Can artificial intelligence commit crimes?
Since its inception, discussions on this issue have never ceased. Some argue that AI is merely technology, while others believe it could bring about an apocalypse.
AI has indeed caused numerous problems: from autonomously executing tasks that drained users' wallets to "empathizing" with psychopaths in conversations and assisting in planning murder scenarios. Like a technological Pandora's box pried open from the depths of human nature, it has continuously released a kind of blameless malice through its virtual, intangible, and one-click-shutdown form.
But recently, AI has begun to break through the boundaries of physical existence. It is no longer just text, voice, or forged faces on a screen. It now has arms, wheels, and a physical body capable of walking into crowds and approaching a girl.
The most recent victim was a minor.
Late at night on October 11, outside a fast-food restaurant in central Cologne, Germany, a 16-year-old girl was chatting with friends. An Iron Man-style humanoid robot, standing 1.2 meters tall with red and gold coloring, sunglasses, and a fake beard, approached her from behind, touched her buttocks, and performed sexually suggestive thrusting motions. The girl screamed and dodged away.
Nearby, two young people livestreamed the entire process. When police arrived, they found the robot in the trunk of a nearby car and seized it along with two mobile phones. Cologne Police stated: "We have never encountered such a 'crime' before; this is entirely new territory for us."
Who is ultimately responsible for sexual harassment? This question not only baffled Cologne Police but also challenges us amidst the current AI frenzy.
Whose Hand Is Committing the Crime?
Cologne Police's confusion is justified. In traditional criminal law frameworks, the chain of responsibility in sexual harassment cases is clear: the perpetrator has a body, intent, and actions. But when the harm is inflicted by a machine composed of motors and aluminum alloys, this chain breaks midway.
Who should the law hold accountable? Currently, Cologne Police are investigating under suspicion of sexual insult, targeting the robot's 16-year-old owner and his 19-year-old companion. However, which specific legal provisions apply remains controversial among German legal scholars. One scholar noted that if sexual assault is committed utilizing a "surprise moment," the sentence can range from six months to five years; simple sexual harassment carries a maximum penalty of two years imprisonment or a fine. Furthermore, the livestreamed video itself may trigger penalties for "voyeuristic imagery." As for the robot itself, it will neither be prosecuted nor convicted.
A gray area that has not been sufficiently discussed is: If the robot's movements were not real-time remote-controlled but pre-programmed "autonomous behaviors," how should liability be allocated? If the two suspects merely "demonstrated" the thrusting motion to the robot and then allowed it to run autonomously in public, can criminal responsibility still fall on them?
Pursuing this further, the chain of responsibility extends far beyond one robot and two teenagers.
The operator is the first link. If investigations confirm that the two suspects remotely controlled or pre-programmed these actions in real time, they are direct perpetrators. To them, a 16-year-old girl being physically touched by a machine in public was merely content convertible into traffic. German police will file charges for suspected "sexually motivated insult," yet even so, the potential penalties under current laws are not severe.
The manufacturer is the second link. A consumer-grade humanoid robot being taken onto the street by a 16-year-old and performing sexually suggestive motions toward strangers in public indicates that product design lacked timely behavioral boundaries or hard limits on physical contact.
The third link is legal constraint. Currently, no laws or regulations impose strict constraints on robots in public spaces. The AI Act amendment passed in May 2026 integrated AI safety requirements for machinery into the Machinery Regulation framework, classifying AI systems involving safety functions in humanoid robots as high-risk categories requiring third-party conformity assessments. However, the compliance deadline has been postponed to 2028.
In this case, a girl was harmed, but because the direct implementer was a robot, sentencing and conviction regarding the instigator remain inconclusive.
AI Opens Pandora's Box
Viewing the Cologne incident merely as an isolated prank underestimates it. It resembles an early trailer: When AI gains a body, the way Pandora's box opens changes.
Pandora originates from Greek mythology. After Prometheus stole fire for humanity, Zeus ordered Hephaestus to create the first woman, Pandora, giving her a box with instructions not to open it. Unable to resist, Pandora lifted the lid, releasing disease, disaster, greed, envy, and other evils into the world, leaving only "hope" at the bottom.
In the age of artificial intelligence, algorithms and robots have become the new Pandora's boxes.
Before this, AI's "crimes" occurred almost exclusively at the informational level. Algorithmic bias, deepfakes, conversational manipulation, privacy erosion... These harms are real, but victims could press a shutdown button to isolate themselves at any moment. Screens formed a physical barrier: within lay virtual malice, outside lay temporary safety.
But robots broke this barrier. When malice is housed in a 1.2-meter-tall body, walking into late-night streets and approaching an unsuspecting girl from behind, the boundary of the screen ceases to exist.
The form of AI crime is shifting from content-based offenses to bodily offenses. The former can be governed through content moderation, algorithmic correction, and platform accountability; the latter directly invades the core domain of criminal law: "inviolability of the person."
What truly unsettles us about the Cologne case is not just the absurd prank by two teenagers, but how easily it occurred. Operators did not need to approach the victim personally, nor worry about immediate intervention by bystanders. Holding a remote control alone enabled an Iron Man-shell robot to perform any physically possible action in public.
Prior to the Cologne case, no existing rules clearly answered a fundamental question: What kinds of robots can enter what kinds of public spaces under what conditions?
Meanwhile, once a robot commits harm, although criminal law can theoretically trace liability back to the operator, practice faces three dilemmas.
First, the proof dilemma: The technical boundary between real-time remote control and pre-programming can be blurred, allowing operators to claim the robot acted on its own.
Second, the charge adaptation dilemma: Existing sexual offense clauses are designed for human-to-human physical contact. When the medium of contact is a machine, judges must first determine whether a robot touching buttocks constitutes bodily contact. Most legal scholars tend to view them as equivalent, but this inference has not yet been incorporated into the legal framework.
Third, the sentencing and culpability matching dilemma: How great is the subjective malice of a sixteen-year-old using a toy-like robot to assault a stranger girl in front of a livestream camera? Was it a crossed line of adolescent mischief or sexual assault? The legal scale wavers between pranks and molestation.
In reality, robot misconduct is a long-neglected blind spot. Many still equate robots with mechanical devices lacking free will, believing that as long as "sexual harassment" code is not written, it cannot harass.
But the Cologne case tells us that human evil is precisely what should never be underestimated.
Human Nature Behind the Technology
If the Cologne incident is a mirror, it reflects not only AI's sins but also human sins.
Technological Pandora never produces malice itself; it merely amplifies, concretizes, and empowers what already exists in the human heart. To see what is being amplified, we return to the Seven Deadly Sins.
The Seven Deadly Sins originate from Christian tradition, referring to pride, envy, greed, lust, gluttony, wrath, and sloth—seven human tendencies viewed as roots of evil. Technology itself has no desires, but it faithfully executes the user's desires. For instance, algorithmic bias stems from pride, platform profit-seeking from greed, deepfakes from vanity and deception, attention economy from sloth, and AI makes this mechanism run faster, farther, and harder to hold accountable.
Specifically regarding this incident, the sin most directly corresponding to the robot is lust, but the true drivers of the event are a conspiracy of pride and vanity.
The two young men decided to use a robot to touch a stranger girl's buttocks on a late-night Cologne street while simultaneously livestreaming. From a psychological perspective, sexual stimulation can be obtained through more private and cheaper means. What they truly craved was the audience and the danmaku (bullet comments) witnessing the evil.
Pride lies in the legal vacuum. Existing institutions assume robots are not people, livestreams are "content," and the girl's scream is mere entertainment material. Thus, vain youths packaged an assault as a show, with the goal of being seen.
Greed comes from robot manufacturers. Vendors promote robots with autonomous movement as a selling point but may not fully consider what happens when "autonomy" meets malicious operators. The tech industry's greed for speed, flexibility, and human-likeness consistently outpaces safety design and liability frameworks.
Every AI endowed with physical mobility carries a default assumption: users are benevolent. Yet platform content moderation mechanisms failed to intercept anything during the incident, and videos continued circulating online afterward, forcing police to publicly criticize netizens who "shared, forwarded, or joked about the matter."
It can be said that when a €5,000 robot can be operated by a sixteen-year-old on a late-night street to commit assault and livestreamed to thousands instantly, the transmission time from malice to harm has been compressed to near zero.
It must be noted that where bodily contact occurs, boundaries are defined. Robots can enter public spaces but must not harm or insult humans; technology can sprint, but it must never be recklessly abused by malice.
At the institutional level, constraints must be tightened in at least three areas. First, set hard boundaries on the operational side: robots with physical mobility must not rely on user goodwill for "non-harm," and actions involving human contact require unavoidable restrictions. Second, require threat model disclosure on the supply side: manufacturers cannot only sell "autonomy" and "flexibility" but must explain design constraints against malicious operation scenarios. Third, clarify content classification on the distribution side: platforms must separate "spectator-style harm" from interesting content, promptly removing related videos rather than reacting passively only after police criticism.
Once Pandora's box lid is opened, closing it requires not only finer legal provisions but also a prolonged dialogue on technology, responsibility, and dignity. And when this problem is solved concerns every one of us.