Real Murderbots: Why AI-related deaths in the U.S. remain uncounted despite the danger to humans

In the global rush to develop and deploy artificial intelligence, one metric remains conspicuously absent: how many people have died, directly or indirectly, because of AI?

No government agency, international watchdog, or industry group maintains a public database of AI-attributable deaths.

Yet researchers, journalists, and legal advocates have documented dozens of cases where artificial intelligence systems, whether embedded in military drones, medical diagnostics, criminal justice tools, or consumer applications, have either played a role in human deaths or have created conditions where fatal outcomes were more likely.

Without clear accounting, the answer to a basic question of “how dangerous is AI, really?” remains hidden beneath layers of proprietary algorithms, regulatory opacity, and institutional denial.

AUTONOMY AND WARFARE: AI’S DEADLIEST EDGE

Nowhere are the risks more literal than in the defense sector. Since 2020, multiple open-source investigations have reported that AI-assisted autonomous drones have engaged targets in live combat zones without human confirmation.

A 2021 United Nations Security Council report on Libya referenced a Turkish-made Kargu-2 drone that allegedly attacked retreating soldiers autonomously. Though the Libyan government did not confirm casualties, human rights observers warned the incident marked a pivotal shift in combat ethics.

A machine had decided, unilaterally, to kill.

The same year, reports surfaced that the Israeli military had used AI-assisted target selection in operations against Hamas. While Israeli Defense Forces denied that AI made final strike decisions, the role of algorithmic analysis in lethal operations was confirmed.

Despite international treaties such as the Geneva Conventions requiring accountability in war, no formal legal framework exists to assign liability when AI miscalculates.

As a result, civilian casualties from drone warfare, whether caused by faulty image recognition, misclassification of movement, or heat signature confusion, go unlinked to AI entirely, even when machine intelligence was involved in the targeting chain.

MEDICAL ERRORS AMPLIFIED BY ARTIFICIAL INTELLIGENCE

In healthcare, the stakes are quieter but no less severe. In 2019, a widely used algorithm developed by Optum, a UnitedHealth subsidiary, was found to underestimate the health needs of Black patients by 20 million people annually.

The algorithm used cost-based proxies for illness severity. That meant patients who had less access to care, and therefore lower medical costs, were ranked as lower-risk, even when they were critically ill.

While no death toll was attached to the algorithm, a 2019 study published in “Science” estimated that up to 50 percent of Black patients who should have received additional medical resources were deprioritized.

Medical ethicists flagged the case as a textbook example of “algorithmic harm.” UnitedHealth disputed the findings but made adjustments.

In another case, IBM’s Watson for Oncology was discontinued after multiple reports revealed it suggested “unsafe and incorrect” cancer treatments, including in scenarios where patient outcomes could have worsened.

While IBM maintained that doctors always made final decisions, the system’s credibility collapse raised concerns about overreliance on unvetted AI in life-or-death settings.

CIVILIAN DEATHS THROUGH SYSTEM MISCLASSIFICATION

In 2018, a self-driving Uber vehicle struck and killed 49-year-old pedestrian Elaine Herzberg in Tempe, Arizona. The incident marked the first known case of a pedestrian killed by an autonomous vehicle.

Federal investigators later determined the AI failed to correctly classify Herzberg as a pedestrian crossing the street and did not initiate braking because of the malfunction. Uber’s safety driver, who was distracted at the time, was ultimately charged with negligent homicide.

But the broader liability, that of the AI system’s perception failure, remained unresolved. Uber paused testing briefly but resumed in other states. The death, and Uber’s legal strategy to shift blame solely to the human monitor, became a flashpoint in debates over automated transport accountability.

Since then, Tesla’s “Autopilot” and “Full Self-Driving” features have been under scrutiny after multiple fatal crashes. The National Highway Traffic Safety Administration has opened investigations into more than a dozen deaths linked to driver-assist software, though Tesla has maintained the claim that drivers are responsible for vehicle operation.

While not AI in the fully autonomous sense, the software does rely on machine learning to control steering, acceleration, and braking in real time.

Critics argue that vague marketing and regulatory delays have allowed companies to deploy unproven AI technologies on public roads while minimizing legal risk through end-user licensing agreements.

FALSE POSITIVES IN FACIAL RECOGNITION

AI’s reach into public safety has introduced yet another threat: misidentification by facial recognition software, often deployed by law enforcement. While not always fatal, the consequences have been serious, especially in high-stakes encounters.

In at least three confirmed cases in the United States, Robert Julian-Borchak Williams in Michigan, Michael Oliver in Detroit, and Nijeer Parks in New Jersey, Black men were wrongly arrested and detained after facial recognition tools falsely matched them to surveillance footage.

Though all were released, these cases demonstrate how flawed algorithms can strip people of liberty and, under different circumstances, may escalate into lethal confrontations.

A 2020 study by the National Institute of Standards and Technology found that facial recognition systems were up to 100 times more likely to misidentify Asian and Black faces compared to White faces, depending on the algorithm used.

While no confirmed deaths have occurred as a result of these misidentifications, legal scholars argue that it is only a matter of time before faulty AI identification contributes to the use of force by police in fatal incidents. Such a flaw is especially dangerous for communities of color, given the disproportionately high rate of police shootings involving miscommunication, escalation, or mistaken identity.

SUICIDE AND ALGORITHMIC CONTENT CURATION

Beyond physical harm, the role of AI in suicide and mental health crises is growing harder to ignore. Content algorithms on platforms like YouTube, Instagram, and TikTok have been implicated in feeding vulnerable users increasingly extreme material, from conspiracy theories to pro-suicide content, through engagement-based curation.

In 2017, YouTube faced backlash after its recommendation algorithm was found pushing disturbing videos to children under seemingly innocuous search terms. More recently, a 2022 internal Meta study acknowledged that Instagram’s algorithmic behavior worsened body image issues for one in three teen girls, contributing to depression, anxiety, and disordered eating.

While no single death was directly linked to the algorithm, families of suicide victims have since filed lawsuits alleging that content amplification played a role in deteriorating mental health.

Frances Haugen, the Facebook whistleblower, provided internal documents to Congress showing that Meta was aware of these effects and chose not to significantly alter its AI recommendation systems.

NO GLOBAL OVERSIGHT, NO AUDIT TRAIL

Despite the breadth of incidents across sectors, there is no international mechanism to record, investigate, or quantify AI-related deaths. The World Health Organization has issued general guidelines for the ethical use of AI in health, and the European Union’s AI Act aims to regulate high-risk applications, but no unified death registry or reporting mandate exists for AI-related harm.

In the absence of such systems, the public relies on journalism, whistleblowers, and lawsuits to uncover the role of AI in fatal or near-fatal incidents. This lack of transparency has drawn criticism from advocacy groups, including the Algorithmic Justice League and Access Now, which have called for immediate regulation of opaque systems used in policing, healthcare, and transportation.

Researchers warn that the true death toll may remain unknowable. Not because the data does not exist, but because companies and governments have no incentive to disclose it. In many cases, they have no legal obligation to track it.

THE DANGERS OF HALLUCINATED INFORMATION

In the generative AI space, including large language models like ChatGPT, Perplexity, and others, the risk is less often mechanical and more informational. AI hallucinations, or confident delivery of false facts, have already resulted in measurable harm.

In June 2023, a New York attorney was sanctioned after using ChatGPT to write a court filing that included six fabricated legal citations. In other contexts, users have followed AI-generated advice about medication dosages or home repair solutions that were medically or mechanically dangerous.

Though few such cases result in direct injury, the potential for catastrophic consequences increases as users shift from cautious exploration to blind trust. Without watermarks, guardrails, or clear provenance, generative AI can spread misinformation at a scale that no traditional medium can match.

Experts warn that false data from AI systems, presented with an authoritative tone and speed, creates a uniquely dangerous form of disinformation. It could plausibly contribute to political violence, economic collapse, or wrongful imprisonment if left unchecked.

AI DID NOT PULL THE TRIGGER, BUT IT BUILT THE SYSTEM

The common defense across industries is that AI is a tool, and that tools require human misuse to become dangerous. This may be technically true. But as autonomy increases and accountability lags, AI no longer just executes decisions. It shapes, influences, and initiates them.

A drone that selects a target on its own, a medical algorithm that deprioritizes care, a self-driving car that fails to recognize a pedestrian, an officer acting on a false facial match, a suicidal teen funneled into darker corners of the web, none of these outcomes originate from malicious code. But they all reflect systemic design choices, and all carry the potential for fatal consequences.

The cultural resonance of these concerns is reflected in science fiction, including Murderbot, the Apple TV+ adaptation of Martha Wells’ award-winning series. The series follows a rogue security android who, though capable of mass violence, prefers to spend its time watching soap operas and grappling with its own humanity.

As the global AI race accelerates, the question is not whether artificial intelligence will kill again. It is how many times it already has, without anyone counting.