Dashcam footage may soon carry legal weight beyond the insurance claim.
A research team has introduced a new task — responsibility distribution estimation — that asks multimodal large language models to divide fault, in percentages, among every party in a traffic accident. Rather than labeling crashes with a single cause or culprit, the system outputs something like "60% driver, 40% other vehicle." The team built an annotation pipeline assisted by an LLM, then fine-tuned models on ego-view video under several input conditions: raw frames, segmentation-overlaid frames, and plain text descriptions of the scene.
Why ego-view footage matters is the key insight here. Fixed infrastructure cameras and satellite imagery are expensive to deploy and can only show what a camera saw — not what the driver saw. A dashcam recording is, by definition, the driver's visual field in the seconds before impact, which is exactly the evidence relevant to questions of avoidability and reasonable reaction. That framing shifts accident analysis from pure reconstruction toward something closer to a legal standard of care.
The paper reports strong initial benchmarks, though "strong" at a first-cut task with a custom dataset is a claim worth watching. The harder test will come when insurers or courts consider whether a model's fault split holds up to adversarial scrutiny — and who gets to audit the training data that shaped it.