Assessment of AI-Generated HARA in Automotive Functional Safety
This study evaluates the capability of four general-purpose AI tools to support ISO 26262 functional safety engineering by comparing their Hazard Analysis and Risk Assessment (HARA) outputs, revealing that while AI offers a promising complement to safety processes, its implementation requires rigorous scrutiny regarding traceability, reliability, and regulatory adherence.
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
In the modern car, safety is no longer just about strong steel and airbags; it is about the invisible logic that keeps a vehicle from harming its passengers. Engineers rely on a rigorous process called hazard analysis to imagine every way a system could fail, from a brake that won't release to a sensor that sees a ghost. They then assign a risk level to each potential failure, deciding how strictly the system must be built to prevent injury. This process is the backbone of automotive safety, a discipline so serious that it has its own international rulebook. As artificial intelligence begins to write code and design systems for cars, a critical question has emerged: can a machine, trained on vast amounts of human text, perform this high-stakes thinking for us? Can an artificial intelligence look at a description of a car part and correctly identify what could go wrong, and then decide how dangerous that mistake would be?
Two researchers from a major automotive engineering firm set out to test this idea. They gathered four different artificial intelligence tools, ranging from assistants built into standard office software to specialized models designed for complex problem-solving. They gave each tool the exact same set of documents describing a specific car safety system: how the vehicle is supposed to work, the conditions under which it operates, and the scenarios in which it might be used. They asked the machines to act as safety experts and produce a hazard analysis, a document that lists every possible failure and rates its danger. To see if the instructions mattered, they tried two different ways of asking: one was a detailed, step-by-step conversation where the AI was told to adopt a specific professional role, and the other was a single, direct command to do the job.
The results showed that while the machines could produce the requested documents, the quality of their thinking was inconsistent. Every tool generated a list of safety goals, which are the high-level rules meant to prevent accidents, but the lists looked very different from one another. When the researchers compared the outputs, they found that only one specific safety goal appeared in every single list: the need to prevent the airbags from deploying when they should not. For everything else, the tools disagreed. One machine might identify a specific failure mode that another completely missed, or it might describe the same problem in a way that was too vague to be useful. This lack of agreement means that if an engineer relied on just one of these tools, they might miss a critical danger simply because that particular machine did not think to look for it.
The inconsistency became even more apparent when the tools tried to assign a risk rating to the problems they found. In the world of car safety, these ratings determine how much effort and testing a system requires. For the one safety goal that all four tools agreed on, their risk ratings still varied wildly. One tool rated the danger as moderate, while another rated it as the highest possible level of risk. This gap is significant because a difference of one level in the rating can change the entire development process, requiring more expensive components or stricter testing. The study also revealed that the way the question was asked changed the answer. When the researchers gave the tools a simple, one-sentence instruction, the machines tended to assign higher risk ratings than when they used a more detailed, conversational approach. This suggests that the AI is not calculating risk based on a fixed internal truth, but is reacting to the tone and structure of the prompt.
Perhaps the most telling finding was how the machines struggled to write the safety goals themselves. A proper safety goal must be written at a specific level of detail: it must state what bad thing to avoid without getting bogged down in the technical details of how to fix it. The researchers found that most of the AI-generated goals were written incorrectly. Many were too detailed, sounding like instructions for a specific part rather than a general rule for the whole system. Others were too vague, failing to clearly link the failure to the danger. In several cases, the tools confused the high-level safety goal with the lower-level technical requirements, mixing up the "what" with the "how." This indicates that while the machines can mimic the language of safety engineering, they do not yet fully understand the conceptual difference between a rule and a specification.
Despite these flaws, the study does not suggest that artificial intelligence is useless for safety work. The tools were able to generate a wide range of ideas and identify potential hazards that a human engineer might consider. The researchers concluded that these machines can serve as a helpful starting point, a way to brainstorm a first draft or to check if a human team has overlooked an obvious scenario. However, the output cannot be trusted on its own. Because the tools disagree on what the risks are, how dangerous they are, and how to write the rules to prevent them, a human expert must review every single line. The artificial intelligence can offer a spark of insight, but in the high-stakes world of keeping people safe in moving vehicles, the final judgment must remain firmly in human hands.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.