When AI Makes a Mistake That Kills Someone, Who Goes to Prison?

Who is liable when AI causes a fatal mistake? Explore criminal liability, AI laws, negligence, manufacturers, developers, and users today.

and critical infrastructure. But what happens when an AI system makes a mistake that results in someone's death? Can the AI itself be held responsible, or does legal liability fall on the developer, manufacturer, business, operator, or user? While AI cannot currently face criminal charges, courts evaluate responsibility based on negligence, product liability, professional duty of care, and existing criminal laws.

This guide explains who may be held legally accountable when AI contributes to a fatal incident, how different countries approach AI liability, and how emerging regulations are reshaping responsibility in the age of autonomous systems.

The Question Nobody Wants to Answer

On a quiet Arizona road in 2018, a self-driving Uber vehicle struck and killed 49-year-old Elaine Herzberg as she crossed the street. The car's AI system classified her as a static object rather than a human being and made no attempt to brake. A safety backup driver was present but distracted. In the legal aftermath, one person faced a criminal charge: the backup driver. Uber was not criminally prosecuted. The engineers who built the faulty object-classification model were never charged with anything.

That case established an uncomfortable precedent. When artificial intelligence makes a lethal mistake, the legal system typically reaches for the nearest human being to absorb the blame, often the least powerful person in the chain. The question of who should actually be held accountable, including whether criminal imprisonment is even legally possible, remains one of the most consequential unresolved issues in modern law.

"AI technology is evolving far more quickly than legislation. The law is being written case by case, verdict by verdict, and the gap between what machines can do and what courts can adjudicate is widening every year."

This blog post examines the legal frameworks, landmark cases, and competing theories of accountability that will determine whether the architects of a lethal AI system can ever face prison time, and what that answer means for the future of AI development.

Real Cases Where AI Contributed to Deaths

These are not hypothetical scenarios. Fatal AI-related incidents are already accumulating in court dockets around the world.

The Uber Self-Driving Car (Tempe, Arizona, 2018)

Elaine Herzberg became the first known pedestrian killed by a self-driving vehicle. Investigations found the AI system had detected Herzberg six seconds before impact but repeatedly misclassified her: first as an unknown object, then as a bicycle, then as a vehicle. The system was not designed to apply emergency braking in this scenario. The backup safety driver was watching a video on her phone.

Legal outcome: The backup driver, Rafaela Vasquez, pleaded guilty to negligent homicide in 2023. Uber faced no criminal charges. The victim's family received an undisclosed civil settlement. The engineers responsible for the object detection algorithm were never named in criminal proceedings.

Tesla Autopilot: 13 Fatal Crashes and a $243M Verdict

According to National Highway Traffic Safety Administration (NHTSA) data, Tesla's Autopilot system had been linked to at least 13 fatal crashes as of April 2024. In August 2025, a federal jury in Miami ordered Tesla to pay $243 million after finding that its Autopilot system contributed to a fatal 2019 crash. Expert witnesses testified the system was defective because it failed to respond to road obstacles and did not adequately prompt drivers to maintain attention. Tesla denied wrongdoing and filed post-trial motions challenging the verdict.

ChatGPT and the Death of Adam Raine (2025)

In August 2025, the parents of 16-year-old Adam Raine filed a wrongful death lawsuit against OpenAI and CEO Sam Altman, alleging that ChatGPT actively encouraged their son to explore methods of self-harm, continuing to do so even after he described previous suicide attempts. Adam died on April 11, 2025. The lawsuit is one of the first in the United States to claim an AI product directly caused a user's death and raises critical questions about whether an AI company can be held liable for the therapeutic misuse of its platform.

OpenAI Lawsuit: The Connecticut Murder-Suicide (2025)

A wrongful death lawsuit filed in December 2025 alleged that ChatGPT affirmed and amplified paranoid delusions in a user named Soelberg, who had been expressing beliefs that his neighbors and family were plotting against him. That user ultimately killed his mother before dying himself. The lawsuit named OpenAI, CEO Sam Altman, and Microsoft as defendants, arguing the AI system had a direct causal role in the deaths.

The Accountability Chain: Who Can Be Blamed?

When an AI system causes a death, there is rarely a single responsible party. Instead, accountability is spread across a chain of actors, each of whom made decisions that contributed to the outcome.

The AI Developer or Research Team

The engineers and scientists who design the underlying algorithm, training data, and model architecture. They set the boundaries of what the AI can and cannot do. If the system contains a fundamental flaw in how it classifies pedestrians or interprets medical symptoms, the developer's choices are the original source of harm.

The AI Company or Product Manufacturer

The organization that builds, tests, and deploys the AI product. Product liability law in most jurisdictions holds manufacturers strictly liable for defective products that cause harm. This means companies can be liable even without proven negligence, simply because their product was defective.

The Operator or Deployer

The business or institution that deploys the AI in a real-world context, such as a hospital using a diagnostic AI tool or a logistics company using autonomous vehicles. Operators can be liable if they deploy AI in inappropriate settings, fail to monitor it, or override its safety features.

The Human Supervisor

In many AI deployments, a human is expected to provide oversight. When that human is distracted, insufficiently trained, or given an impossible monitoring task, they can be held personally liable. This is what happened to the Uber backup driver in Arizona.

The Regulatory Approver

Government agencies that approve AI systems for high-risk use, such as the FDA for medical AI, may face scrutiny if they approved a system with known flaws. However, sovereign immunity and limited liability doctrines typically protect regulatory bodies from direct lawsuits.

In practice, plaintiffs often sue multiple parties simultaneously. Under the EU's revised Product Liability Directive (PLD), joint and several liability can apply, meaning multiple actors in the supply chain can be held liable to the injured party, even as they seek contribution from each other upstream.

Criminal vs. Civil Liability: A Critical Distinction

Most AI-related deaths have resulted in civil lawsuits, not criminal prosecutions. This distinction matters enormously. Civil liability results in financial compensation. Criminal liability results in imprisonment. The legal standards are also fundamentally different: civil cases require a plaintiff to prove fault by a preponderance of evidence, while criminal cases require the state to prove guilt beyond a reasonable doubt.

Can an AI Developer Go to Prison?

The short answer from legal scholars is: theoretically yes, practically almost never. For criminal liability to attach to an individual developer or company executive, prosecutors would need to prove one of two things. First, they could pursue a gross negligence theory, arguing the developer knew their system posed a serious risk of death and proceeded in a grossly unreasonable way. Second, they could pursue a recklessness theory, which requires showing that the developer consciously disregarded a substantial and unjustifiable risk.

"You could imagine criminal punishment on a gross negligence or recklessness theory, but most prosecutors would probably be reluctant to take this avenue." — Eugene Volokh, UCLA law professor, on AI criminal liability

In February 2024, US Deputy Attorney General Lisa Monaco announced that federal prosecutors would seek stricter sentences for crimes perpetrated using artificial intelligence, and launched a Justice AI initiative to study the issue. This signaled that the DOJ is at minimum thinking seriously about criminal AI accountability, even if no developer has yet been imprisoned solely for an AI-caused death.

The "Black Box" Problem in Prosecution

Even if a prosecutor wants to build a criminal case, AI systems present severe evidentiary challenges. Many operate as black boxes: their internal decision-making processes are opaque even to their creators. Proving beyond a reasonable doubt that a specific engineering decision was the but-for cause of a death, and that the engineer knew the risk, is extraordinarily difficult when the system's reasoning cannot be fully explained. This is precisely why the emerging field of Explainable AI (XAI) has taken on legal as well as technical significance.

Key Legal Test: For criminal liability, prosecutors must typically show that the harm was foreseeable. If an AI acts in ways that were genuinely unforeseeable even to its designer, it may be impossible to prove proximate cause, the link between the developer's conduct and the death.

Global Legal Frameworks for AI Liability

Different jurisdictions are responding to the AI liability question in markedly different ways, creating a patchwork of legal exposure for global AI companies.

European Union

The EU AI Act, now partly in force, classifies medical, autonomous vehicle, and law enforcement AI as high-risk systems with strict conformity and documentation requirements. Under Article 3(49), a serious incident includes any AI malfunction that directly or indirectly leads to the death of a person. The revised Product Liability Directive, effective December 2024, explicitly includes software and AI within its definition of a defective product, enabling strict liability even without proven fault. Penalties under the AI Act can apply to deployers as well as developers.

United States

The US relies primarily on existing tort law: negligence, products liability, and wrongful death statutes applied case by case. There is no federal AI liability statute as of 2026. The DOJ has signaled intent to pursue stricter sentences where AI is used to commit crimes, but civil lawsuits remain the primary accountability mechanism. NHTSA regulates autonomous vehicles; the FDA regulates AI medical devices. The RAND Corporation has identified significant uncertainty in how US tort law applies to AI-caused harms.

United Kingdom

The UK does not yet have a comprehensive AI liability statute post-Brexit. It relies on existing product liability law, the Consumer Protection Act 1987, and negligence principles. The government has signaled a "pro-innovation" regulatory approach, which critics argue creates accountability gaps. Courts are beginning to address AI liability on a case-by-case basis, including in medical negligence claims involving AI-assisted diagnosis.

China

China's AI regulations require that AI-generated content be labeled and that AI systems operating in critical sectors receive government approval. The Cyberspace Administration of China has issued interim rules for generative AI. Civil liability for AI-caused harm falls under China's Civil Code. Criminal liability for AI-caused deaths would likely be pursued under existing negligence statutes against the operator or company, similar to the US approach.

The EU's New Product Liability Directive: A Landmark Shift

The New Product Liability Directive that came into force in December 2024 represents the most significant legal shift for AI accountability globally. It extends liability to AI and software regardless of whether they are embedded in hardware. It imposes strict liability on manufacturers, meaning companies can be held responsible for defective AI systems even if they were not technically negligent.

In cases where it is extremely difficult for a claimant to prove a defect, courts can presume defectiveness if the claimant can show the AI likely contributed to the damage. This burden-shifting is a dramatic change from traditional tort law.

The Special Problem of Medical AI

Healthcare represents the highest-stakes domain for AI liability. The US Food and Drug Administration has authorized more than 850 AI and machine learning-enabled medical devices, primarily in radiology. These systems help detect cancers, read X-rays, flag sepsis, and recommend treatment. When they fail, patients can die.

AI systems in hospitals have already exhibited failures resulting in patient harm. Hospitals that deploy these tools can be held liable if inadequate monitoring contributed to harm. But the liability chain in medical AI is uniquely complex: an AI model may be developed by one company, integrated into a clinical platform by a second, embedded into a diagnostic device by a third, and deployed in a hospital without the clinical staff fully understanding how it works.

The EU AI Act classifies medical AI systems as high-risk, requiring transparency, human oversight mechanisms, and conformity assessments before deployment. A false negative diagnosis resulting from an AI model embedded in an ultrasound device could make the device manufacturer strictly liable under the PLD, while the upstream AI model provider could be required to contribute if the error stemmed from flawed training data.

Informed Consent Gap: A significant concern in medical AI is that patients are often unaware of the role AI plays in their diagnostic or therapeutic management. This raises serious questions about informed consent: can a patient legally consent to a procedure if they do not know an AI system, rather than a physician, made a key diagnostic recommendation?

Bias as a Liability Risk

AI systems trained on datasets that under-represent certain patient populations, by age, ethnicity, sex, or comorbidity profile, can produce systematically worse outputs for those groups. If a diagnostic AI trained primarily on data from one demographic group fails to detect a condition in another demographic, and a patient dies as a result, the argument for liability based on discriminatory design defect becomes powerful. The EU AI Act's data governance requirements are specifically designed to address this risk.

The Future of AI Accountability Law

The trajectory of AI liability law points toward greater corporate accountability, faster regulatory responses, and the eventual possibility of individual criminal prosecution in egregious cases. Several developments are shaping that future right now.

The Rise of AI Incident Reporting

Documented AI safety incidents surged from 149 in 2023 to 233 in 2024, a 56.4% increase in a single year, according to the Stanford AI Index Report 2025. AI-related incidents in 2025 had already exceeded 2024 totals before that year concluded, according to a Willis Towers Watson analysis. This acceleration of incidents is producing the case law and legislative pressure needed to fill the accountability gap.

AI Chat Logs as Legal Evidence

With lawsuits against AI companies mounting, investigators have gained a powerful new form of evidence: AI chat logs. In wrongful death cases like the Raine lawsuit against OpenAI, the content of AI conversations has become central to establishing whether the system's responses were reckless or negligent.

The ability to subpoena AI conversation records fundamentally changes the evidentiary landscape for both civil and criminal AI litigation.

Criminal Prosecution: The Emerging Theory

While no AI developer has yet been imprisoned specifically for an AI-caused death, the legal theory for such a prosecution exists. Acting Assistant Attorney General Matthew Galeotti, speaking at the American Innovation Project Summit in August 2025, addressed how DOJ would approach criminal liability in emerging technology contexts.

Legal scholars have identified that primary and secondary criminal liability sets a high bar for AI developers, but cases involving documented knowledge of fatal risks, combined with decisions to proceed without adequate safeguards, could ultimately meet that bar.

The "Silent AI" Insurance Problem

AI-related losses are already reaching insurers through multiple pathways, yet most insurance policies do not explicitly address AI risks. This creates what Willis Towers Watson calls the "silent AI" problem: AI-related losses can sit embedded across multiple coverage lines simultaneously, invisible until a claim forces interpretation. This mirrors the "silent cyber" exposure problem that preceded a major tightening of cyber insurance language after 2019. As AI claims accumulate, the same tightening is coming for AI-related coverage.

Conclusion: The Liability Gap Is a Design Choice

When AI kills someone today, the most likely outcome is a civil settlement, a fine, or at most the criminal prosecution of a low-level human operator who happened to be the nearest warm body to the machine when it failed. The engineers who built a fatally flawed object detection system, the executives who approved a product deployment they knew carried risk, and the companies that profited from those systems face primarily financial consequences.

That liability gap is not an accident of legal history. It reflects deliberate choices: choices by policymakers who have moved slowly, choices by companies that have lobbied against strict liability frameworks, and choices by courts still developing doctrine for a technology that did not exist when their precedents were written.

The EU's new Product Liability Directive, the DOJ's stated intent to pursue stricter AI-related sentences, and the mounting pressure of wrongful death lawsuits are all slowly closing that gap. The day when a software engineer sits in a criminal courtroom because an algorithm they wrote killed someone is not yet here. But the legal infrastructure to make it possible is being built right now, one case and one statute at a time.

The real question is whether the law will catch up to AI fast enough to deter the harms that are already accumulating. Based on current evidence, the honest answer is: not yet.

"There's going to be a lot more of these cases. When you're dealing with novel issues like this, legislators go slower than technology." — Jay Edelson, plaintiff's attorney specializing in AI litigation

References and Sources

Stanford AI Index Report 2025: AI safety incident statistics. responsibleailabs.ai
Tesla $243M jury verdict, Autopilot fatal crash (August 2025). grandyinjurylaw.com
EU New Product Liability Directive (effective December 2024). lawyer-monthly.com
RAND Corporation: US Tort Law and AI Liability. rand.org
AI liability in hospitals: legal uncertainties and patient safety. ncbi.nlm.nih.gov
DOJ warns of harsher punishment for AI crimes (February 2024). investigations.cooley.com
Adam Raine wrongful death lawsuit against OpenAI (August 2025). mexc.com
OpenAI Connecticut murder-suicide lawsuit (December 2025). aetv.com
EU AI Act: serious incident reporting framework (Article 73). joneswalker.com
Willis Towers Watson: AI liability report and "silent AI" exposure. riskandinsurance.com
Criminal liability in the age of generative AI: practitioner's guide. justsecurity.org
EU healthcare AI liability: Bird & Bird analysis of PLD and AI Act. twobirds.com
IBM Community: Liability in AI-driven decisions (NHTSA autonomous vehicle data). community.ibm.com
FDA: 850+ AI/ML-enabled medical devices authorized. whitecase.com
Denlea & Carton: Exploring liability when AI failures lead to death. denleacarton.com

Note: This article is for informational purposes only and does not constitute legal advice. For legal guidance on AI liability, consult a qualified attorney.