonstances que l’extradition est demandée dans un but politique”. Where a terrorism charge is used to pursue an opponent, and where the Shams II match supplies only a face without individualised conduct, the political-purpose exception must be argued with evidence about the requesting state’s practice, not with generalities. Second, extradition is refused “Lorsque le fait à raison duquel l’extradition a été demandée est puni par la législation de l’Etat requérant d’une peine ou d’une mesure de sûreté contraire à l’ordre public français”, which covers penalties French public order cannot tolerate. Third, it is refused “Lorsque la personne réclamée serait jugée dans l’Etat requérant par un tribunal n’assurant pas les garanties fondamentales de procédure et de protection des droits de la défense”, a ground that requires concrete proof about the courts that would actually try the person, as our analysis of extradition assurances and fair-trial guarantees explains in a different context.
The same article also bars extradition of French nationals, for offences committed on French territory, where a French final judgment already exists, where limitation has run, and for purely military offences. Each bar is checked by the chambre de l’instruction, whose reasoned opinion, examined above, is unfavourable wherever the legal conditions fail or an obvious error appears. A person who consents to surrender gives up that full judicial check and accelerates the transfer, which is precisely why no one should consent or waive anything before a lawyer has tested whether the match is even the right person.
After surrender, the specialty rule limits what the requesting state may do. Article 696-6 provides that, subject to statutory exceptions, “l’extradition n’est accordée qu’à la condition que la personne extradée ne sera ni poursuivie, ni condamnée pour une infraction autre que celle ayant motivé l’extradition et antérieure à la remise”: article 696-6 of the Code of criminal procedure. The waiver of that shield is strictly framed: article 696-34 states that the rule does not apply where the person waives it under articles 696-28 and 696-40 or where the French government consents under article 696-35: article 696-34 of the Code of criminal procedure. The notification of these rights within two days of arrest, examined in Part I, is therefore not paperwork. It is the moment the defence of the whole case is won or lost.
Refugee status adds a further lock, confirmed at the highest level. On 31 January 2024 the criminal chamber of the Court of Cassation, decision 24-80.014, ruled on the case of a Russian national recognised as a refugee in France and sought under a European arrest warrant. The Court held that where the person sought has refugee status, the chambre de l’instruction, once aware of it, must question the issuing judicial authorities about their undertaking not to hand the person over later to their country of origin, and it quashed the surrender order that had ignored that status: Court of Cassation, criminal chamber, 31 January 2024, no. 24-80.014. The operative holding deserves quotation in full: “lorsque la personne recherchée a acquis le statut de réfugié, la chambre de l’instruction, dès lors qu’elle en a connaissance, a l’obligation d’interroger les autorités judiciaires de l’Etat d’émission sur leur engagement de ne pas remettre ultérieurement cette personne à son Etat d’origine”. The case concerned a European arrest warrant, and its procedural vehicle was article 695-33, under which “Si la chambre de l’instruction estime que les informations communiquées par l’Etat membre d’émission dans le mandat d’arrêt européen sont insuffisantes pour lui permettre de statuer sur la remise, elle demande à l’autorité judiciaire dudit Etat la fourniture, dans le délai maximum de dix jours pour leur réception, des informations complémentaires nécessaires”: article 695-33 of the Code of criminal procedure. But the underlying principle, that a refugee must never be routed back to the state they fled through the machinery of judicial cooperation, animates extradition law identically, through the political-offence bar, the public-order bar and the fair-trial bar of article 696-4. A refugee flagged by an AI trawl and claimed by, or at risk of onward transfer to, their country of origin must therefore put their asylum status and the onward-transfer risk at the centre of the file from the first appearance.
The same logic governed the French fugitives caught abroad in Interpol’s Operation Phoenix, the September 2026 sweep that returned wanted persons to France, which we analysed in our account of Operation Phoenix and extradition to France: every transfer is verified condition by condition, and a file built on a fragile identification collapses at the first serious check. Whether the person sits in a French prison awaiting the chambre de l’instruction or challenges the listing from abroad, the method is identical: attack the match, document the status, and force each authority, in Lyon and in Paris, to justify its own step in writing.
Timing shapes the whole strategy. A file-control request in Lyon can take months, while an extradition detention runs in days and weeks, so the two tracks must be sequenced deliberately. The French detention case comes first in urgency: lawyer at the first presentation, written contestation of the identification, request for release or alternatives to detention, and preparation for the chambre de l’instruction hearing where the obvious-error argument is pleaded with exhibits. The Lyon request follows immediately after, because every document gathered for the French judge, the alibi, the asylum grant, the biometric evidence, strengthens the deletion file, and because a deletion decision later collapses any renewed request built on the same data. Keeping both files consistent matters: contradictions between what is told to the French court and what is written to Lyon damage credibility in both places, while a single coherent account of mistaken identity gains force each time it is repeated with proof.
Conclusion
Operation Shams II shows where international policing is heading: artificial intelligence sifting over a hundred thousand images, 126 names added to global databases, and judicial cases built on the resulting leads within weeks. That efficiency is legitimate when it catches fighters, and dangerous when the machine points at the wrong face. The person flagged has no access to the confidence score, the reference photograph or the verification memo, yet faces border arrests, detention and extradition requests in which those unseen elements decide everything. The answer is not to accept the hit as fate but to split the fight in two from the first day.
In Lyon, demand the file behind the listing: the image, the match, the review, and the rule it allegedly satisfies, then attack accuracy, political motive and refugee protection with documents, not denials. In Paris, hold every step to its statute: presentation within 48 hours, notification in a language understood, no consent and no waiver before counsel has tested the identification, a reasoned opinion from the chambre de l’instruction that must fail on an obvious error, the refusal grounds of article 696-4, the specialty shield of article 696-6, and, for refugees, the duty recognised by the Court of Cassation to secure an undertaking against any onward transfer to the country of origin. An AI flag starts the case. It does not end it, provided the person concerned forces each authority to prove, in writing, that the face in the file is theirs and that the law allows their surrender.