The Embodiment Gap: What Mandatory, Regulatory and Media-Based Incident Regimes Reveal About Robotic AI Failures, 2021–2026
This paper analyzes incident data from 2021 to 2026 across mandatory regulatory and media-based regimes to demonstrate that each system captures a distinct and incomplete subset of robotic AI failures, revealing significant gaps in visibility for non-crash operational errors and emerging humanoid technologies.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
When a machine with artificial intelligence makes a mistake in the real world, the story of that mistake depends entirely on who is listening. If a self-driving car crashes, a government agency with a legal mandate to collect data will record it. If a surgical robot malfunctions inside a patient, a health regulator will file a report. But if a robot stumbles in a museum or a humanoid machine behaves strangely at a trade show, the only record might be a video posted by a bystander or a news article. These different ways of collecting information are called incident regimes. They act as the eyes and ears of society, deciding which failures become part of our collective memory and which vanish into silence. The problem is that each regime sees a different slice of reality. One sees only crashes, another sees only medical errors, and a third sees only the stories that make the news. Because these systems do not talk to each other, we often think we understand how safe robots are, when in fact we are looking at a fragmented picture where the most dangerous failures might be hidden, and the most common ones might be invisible.
A researcher named Alfredo Merlet set out to map these blind spots by comparing three distinct ways of tracking robotic failures between 2021 and 2026. He looked at the mandatory crash reports filed with the US National Highway Traffic Safety Administration, the medical device reports filed with the US Food and Drug Administration, and a voluntary database that gathers stories from news media. By linking these sources together, he discovered that the way we learn about robot failures is deeply flawed. The mandatory system for self-driving cars recorded over 8,000 incidents, yet the media database, which is often treated as a comprehensive list of AI problems, contained only a tiny fraction of those crashes. In fact, less than one percent of the crashes reported to the government ever made it into the media database. This suggests that the public is seeing a highly distorted view of robotic safety, missing the vast majority of routine failures while focusing on a select few.
The study also revealed that the media does not simply ignore serious accidents. Contrary to the idea that news outlets prefer harmless, quirky robot stories over tragic ones, the data showed that severe crashes were actually much more likely to be reported. When a self-driving car crash resulted in a fatality or a serious injury, it was nearly nine times more likely to appear in the media database than a minor fender-bender. However, the media database captured a different kind of failure that the government system completely missed: operational failures without a crash. These are incidents where a robot behaves dangerously but does not hit anything, such as a robotaxi blocking an ambulance, a fleet of vehicles freezing in traffic because of a power outage, or a robot becoming stuck in a flood. Because the government's reporting rule is triggered only by a physical crash, these dangerous situations remain invisible to regulators, even though they pose a real risk to public safety.
The research extended to surgical robots, where a different kind of distortion was found. The database of medical device reports showed a massive, fourfold jump in malfunction reports in 2025. At first glance, this looked like a sudden spike in dangerous failures. However, a closer look at the data revealed that the surge was not caused by more injuries or deaths. Instead, the manufacturer began filing reports for a specific type of wear and tear on their instruments, a routine maintenance issue that had previously gone unrecorded. The number of actual injuries to patients barely changed. This finding highlights how the behavior of the person filing the report can create the illusion of a crisis, even when the physical reality of harm remains stable.
Perhaps the most striking finding concerns the newest generation of robots: humanoids. These are the machines that walk on two legs and are beginning to appear in public spaces, from amusement parks to city streets. The study found that these robots fall through the cracks of every reporting system. There is no mandatory rule for them to report crashes, no medical device category for their failures, and no consistent way for the media to track them. The researcher compiled a list of five specific incidents involving humanoids in 2026, including a robot that caught fire at a competition and another that was seized by police after an alleged inappropriate touch. Only one of these events appeared in the media database, and none were captured by any official government or health regulator. This creates a dangerous gap where the most visible and rapidly deploying form of artificial intelligence is operating without a systematic record of its failures.
The conclusion is that no single source of information can tell us the truth about robotic safety. The government sees the crashes but misses the near-misses. The media sees the dramatic failures but misses the routine ones. The medical regulators see the reports filed by manufacturers but may not see the actual harm. Until these systems are connected and expanded to cover the new types of robots entering our world, our understanding of risk will remain incomplete. The study suggests that regulators need to update their rules to capture operational failures, not just crashes, and to create new reporting channels for the humanoid robots that are already walking among us. Without these changes, we will continue to learn about the dangers of artificial intelligence only after the fact, and only when the story is dramatic enough to be told.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.