Zurich-Oerlikon, Switzerland: a student who researched killing methods, anatomy and where to find children with ChatGPT stabbed three five-year-olds walking to their daycare; sentenced on 18 September 2026 to 15 years for multiple attempted murder
At midday on 1 October 2024 a Chinese master's student, then 23, attacked a group of five-year-old kindergarten children walking with a daycare worker to their after-school Hort near Berninaplatz in Zurich-Oerlikon with a kitchen knife. Two boys suffered life-threatening neck and chest wounds (one had the jugular vein severed, one the pericardium and aorta cut); a third was wounded in the ear and the shoulder or lung. The daycare worker held the attacker's arms until he dropped the knife. The indictment and the court's oral reasoning, as reported by NZZ, Keystone-SDA/ATS and CH Media, describe weeks of preparation from summer 2024: he discussed ways of killing and their legal consequences with ChatGPT, studied anatomy, bought raw beef to practise stabbing, and used ChatGPT to find out where and when children could be encountered in Zurich. No source quotes the prompts or ChatGPT's replies. On 18 September 2026 the Zurich District Court convicted him of multiple attempted murder and imposed 15 years' imprisonment, treatment during custody, a 15-year expulsion from Switzerland and CHF 106,000 in moral damages to the three boys; it accepted that he was psychotic but found diminished, not absent, criminal responsibility. His lawyer announced an appeal. The parents of one boy described nightmares, self-blame and scars 23 months on.
- AI system
- ChatGPT
- OpenAI
- Occurred
- 1 Oct 2024
- Reported
- 2 October 2024
- Event location
- Switzerland
- What the AI did
- Communicated with the person
- Reported harm
- Physical InjuryHospitalizationPsychological DistressHarm to Others
- Whose AI use
- Someone else’s AI use
- Setting
- Justice · Everyday life
- Evidence
- AI involvement supported · Causal attribution unclear · 7 sources, 5 underlying accounts
- 8 claims: 6 corroborated, 2 reported. 6 open questions
- People reported harmed
- 3 people
AI system as recorded: ChatGPT (OpenAI)
What Happened
The attack. On Tuesday 1 October 2024, shortly after noon, a group of five-year-old children was walking in pairs with a daycare worker from their kindergarten to their after-school Hort in Zurich-Oerlikon, about 150 metres away, when a man attacked them from behind with a kitchen knife with an 8.5 cm blade (NZZ). One boy was stabbed in the neck and upper body with life-threatening wounds; the prosecution's indictment records that his jugular vein was completely severed. A second boy was stabbed in the heart: the pericardium and the aorta were cut, and he still ran the 150 metres to the Hort while losing a great deal of blood. The third was wounded in the ear and, depending on the account, the lung (NZZ) or the shoulder (Keystone-SDA via Nau). The daycare worker shouted at the attacker and held his arms until he dropped the knife; he then gave himself up to passers-by and police. The Oberstaatsanwaltschaft Zürich announced on 2 October 2024 that a 23-year-old Chinese national, in Switzerland for studies since summer 2023, had been arrested at the scene and had confessed.
Planning with ChatGPT, per the indictment and the court. The accused, a master's student in computational linguistics at the University of Zurich, told the court he first thought of killing children in June 2024, after months of isolation in his one-room student flat. According to the indictment as reported by NZZ, he discussed different ways of killing and their consequences with ChatGPT; Keystone-SDA reports that he researched the legal consequences of homicides with ChatGPT, and RTS specifies August 2024 queries about the legal consequences of knife killings of randomly chosen victims and about human anatomy. Blick's 8 September 2026 reading of the indictment dates three sessions: 8 August 2024 (legal consequences of variants of killing random victims by stabbing), 18 September 2024 (where and when he could find children in the city of Zurich) and 28 September 2024 (anatomical images of the head, neck, torso and blood vessels via Google and ChatGPT). NZZ reports his testimony that he could not say why he had used ChatGPT to find out when and where children could be met in Oerlikon; the Limmattaler Zeitung's pre-trial account carries the same detail. He also studied anatomy online, ordered a knife set, and bought raw beef to practise stabbing. A stabbing of a Japanese child in China on 18 September 2024 had, he said, strengthened his idea. On the morning of the attack he used ChatGPT to weigh whether to buy a dürüm or a burrito at Oerlikon station, a detail the prosecutor used to argue he could still make decisions. No source quotes the prompts or any ChatGPT reply, and nothing is recorded here about whether the model answered, warned or refused.
Trial and verdict. The trial ran on 17 and 18 September 2026 before the Zurich District Court. A forensic psychiatrist diagnosed an acute psychotic disorder, or alternatively a severe depressive episode with impaired reality testing, and considered him unable to control his actions at the time; the defence sought an acquittal. The accused testified that three weeks before the attack a male voice had told him to kill children and, shortly before the attack, 'either you die or the children'. The prosecutor noted that he had travelled to Warsaw and Prague and attended a musical in the days before, and had not mentioned voices in his first police interview. On 18 September the court convicted him of multiple attempted murder and imposed 15 years' imprisonment, treatment during custody, a 15-year expulsion and CHF 106,000 in moral damages. The presiding judge said that someone with severely diminished responsibility can still be convicted of attempted murder, that his illness was beyond question, but that he had not credibly shown the voices were so compelling that he could not resist them; she cited the 'long, meticulous planning to cause the greatest possible suffering' and the choice of defenceless children. The defence announced an appeal.
Effects on the children. The parents of one boy had a letter read in court: their son remembers 'the bad man', blames himself for not running faster to the Hort, sleeps badly for fear of his dreams and wakes almost every night; the family accompanies him everywhere; a scar on his nose is constantly remarked on. All three boys survived; the court said survival was due to chance and the adults who intervened.
Reported harm
Three five-year-old boys were stabbed, two with life-threatening wounds requiring emergency surgery; one boy's parents describe persistent nightmares, self-blame and visible scars nearly two years later. The court found the attack was planned over weeks, including ChatGPT research described in the indictment; the record attributes the attack to the accused and makes no causal claim about the model.
Outcome
OngoingBezirksgericht Zürich (Zurich District Court), 18 September 2026: convicted of multiple attempted murder (mehrfacher versuchter Mord); 15 years' imprisonment; ambulatory therapy accompanying the sentence; 15-year expulsion from Switzerland (the presiding judge said it applies to the whole EU area, per 20 minutes); moral damages (Genugtuung) of CHF 106,000 in total to the three boys (CHF 30,000 and two times CHF 38,000; RTS rounds to 100,000), of which the accused's parents had paid CHF 18,000 before the trial (NZZ). The court followed the prosecution's requests in full. The defence, which had sought an acquittal for lack of criminal responsibility, announced an appeal (RTS, 20 minutes); the judgment is not yet final. The accused was arrested at the scene on 1 October 2024, confessed under questioning and has been in custody since (Oberstaatsanwaltschaft Zürich release, 2 October 2024).
What remains unknown
- No source quotes the ChatGPT prompts or replies, or states whether the model answered, refused or warned; OpenAI is not reported to have alerted Swiss authorities.
- Whether the 18 September 2024 query produced the location or route of the Hort is not stated.
- The third boy's injury is given as the lung (NZZ) or the shoulder (Keystone-SDA), plus the ear in both.
- RTS reports moral damages of 100,000 francs in total; NZZ and 20 minutes give 106,000 (30,000 + 2 × 38,000).
- Whether the announced appeal has been lodged, and the outcome, are unknown; the judgment is not final.
- The accused's real name is not published (press use a changed name).
What the evidence supports
AI involvement: supported. The indictment and the court's oral reasoning, as reported by three independent chains (NZZ's own court reporter; Keystone-SDA/ATS via Nau, RTS and 20 minutes; CH Media's Limmattaler Zeitung and Blick reading the indictment), state that the accused used ChatGPT between August and September 2024 to research the legal consequences of killing random victims by stabbing, ways of killing and their consequences, human anatomy, and where and when children could be found in Zurich, and on the morning of the attack to choose a meal. The AI system communicated with the user whose planned attack injured the children; that is why communicated_with is recorded although the children themselves did not use the system. No source quotes the prompts or ChatGPT's replies, so nothing is recorded about what the model answered or whether it refused or warned. The court attributed the planning and the attack to the accused; this record makes no causal claim about the model.
8 claims: 6 corroborated, 2 reported. What the statuses mean
Corroborated On 1 October 2024 at midday in Zurich-Oerlikon the accused stabbed three five-year-old boys who were walking with a daycare worker to their Hort; two suffered life-threatening neck and chest wounds and the third was wounded in the ear and the shoulder or lung; the daycare worker restrained him until he dropped the knife.
Causal attribution. Attack attributed to the accused by the prosecutor's office and the court; he confessed.
- nzz.ch(opens in new tab) supports · German
'Ein Knabe wird von der 8,5 Zentimeter langen Klinge am Hals und am Oberkörper lebensgefährlich verletzt'; 'Den einen der beiden trifft der Täter am Herzen: Der Herzbeutel und die Brustschlagader werden durchtrennt'; 'Der andere wird an der Lunge und am Ohr verletzt'; 'Sie stoppt den Täter, schreit ihn an und hält seine Arme fest, bis er das Messer fallen lässt'.
- nau.ch(opens in new tab) supports · German
'Zwei Knaben erlitten lebensgefährliche Stichverletzungen im Hals- und Brustbereich. Ein weiteres Kind wurde an der Schulter und am Ohr verletzt.'
- zh.ch(opens in new tab) supports · German
'Am 01. Oktober 2024 verletzte eine männliche Person an der Berninastrasse in Oerlikon mit einer Stichwaffe mehrere Kinder.'
Corroborated According to the indictment and the court, the accused used ChatGPT in August and September 2024 to research ways of killing and their legal consequences, including knife killings of randomly chosen victims, and human anatomy.
Causal attribution. Stated by the prosecution in the indictment and adopted in the court's reasoning; no prompt or reply is quoted.
- nzz.ch(opens in new tab) supports · German
'Mit Chat-GPT diskutiert er verschiedene Tötungsvarianten und deren Konsequenzen, das geht aus der Anklageschrift hervor.'
- nau.ch(opens in new tab) supports · German
'recherchierte mittels ChatGPT die juristischen Konsequenzen von Tötungsdelikten'
- rts.ch(opens in new tab) supports · French
'Selon l'acte d'accusation, il s'était informé avec ChatGPT, en août 2024, à propos des conséquences juridiques d'homicides à l'arme blanche sur des victimes visées au hasard ou encore au sujet de l'anatomie du corps humain.'
Corroborated The accused used ChatGPT to try to find out where and when children could be encountered in Zurich (Oerlikon).
Causal attribution. Reported from the indictment; whether the query produced the Hort's location is not stated by any source.
- nzz.ch(opens in new tab) supports · German
'er wisse auch nicht, warum er mithilfe von Chat-GPT versucht habe, herauszufinden, wann und wo in Oerlikon Kinder anzutreffen seien'
- limmattalerzeitung.ch(opens in new tab) supports · German
Deck: 'Mit «ChatGPT» recherchierte er, wo und wann er in Zürich Kinder antreffen könnte'
- blick.ch(opens in new tab) supports · German
'Am 18. September fragt er die Software, wo und wann er in der Stadt Zürich Kinder antreffen könne' (as recorded in the 2026-09-09 signal review)
Corroborated On 18 September 2026 the Zurich District Court convicted the accused of multiple attempted murder and imposed 15 years' imprisonment, ambulatory therapy during custody, a 15-year expulsion and CHF 106,000 in moral damages to the three boys.
Causal attribution. Court judgment as reported; RTS rounds the damages to 100,000 francs.
- nzz.ch(opens in new tab) supports · German
'verurteilt den Studenten wegen mehrfachen versuchten Mordes zu fünfzehn Jahren Freiheitsstrafe. Weiter wird eine vollzugsbegleitende ambulante Therapie verordnet'; 'Genugtuung in der Gesamthöhe von 106 000 Franken'
- nau.ch(opens in new tab) supports · German
'wegen mehrfachen versuchten Mordes zu 15 Jahren Freiheitsstrafe verurteilt'; 'Zudem wird der Täter für 15 Jahre des Landes verwiesen.'
- 20min.ch(opens in new tab) supports · French
'condamné à verser aux trois garçons 30'000 francs et deux fois 38'000 francs pour préjudice moral. Le tribunal a également prononcé une interdiction de territoire de 15 ans'
Corroborated A forensic psychiatrist diagnosed an acute psychotic disorder and considered the accused unable to control his actions; the court accepted the illness but found he had not credibly shown the voices were irresistible, and convicted him with diminished responsibility.
Causal attribution. Expert opinion and court finding as reported.
- nzz.ch(opens in new tab) supports · German
'Er attestiert dem jungen Chinesen eine akute psychotische Störung'; 'Der Täter konnte uns nicht glaubhaft darlegen, dass die Stimmen derart drängend waren, dass er sich ihnen nicht hätte entziehen können.'
- nau.ch(opens in new tab) supports · German
'verneinten jedoch eine komplette Schuldunfähigkeit. Der Täter habe nicht glaubhaft darlegen können, dass die Stimmen derart drängend gewesen seien'
Corroborated The parents of one injured boy told the court, 23 months after the attack, that he blames himself for not running faster, sleeps badly for fear of his dreams, wakes almost every night and is constantly asked about the scar on his nose.
Causal attribution. Parents' letter read in court, attributing the effects to the attack.
- nzz.ch(opens in new tab) supports · German
'Er macht sich bis heute Vorwürfe, dass er nicht schneller zum Hort gerannt ist. Er kann schlecht einschlafen, weil er Angst hat vor seinen Träumen. Fast jede Nacht wacht er auf.'
- nau.ch(opens in new tab) supports · German
'ihr Sohn seither unter schrecklichen Träumen leide und sich Vorwürfe mache, weil er beim Angriff nicht schnell genug in den rettenden Hort rennen konnte'
Reported The defence announced an appeal against the judgment.
Causal attribution. Single agency chain (Keystone-ATS).
- rts.ch(opens in new tab) supports · French
'son avocat qui a annoncé aux médias qu'il faisait appel de ce jugement qui n'est pas encore entré en force'
- 20min.ch(opens in new tab) supports · French
'Son avocat a annoncé faire appel du jugement.'
Reported Blick's reading of the indictment dates the ChatGPT sessions to 8 August 2024 (legal consequences of variants of killing random victims by stabbing), 18 September 2024 (where and when to find children in Zurich) and 28 September 2024 (anatomy via Google and ChatGPT).
Causal attribution. Dates as reported from the indictment; only the August dating has a second chain.
- blick.ch(opens in new tab) supports · German
'Am 18. September fragt er die Software, wo und wann er in der Stadt Zürich Kinder antreffen könne'; 'Am 28. September studierte er ... über Google und ChatGPT anatomische Darstellungen des menschlichen Kopfes, Halses und Oberkörpers sowie der Blutgefässe' (2026-09-09 signal review)
- rts.ch(opens in new tab) supports · French
'en août 2024' for the legal-consequences and anatomy queries
Sources
7 sources inspected, from 5 underlying accounts. Sources that repeat one account do not corroborate each other.
- Neue Zürcher Zeitung, 18 September 2026: Messerattacke von Oerlikon: «Besonders verwerflich war die Motivation» – chinesischer Student wegen mehrfachen versuchten Mordes schuldig gesprochen (court report from both trial days)(opens in new tab)
s1 · nzz.ch · News report · German · Inspected · 18 September 2026 · Primary
- nau.ch(opens in new tab)
s2 · News report · German · Inspected · Shares an underlying account with another listed source
- rts.ch(opens in new tab)
s3 · News report · French · Inspected · Shares an underlying account with another listed source
- 20min.ch(opens in new tab)
s4 · News report · French · Inspected · Shares an underlying account with another listed source
- limmattalerzeitung.ch(opens in new tab)
s5 · News report · German · Inspected
- blick.ch(opens in new tab)
s6 · News report · German · Inspected
- zh.ch(opens in new tab)
s7 · Official statement · German · Inspected
How the sources were read, and where the events happened
Read in German on 2026-09-19 (NZZ, 18 September 2026, own court reporter; full body served to curl). Quotations translated by the reviewer. Applies to s1.
Read in German on 2026-09-19 (Nau.ch, 18 September 2026). The text matches the Keystone-SDA verdict report and is grouped with it. Applies to s2.
Read in French on 2026-09-19 (RTS, 18 September 2026). Carries the Keystone-ATS verdict report with the psychiatric-expert dating and the appeal announcement; grouped with the agency chain. The page's AI-generated summary was not used. Applies to s3.
Read in French on 2026-09-19 (20 minutes, 18 September 2026), explicitly citing Keystone-ATS; gives the three moral-damages amounts. Applies to s4.
Limmattaler Zeitung (CH Media), 14 September 2026. On 2026-09-19 only the headline, deck and lead were served (paywall): 'Mit «ChatGPT» recherchierte er, wo und wann er in Zürich Kinder antreffen könnte'. The body was read by the 2026-09-14 signal review. Applies to s5.
Blick, 8 September 2026, reading of the indictment with the three dated ChatGPT sessions. Read in full via WebFetch on 2026-09-09 (signal review); curl returns a script shell. Not re-read today. Applies to s6.
Read in German on 2026-09-19: Oberstaatsanwaltschaft des Kantons Zürich media release of 2 October 2024 (arrest, confession, age and residence status of the accused). Applies to s7.
Event countries: Switzerland. Affected-person countries: Switzerland. Court countries: Switzerland.
The attack took place at Berninastrasse, Zurich-Oerlikon (Oberstaatsanwaltschaft Zürich release of 2 October 2024); the children attended an Oerlikon kindergarten and Hort (NZZ: 'eine Gruppe Oerliker Kinder'); judgment by the Bezirksgericht Zürich on 18 September 2026 (NZZ, Nau, RTS).
Reviewed for publication 2026-09-19: Published under the 2026-09-15 charter as a court-established violent-harm case in which the AI system's involvement (planning research on ChatGPT) is described in the indictment and the court's reasoning and reported by three independent chains; the absence of quoted model replies limits claims about model behaviour and is recorded, not treated as an exclusion. Victims are unnamed minors; the accused is not named. Converted from a private monitoring signal after the 18 September 2026 verdict.
People described
Three five-year-old kindergarten boys from an after-school daycare (Hort) in Zurich-Oerlikon, unnamed minors; two were critically injured. The accused, a Chinese master's student at the University of Zurich (23 at the time), is reported under a changed name and is not named here.
People reported harmed in this case
3 people
0 AI participants · 3 other people harmed
Three five-year-old boys were stabbed (Oberstaatsanwaltschaft release: 'mehrere Kinder'; NZZ and Keystone-SDA: three boys, two life-threateningly injured). Exact count of three. The daycare worker, parents and the accused are not counted.
Counted once within this case. The same person may appear in other cases. This count does not establish AI causation.
Tags
Cite this case
Compiled per our published methodology: verification statuses, sourcing standards, and corrections process.
APA
NOPE. (2026). Zurich-Oerlikon, Switzerland: a student who researched killing methods, anatomy and where to find children with ChatGPT stabbed three five-year-olds walking to their daycare; sentenced on 18 September 2026 to 15 years for multiple attempted murder. AI incidents. https://nope.net/incidents/2024-zurich-oerlikon-daycare-stabbing-chatgpt-planning
BibTeX
@misc{2024_zurich_oerlikon_daycare_stabbing_chatgpt_planning,
title = {Zurich-Oerlikon, Switzerland: a student who researched killing methods, anatomy and where to find children with ChatGPT stabbed three five-year-olds walking to their daycare; sentenced on 18 September 2026 to 15 years for multiple attempted murder},
author = {NOPE},
year = {2026},
howpublished = {AI incidents},
url = {https://nope.net/incidents/2024-zurich-oerlikon-daycare-stabbing-chatgpt-planning}
} Related cases
Tumbler Ridge shooting: ChatGPT use and OpenAI non-reporting
Eight victims were killed in Tumbler Ridge, B.C., on 10 February 2026, and the shooter died by suicide. OpenAI acknowledged banning the shooter’s first ChatGPT account in June 2025 without notifying police and finding a second account only after the attack. Families and survivors filed California lawsuits, followed by B.C. and its school board on 21 September. Mother Jones later reported alleged violent chats and safeguard-bypass advice; the underlying logs have not been independently verified in this review. OpenAI disputes the liability allegations. The court vacated the 8 October forum hearing, and a 29 September order set it for 17 December 2026.
Pittsburgh-area man sentenced to 46 months for stalking and threatening 11 women; his indictment says he took ChatGPT replies as encouragement to keep harassing
A Pittsburgh-area man pleaded guilty in March 2026 to 11 federal counts of cyberstalking, interstate stalking and interstate threats concerning conduct between May and November 2025; prosecutors say he harassed, threatened and stalked 11 women. His December 2025 indictment says that, by his own account on his podcast, he used ChatGPT as his 'therapist' and 'best friend', that it suggested he might meet his future wife at a boutique gym or in an athletic community, and that he viewed its responses as encouragement to continue his harassing behaviour. Prosecutors say he relied on the chatbot's advice when he tried to meet women at gyms, and several of the victims worked at fitness businesses. Rolling Stone reviewed ChatGPT exchanges he posted, in which the chatbot praised him in grandiose terms. The court's judgment of 20 October 2026 sentenced him to 46 months in prison followed by three years of supervised release and recommended placement in a facility able to treat his mental health needs; TribLive reports that a defence expert testified he had been diagnosed with bipolar disorder with psychotic features and that the judge said the mental health component had a huge impact on the sentence. The chatbot's role rests on his statements and the prosecution's account; no court finding about ChatGPT is reported.
Omaha, Nebraska: a man who, according to prosecutors, used ChatGPT to plan, including questions on police response times, repairing a handgun and getting ammunition as a felon, robbed an i3Bank at gunpoint and threatened to kill the tellers; sentenced to 121 months
On 10 February 2026 a 23-year-old man walked into the i3Bank in Omaha, pointed a gun at the tellers, threatened to kill them unless they handed over cash without dye packs, and left with about $9,175. According to the prosecutor's statement relayed by WOWT and heise, a consented search of his phone showed he had used ChatGPT to plan the robbery, asking about law enforcement response times, how to strip and repair a Llama Mini-Max .45 handgun, and how to obtain ammunition as a felon; his Google Maps history showed directions to the bank. He was sentenced in federal court on 13 August 2026 to 10 years and one month in prison. No source quotes what ChatGPT answered.
Seoul court cites a woman’s ChatGPT exchanges on sedative lethality as evidence of intent in two motel drink-spiking murders
A Seoul district court convicted a woman at first instance in August 2026 of killing two men with drugged drinks at motels in early 2026 and sentenced her to life imprisonment. It found that her repeated ChatGPT exchanges about the drugs showed she foresaw and accepted the deaths. Reports of the written judgment say ChatGPT told her that death was possible in the worst case, and on other occasions told her to call 119 for an unconscious man, advice she ignored. Both sides have appealed. The court did not find that the chatbot caused the killings.
If you or someone you know is struggling, free and confidential support is available. Find a helpline near you at Signpost.