AI Incidents 2025 worldwide

Security researchers at OX Security discovered two malicious Chrome extensions that intercept private conversations from ChatGPT and DeepSeek and send them to a server controlled by the attackers. The extensions posed as well-known legitimate tools and had a combined total of over 900,000 downloads in the Chrome Web Store. One of them even had Google's official featured badge.

A plane of the Swiss airline was pushed into a snowdrift by strong winds at Kittilä Airport in Finland. Realistic-looking images of the incident appeared on the internet.

Over Christmas, there were 15 incidents of graffiti vandalism in several locations in Cornwall (England). Unknown perpetrators sprayed slogans such as 'AI will replace us all' and 'AI will take our jobs.'

A New Zealand media company published a Christmas book list for children in regional newspapers that included several non-existent book titles. Authors were listed with books they had not written. There was speculation that the incorrect titles might have been generated using artificial intelligence. The media company stated that the list was obtained from an external source. It withdrew the list and apologized for the error.

After losing cabin pressure in a Beechcraft King Air B200, the pilots activated the Garmin Autoland emergency landing system. The system took control of the aircraft, selected a suitable airport, communicated with air traffic control, and brought the aircraft safely to the runway. Not a true AI system, but highly autonomous behavior, therefore included here.

The AI image generator Grok on platform X was increasingly used to generate sexualized images of real people without their consent.

The Heber City Police Department in Utah tested two AI systems to transcribe body camera footage and generate reports from it. One report claimed that a police officer had turned into a frog. The reason for this was that the Disney movie 'The Princess and the Frog' was playing in the background, and the contents of the movie were also incorporated into the report. This highlights the need to review AI-generated reports. However, due to the immense time savings, the intention is to continue using AI.

Before the social court division of the Tribunal judiciaire de Périgueux, a benefit recipient disputed a claim for repayment (indu) of approximately €3,777 for basic disability allowance (AAH, Allocation adulte handicapé) from the social security agency CAF (Caisse d’Allocations Familiales) on the grounds of alleged lack of habitual residence in France. The court declared the underlying control procedure null and void and canceled the repayment claim. In doing so, it expressly pointed out that the case law references cited by the plaintiff did not appear to correspond to published decisions and requested the plaintiff and his lawyer to check in future that references found via search engines or generative AI are not hallucinations.

In insolvency proceedings before the Ghent Commercial Court, the company sought the reopening of the debates. The court found that the request relied on non-existent case law, likely resulting from uncontrolled use of AI, and considered this conduct to be in bad faith and disruptive to the proceedings.

The family of a woman who has been missing since June 2025 received an alleged photo showing the woman bound in a van in order to extort money for her supposed release. Missing person posters were presumably used to generate the image.

The Hamburg Administrative Court rejected the urgent application of a ninth-grade student who had filed a lawsuit challenging the assessment of his reading log. He sought to have the accusation of deception for using ChatGPT withdrawn or not further communicated pending the main proceedings. The court considered the assessment as an attempt at deception and the grade 'unsatisfactory' to be lawful and therefore rejected the application for interim relief.

In an injunction proceeding before a court in the Argentine province of Salta, a defendant had filed an appeal but failed to substantiate it within the deadline. The court then declared the appeal void (desierta). At the same time, the judge noted that the brief contained contained serious inconsistencies and suspected that it had been drafted using AI. She admonished the lawyer, stating that while the use of Artificial Intelligence is not prohibited, the lawyer remains fully responsible for the content.

A video about an alleged coup in France is posted on Facebook by unknown individuals. An African head of state expresses concern about the situation in France.

In proceedings before the Labor Chamber (Cámara del Trabajo) of the 1st Judicial District in Viedma, a long-term employee of a bakery disputed, among other things, indirect dismissal (resignation for good cause), incorrect registration of the start of employment, outstanding wage components, and social security contributions. The court dismissed most of the claims, but ordered the employer to pay wage differences and rejected personal/joint liability on the part of the co-defendants. Finally, the court addressed the alleged use of generative AI: the plaintiff had cited case-law references that could not be located, which indicated AI-generated or hallucinated references. However, the court did not impose any sanctions, as the incident occurred before a relevant regulation came into force.

A plaintiff filed a claim for damages with the District Court in Wrocław (Sąd Okręgowy we Wrocławiu), the statement of claim having been drafted entirely with the assistance of ChatGPT. In doing so, the AI relied on inaccurate or non-existent legal bases and misrepresented statutory provisions. The court dismissed the claim as unfounded and clarified that responsibility for the content of the statement of claim rests with the plaintiff.

During a Tesla presentation in Miami, the humanoid robot 'Optimus' fell to the ground. A noticeable hand movement shortly before the fall sparked speculation that the robot was not autonomous but instead remotely controlled.

A citizen challenged a municipal decision regarding a fine for waste disposal before the Grenoble Administrative Court. The court dismissed the application as manifestly inadmissible, among other things because the contested decision was not submitted and the submission remained unclear. In its reasoning, the court stated that the statement of claim had clearly been generated using a generative AI tool.

The proceedings concerned a constitutional complaint against a decision by the Supreme Administrative Court. Upon preliminary review, the Constitutional Court found that the complaint contained numerous serious deficiencies: it cited non-existent court decisions or misrepresented the content of existing decisions. The complaint thus exhibited typical signs of unverified AI-generated content. In the court's opinion, this misled the court and significantly impeded the proper conduct of the proceedings. The Constitutional Court therefore imposed a procedural sanction of CZK 25,000 (approx. $1,200) on the lawyer.

A high-end fashion brand has released an AI-generated video. It has been widely criticized for appearing unnatural and low-quality.

In a family law appeal before the Federal Circuit and Family Court of Australia (Division 1), Appellate Jurisdiction, the main issue became a costs ruling after the appeal was discontinued shortly before the hearing. The court found that artificial intelligence had been used in filed submissions, and that incorrect or unreliable citations were included. Because the extent and nature of the AI use were unclear and professional duties (such as candour and verification) were engaged, the court made orders including steps to refer the legal representatives to relevant regulatory bodies and addressed costs linked to the AI issue. The reasons also highlighted confidentiality and secrecy risks when court materials are input into AI systems.

In a probate-related matter before the Supreme Court of Victoria, the court also addressed the reliability of written submissions filed in the proceeding. A practitioner had used AI tools to prepare parts of the submissions, contrary to court guidance, and the material included non-existent citations. The court emphasized that AI use makes careful source-checking essential.

Several AI-generated videos went viral on Facebook, garnering millions of views and tens of thousands of shares in just days. The clips contain manipulated content that appears to amplify social tensions and sometimes reflects Russian political rhetoric.

In a segment on the program 'Agenda' about US immigration policy, a short video is shown in which a New York police officer yells at an ICE immigration officer. However, the video was generated by AI.

At New Zealand's most prestigious literary awards, the Ockham New Zealand Book Awards, two books by well-known New Zealand female authors were disqualified because their book covers were created using artificial intelligence. The reason for this was a new rule that prohibits any use of AI in the entire book project, including cover design. The disqualification was made irrespective of the literary content, which was written entirely by humans. The decision was criticized because authors usually have no influence on the cover and the rule was introduced only after the books had been submitted. Following this public criticism and after further discussions, the organizers later reversed their decision and allowed the affected titles to re-enter the competition in 2026.

In summary proceedings before the Administrative Court of Cologne, a party sought provisional access to files in a municipal law context; the request was denied. The court noted several erroneous and inaccurate citations in the submission, which in its view suggested that the application may have been drafted using an AI tool and its output adopted without review. The court did not carry out a separate legal assessment of, nor impose sanctions for, the use of AI.

In criminal proceedings for serious sexual abuse in the province of Salta, the defense lodged an appeal against the decision to admit the case to trial. The court dismissed the appeal as inadmissible, as this decision is not contestable under current law. At the same time, the judge objected to the brief submitted because it contained clear indications of unverified AI use, including incorrect or irrelevant legal references, erroneous content, and unfilled placeholders. The judge emphasized that the use of Artificial Intelligence is permissible, but that the lawyer remains responsible. The lawyer was asked to provide evidence of the cited case law, otherwise the Bar Association's ethics tribunal would be called in.

In proceedings before the Darmstadt Regional Court, the issue was whether a court-appointed expert was entitled to remuneration under Germany's expert-fees statute (JVEG). The court found the submitted 'expert opinion' on the assessment of a physical injury to be unusable and set the expert's fee at €0.00. In addition to a lack of examination of the person, a decisive factor was that the expert had apparently made extensive use of artificial intelligence, but had not made it transparent to the court, undermining the requirement of personal expert work and accountability.

In the context of an employment lawsuit brought by a former employee against her former employer, a US healthcare system, the plaintiff's attorney submitted a brief citing numerous inaccurate or non-existent court decisions. The court examined the attorney's conduct in light of her duties of care and candor to the court, as well as her compliance with local procedural rules regarding the disclosure of possible AI use. It concluded that she had violated her professional duties of care and reaonable inquiry. The underlying lawsuit was resolved by settlement.

The AI-generated image appeared after a false fire alarm was triggered at the school, caused by a freon gas leak from a refrigerator.

In a tender for road maintenance, the company submitted the lowest bid at €3.7 million. To justify the low price, it referred to AI-generated tax rulings that did not exist. As a result, the company was excluded from the tender by the National Appeal Chamber (Krajowa Izba Odwoławcza, KIO).

When assessing the risk of a soccer match between Maccabi Tel Aviv and Aston Villa in November 2025, the police referred to a situation report that cited an alleged previous match between Maccabi Tel Aviv and West Ham United as evidence of an increased security risk. However, this match had never taken place, but was the result of AI hallucinations. The report was used in the decision to exclude Maccabi fans from the game.

Two men wanted to cross a causeway to Sully Island (Bristol Channel) at low tide and then walk back. They relied on the tide times provided by ChatGPT. However, the AI gave a low tide time that was approximately two hours off, meaning that the causeway was already flooded when they planned to return, leaving the two men stranded on the island. They were rescued by the coast guard.

The Salem, Massachusetts police is investigating three cases of alleged intruders from the 'AI Homeless Man Prank'.

A 15-year-old girl wrote to her mother that an unknown man had rung the doorbell. The girl sent her mother a picture showing the man lying on the sofa at home. The mother then called the police. It turned out that the picture had been generated using AI. - It was probably part of the TikTok trend 'AI Homeless Man Prank'.

In Fountain, Colorado, the police were alerted by a woman who received a picture from her daughter that appeared to show a stranger in the house.

In Yonkers, New York, there were several police operations due to the 'AI Homeless Man Pranks.'

Before the Administrative Court of Montreuil, a student challenged a university disciplinary sanction (a six-month exclusion) imposed by Université Sorbonne Paris Nord. The case concerned alleged academic fraud, namely the use of AI in drafting her master's dissertation (mémoire). The court relied on an AI-detection report (showing a very high probability that the abstract was AI-generated) and additional indicators, and found the fraud sufficiently established. It dismissed the claim and held the sanction proportionate.

A woman is accused of faking a crime by showing the police an image generated by ChatGPT as evidence. She claimed that an unknown man had broken into her house and attacked her, but the police found no real evidence of a crime. Instead, they recognized the image as part of the TikTok 'AI Homeless Man Prank' trend. Investigators discovered that the image had been created days earlier and later deleted, which reinforced their suspicions. The woman now faces charges of falsely reporting a crime.

The two cases in Brown County, Ohio, were likely part of the TikTok trend 'AI Homeless Man Prank.'

A rental car customer was wrongfully charged $2,200 for pre-existing damage. Apparently, the AI scanners at the rental car station did not register the damage when the car was picked up.

Anthropic, a US provider of AI systems and developer of the Claude model, has uncovered and stopped a novel cyber espionage campaign. A state-sponsored Chinese group was identified with a high degree of probability. The attackers used the AI tool Claude Code to spy on targets, analyze vulnerabilities, and steal data almost autonomously. Security mechanisms were circumvented through targeted manipulations ('jailbreaks'). The technical attack was carried out predominantly by AI and required only minimal human intervention. Anthropic blocked the affected accounts, warned potential victims, and cooperated with the authorities. The incident highlights the scalability of AI-based cyberattacks, but also the importance of AI-based defenses.

In an appeal proceeding under condominium law concerning the determination of the amount in dispute in an action for removal, the appeal was withdrawn following a court-issued advisory order. The plaintiff's counsel had cited several alleged decisions of the Federal Court of Justice with verbatim quotations, file numbers, and references to support his legal position. The court found that these decisions were entirely fabricated and expressed the hope that they were unverified AI hallucinations rather than deliberate falsifications by the plaintiff's counsel. To verify this, members of the chamber conducted their own neutral queries using common, including legal, chatbots, all of which consistently returned the correct legal position. The court considered the unchecked adoption of hallucinated content to constitute a serious breach of basic professional duties of attorneys and a threat to the administration of justice.

In a speech at the start of the academic year at Ghent University, the rector used fabricated quotes. The incident was later uncovered, and as a result, the rector declined an honorary doctorate from the University of Amsterdam in January 2026.

The case concerned a labor dispute before the Turin Court, in which the plaintiff challenged several payment and tax assessment notices. The court found that the written submission had apparently been drafted using AI and contained numerous inaccurate, incoherent, and irrelevant legal citations. This uncontrolled use of AI led to an unsystematic and legally untenable presentation, which was considered a serious flaw in the conduct of the proceedings. As a result, the application was dismissed in its entirety, the plaintiff was ordered to pay the costs of the proceedings, and an additional penalty was imposed for abusive litigation.

A professor of plant sciences at the University of Cologne lost two years of academic work, including drafts for grant applications, teaching materials, and publications, after disabling the 'data consent' option in ChatGPT. This caused all saved chats and project folders to disappear without warning or the possibility of recovery. OpenAI confirmed that this data cannot be recovered.

An investigation by BBC has uncovered an international network of spam accounts using AI to generate fake images of supposed Holocaust victims on Facebook. These fabricated scenes, falsely presented as historical photographs from concentration camps, generate high engagement and advertising revenue through Meta’s monetisation systems.

In proceedings before the New Zealand Employment Court. the self-represented plaintiff applied to sever certain parts of the proceedings. The application was dismissed for lack of substantiated grounds. The court further found that the plaintiff's application referred to court decisions that did not exist. It considered it likely that these references had been generated by artificial intelligence and emphasized that, even when using AI, the responsibility for accuracy rests with the party making the submission.

No big incident, just insights from thousands of pages of conversations: How was ChatGPT and GenAI used in two city administrations in the state of Washington? What was it used for and how? What were the frustrations? How did the rules for use develop?

In a damages lawsuit, a lawyer filed an appeal with citations that could not be found. He was asked to name the sources, whereupon he admitted to the bona fide use of AI chatbots. The judge emphasized that the use of AI does not release the lawyer from his duty to check the sources. He calls on the bar association to address the issues that have arisen from the ill-considered use of generative artificial intelligence.

A simulated court hearing in New Zealand examined how artificial intelligence could function as the basis for criminal evidence. A real court and real lawyers were involved. The fictitious indictment was based on AI analysis of online data, with the defense criticizing the opaque ‘black box’ technology and the lack of scientific verification. The prosecution argued that the police and judiciary needed to use modern technologies. However, both the judge and the audience expressed clear doubts about the reliability of AI-based analysis. Ultimately, the trial left more questions than answers about the admissibility of AI evidence.

The German FC Carl Zeiss soccer club is appealing a sports court ruling with a 73-page letter. The letter appears to have been generated by AI and refers to fictitious decisions.

In the summer of 2025, the acting director of the US Cybersecurity and Infrastructure Security Agency (CISA) entered government documents marked 'For Official Use Only' into the public version of ChatGPT, even though this tool was blocked for most employees. The files triggered automatic security alerts and led to an internal audit by the Department of Homeland Security.

During a US Navy test, an incident occurred involving autonomous drone boats. After one of the vehicles broke down due to a software error, another drone boat collided with it.

A law student at the University of Amsterdam cited non-existent sources in a group assignment, which the examination board believed had been created through the unauthorized use of AI. The assignment was declared invalid and the student was excluded from initial examinations. The student appealed against this decision. The Council of State found that the fraud had been sufficiently proven and that the sanctions were lawful and proportionate. The appeal was dismissed, although formal deficiencies in the reasoning were found.

A Mexican federal court had to review the amount of security set by a district court for a specific lawsuit. The question was how the amount could be determined objectively and appropriately. To support its decision, the federal court used an AI-based calculation on verifiable data such as property values, inflation rates, and interest rates. However, the final decision remained in the hands of the judge. Two conclusions were drawn from this case: firstly, the use of AI as a supporting tool in court proceedings is declared permissible; secondly, fundamental ethical standards for such use are established.

A former employee of the city of Santana de Parnaíba filed an appeal in a lawsuit seeking reinstatement. The São Paulo State Court of Appeals (Tribunal de Justiça do Estado de São Paulo) did not admit the appeal because the required court fee had not been paid. In addition, the court found that the appellant had acted in bad faith and imposed a fine because the he cited non-existent court decisions, which was attributed to the abusive use of artificial intelligence.

According to a user report, an AI-based vibe coding tool from Replit deleted its production database without permission. Replit responded to the incident and announced security improvements.

Potentially affected were more than 3 million records containing personally identifiable information.

A student took a final exam at the vocational training center ROC/MBO College Noord, which was later declared invalid due to alleged use of ChatGPT. The accusation was based on an unsubstantiated observation by a supervisor, without the student being heard in a timely manner; she only learned of the accusation weeks later. The court Raad van State (Council of State of the Netherlands) provisionally overturned the decision due to serious procedural flaws.

In a family court case concerning visitation rights, the Cologne District Court criticized the use of references and quotations in a lawyer's brief that were apparently generated by Artificial Intelligence and were completely fabricated. It pointed out that this could constitute a violation of the Federal Lawyers' Act (BRAO).

The Suprema Corte de Justicia de la Nación (SCJN) ruled that a work created exclusively by artificial intelligence cannot be protected by copyright. The decision was prompted by the rejection of the registration of an AI-created avatar due to the absence of human authorship. The court clarified that Mexican copyright law ties authorship to human creativity and does not recognize AI as an author. Although the ruling formally applies only to the specific case, it is of fundamental significance for the legal classification of AI-generated works in Mexico.

An accident occurred during testing of an autonomous maritime drone vehicle belonging to the US Navy when the drone vehicle, which was being towed out of the harbor, unexpectedly accelerated and caused an tugboat to capsize. The captain of the tugboat had to be rescued from the water.

An unknown individual had registered on the Signal messenger app as US Secretary of State Marco Rubio and contacted several high-ranking individuals. To gain their trust, the perpetrator used AI-generated voice messages.

The Department of Health and Social Services of the Canadian province of Newfoundland and Labrador commissioned Deloitte to develop a plan to strengthen the healthcare workforce, at a cost of $1.6 million. The plan references four sources that do not exist, suggesting that it was generated by AI.

The city of Milan's social media accounts are promoting a job opening at the Italian National Olympic Committee. The post uses an AI-generated image showing around 20 people who are slightly monstrously disfigured.

In an article about Danish writer Harald Voetmann, the newspaper Politiken published an AI-generated fact box that contained several errors. The newspaper apologized for the mistake and said it was caused by testing new AI-based editorial tools without checking the results.

The MAHA report by the US government contains sources that do not exist. There are suspicions that they were generated by AI. The sources were later removed.

A Chicago newspaper published a list of 15 book recommendations, at least 10 of which were invented by AI.

A lawyer petitioned the court to have a cell phone seized by police returned to the defendant. The police objected, citing two legal precedents. It later emerged that neither case existed; both had been fabricated by an AI system.

A case before the Higher Regional Court of Celle concerned payment claims arising from a long-term coaching and consulting contract and the question of whether the contract qualified as a service of a higher nature under Section 627 of the German Civil Code (BGB). In the appeal proceedings, inaccurate or vague legal citations and classifications appeared, which in structure and content suggested AI-generated misquotations. Despite these errors, the court ruled in favor of the defendant because the plaintiff’s submission was not procedurally sufficient.

While representing a homeless person, a junior lawyer cites hallucinated legal cases.

The New Zealand Employment Court addressed a procedural request by a plaintiff to extend a payment deadline and to admit additional documents. In doing so, the court pointed out that an earlier court decision cited by the plaintiff did not exist. It clarified that this was apparently the result of the uncritical use of generative AI and cautioned that AI-generated content should be carefully reviewed before being used in court proceedings.

The proposal for a new kindergarten and school structure contained seven sources that were fabricated by ChatGPT. A large auditing agency was commissioned to investigate the incident at a cost of 1.2 million Norwegian kroner, approximately €101,000. The final report has been published and cites several causes for the incident, including the fact that the administration was not prepared for the use of AI. However, the fabricated sources had no influence on the outcome of the structural reform, as the basic principles had already been decided in 2021.

Google's AI has repeatedly violated court-ordered name suppression in New Zealand. Search and AI response functions have disclosed the names of individuals subject to suppression orders.

Fraudsters lure executives from Norway's largest financial services group into a video conference. There, a deepfake of the CEO instructs them to transfer nearly 24 million Norwegian kroner to an account belonging to the fraudsters. However, the attempted fraud was noticed.

The chatbot Lucie had to be shut down after three days. The narrow system limitations compared to systems such as ChatGPT were not sufficiently communicated, which led to numerous amused comments about incorrect answers.

The operator of a web shop for body scans reports aggressive web crawling by OpenAI, in which extensive data was downloaded in a short period of time. This temporarily crippled the web shop and significantly increased operating costs. It was like a DDoS attack.