No Humans! Autonomous Development and Execution of a Forensic 3D Workflow by a General-Purpose AI Agent
This study demonstrates that a general-purpose AI agent can autonomously develop and execute a complex forensic 3D stature-reconstruction workflow from raw data to a final report, achieving partial accuracy but highlighting that technical completion does not guarantee scientifically correct results without rigorous human validation.
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
Forensic science often relies on the careful reconstruction of events from physical traces, a process where experts piece together how a scene looked and what happened within it. One particularly challenging task involves determining a person's height from a single photograph or a few frames of video, often taken by security cameras that are not designed for precise measurement. To solve this, experts traditionally combine the image with a three-dimensional map of the room, usually created by a laser scanner, and use specialized software to align a digital human model with the person in the picture. This requires deep knowledge of geometry, anatomy, and the specific quirks of camera lenses. The goal is to find a digital pose that matches the photo while remaining physically possible for a human body, allowing the expert to measure the height from the floor to the top of the head. For decades, this has been a job that demands a human expert to guide every step, from calibrating the camera to adjusting the digital skeleton.
A recent experiment asked a bold question: could a general-purpose artificial intelligence agent, given only a description of the rules and the raw data, figure out how to build this entire process from scratch and run it without human help? Researchers at the Zurich Forensic Science Institute set out to test this by handing a sophisticated AI system a complex forensic task: reconstruct the height of a person from four different images taken in a controlled room. The AI was not given any pre-written code, no ready-made 3D models, and no step-by-step instructions on how to use the software. Instead, it was provided with a laser scan of the room, a reference photo with markers, the four images of a test subject, and a document describing the laboratory's standard quality rules. The AI had to write its own computer code, create a digital human body, figure out the camera's position, and then try to fit that body to the person in the photos, all while generating a full report of its work.
The result was a demonstration of a new kind of autonomy. The AI successfully developed a complete workflow, writing the necessary software to process the laser scan and the images. It built a digital human model with movable joints, calculated the camera's position and lens distortions, and then attempted to match the model to the person in each of the four photos. Throughout the process, the only human intervention required was a simple prompt to continue when the AI ran out of its allotted computing time. The system produced a detailed report with measurements and visual evidence, effectively acting as a forensic technician who could not only follow instructions but also invent the tools needed to do the job.
However, the experiment also revealed the limits of this current technology. While the AI managed to complete the entire technical pipeline, the accuracy of its final measurements was mixed. The test subject's actual height, measured with shoes on, was 175 centimeters. The AI produced four separate height estimates from the four images: 176.65 cm, 175.96 cm, 161.30 cm, and 169.27 cm. Two of these estimates were close enough to the true height to meet the laboratory's standard target, but the average of all four was 170.8 cm, which fell short of the required accuracy. More importantly, the digital body the AI created did not stay consistent. The length of the legs, the width of the shoulders, and the size of the torso changed significantly from one image to the next, even though they were all supposed to represent the same person. In some cases, the AI seemed to adjust the body's proportions to make it fit the photo, rather than finding the correct anatomical position.
This suggests that while the AI can handle the heavy lifting of coding and running complex calculations, it still struggles with the subtle, intuitive judgments that human experts make. A human examiner knows that a person's leg length does not change just because they are standing in a different pose, but the AI sometimes altered these dimensions to solve the puzzle of the image. The researchers found that the AI could not always resolve the ambiguity of depth and posture from a flat picture, leading to a digital model that looked correct from the camera's view but was anatomically unstable when viewed from other angles. The system passed all its internal technical checks, yet it failed to produce a consistently accurate reconstruction.
The study concludes that we are on the verge of a shift in how forensic work might be done. General-purpose AI agents can now take a high-level description of a forensic problem and turn it into a working, executable workflow without needing a human to write the code or operate the software. This capability suggests that in the near future, much of the technical implementation and execution of forensic analysis could be handled by machines. However, the human expert will likely remain essential, not to perform the calculations, but to define the problem, set the boundaries, and, crucially, to validate the results. The AI can build the bridge, but it cannot yet be trusted to walk across it without a human checking the footing. The experiment shows that technical completion does not guarantee a correct answer, and that for now, the final responsibility for interpreting the evidence must stay with a human.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.