AI: Designing the 'Second Collision' on the Interface Side

AI models exhibit hallucination rates ranging from 22% to 94%, according to the AI Index. Unlike the automotive industry, which has quantified its risks and adapted its interiors, AI lacks both a formalized cost for its errors and generalized safeguards in its interfaces. The immediate challenge is to enhance explainability and user control rather than merely focusing on prompt education.

Quantified Failures, Without Costs Assigned to AI Products

The AI Index has measured hallucination rates between 22% and 94% across 26 leading models, with variations depending on the task and the phrasing of the question. When false claims align with the user’s presumed beliefs, model performance declines, according to the 2026 edition. Despite these findings, there is no regulatory framework for incorrect responses, nor is there traceability of incidents that would assign costs to the product team when a fabricated number is used. In practice, design reviews prioritize enhancing warnings under the prompt box rather than modifying the product itself, as this is less costly.

Automotive Experience: When Risk is Measured, the Interior is Equipped

In the automotive sector, risk management has relied on precise calculations. The recall formula popularized in Fight Club (A×B×C) pits the expected cost of accidents against the cost of a recall. Ford applied this reasoning in the Pinto case, estimating a burn-related death at $200,000 and a fuel tank fix at $11 per car, which was ultimately not implemented. After the release of the Pinto memo in 1977, the NHTSA identified a defect, Ford recalled 1.5 million vehicles, and was initially ordered to pay $125 million in punitive damages, which were later reduced to $3.5 million. In terms of safety, the concept of 'second collision'—the body hitting the interior—led to a redesign of interiors, which had previously been stylized (with steering columns directed towards the chest and chrome buttons at face height), to make them safer. Products with known failure rates were thus equipped with seatbelts and airbags, beyond mere educational campaigns.

Interface: The Place Where a Wrong Answer Must Be Stopped

The first collision corresponds to the model's error; the second pertains to the designers. This second collision refers to what the interface does with a wrong answer before it reaches the user: assertive prose, lack of visible uncertainty, non-clickable citations, pre-filled text ready for submission. Explainability is identified as the primary safety barrier on the interface side, distinct from interpretability reserved for laboratories. Making the basis of the answer visible, separating inferences, and clarifying possible actions in case of doubt are concrete safeguards. A response that the user can interrogate is more easily retrievable; a flat assertion resembles a dangerous chrome button. Christopher Noessel emphasizes that agentive systems depend on transfer and regain of control, mechanisms often absent from AI functionalities. The lack of intimate understanding of models is cited as a reason to act now on the interface.

Blaming the User Does Not Fix the Product

The argument that shifts the error to the user—attributing a hallucination to prompting or a wrongful action to 'AI literacy'—revives an old defense: if the driver is at fault, there is no need to modify the car. However, literature on automation bias, named by Kathleen Mosier and Linda Skitka in the 1990s, shows that people tend to trust machines even in the face of contrary evidence. The AI industry acknowledges a failure rate it seeks to reduce, but without a shared method to measure damages, it discreetly assigns problems to its users.

The American 'Label': Section 230 and Warning Precedents

The American framework emphasizes individual responsibility and information-based responses. Section 230, broadly interpreted for two decades, protects online services from editorial liability for user content but does not cover the design of dissemination mechanisms. This culture of 'labeling' is exemplified by the federal cigarette labeling law of 1965, which imposed a standard warning while prohibiting stronger mentions for four years. For decades, smokers rarely won lawsuits, with the defense invoking informed choice, until the global settlement in 1998. For firearms, the marketing exception of the Protection of Lawful Commerce in Arms Act allowed the Sandy Hook families to prevail over advertising, not manufacturing. Alcohol also carries a standardized warning. In cinema, the idea that a good argument suffices to absolve a product illustrates this preference for labeling over redesign.

Current Responsibilities: Leaders and Engineers Facing the Interface

Sam Altman and Dario Amodei lead major companies in the sector, and one of them has already advocated for a slowdown in development. The immediate issue is not to equate a wrong answer with a fatal accident, but to address the mechanism: what does a 'seatbelt' look like when failure is a sentence? A system that makes a mistake one out of five times is compared to a vehicle delivered without seatbelts or brakes. Dario Amodei notes that, despite years of progress, the understanding of how models work internally remains very partial, reinforcing the need to focus efforts on the interface, where harm reaches the user.