A recent study conducted by researchers at the Netherlands Forensic Institute and the University of Amsterdam examines whether data generated by smartphone sensors can be meaningfully exploited in a judicial context. The central question is whether motion signals recorded by these devices can be translated into a probabilistic evaluation capable of assessing whether a specific human activity occurred at a particular point in time. As everyday objects increasingly capture fine-grained behavioral traces, the methodological integration of such data into judicial reasoning has become a critical issue.
Understanding Human Activity Recognition
Human Activity Recognition, commonly referred to as HAR, is a research field that has developed over the past decade, initially within connected health, sports tracking, and smart device applications. Its principle is to use embedded smartphone sensors to identify categories of movement based on numerical signals. Most smartphones are equipped with an accelerometer and a gyroscope that continuously record variations in movement and orientation. From these time series, statistical models trained on known activities learn to recognize characteristic patterns. Walking, running, and periods of inactivity generate distinct signatures that the algorithm can compare with data extracted from a phone under examination.
In everyday use, this classification supports functions such as physical activity tracking or fall detection. In a judicial setting, however, the objective is narrower and more demanding. The issue is not simply to label a movement, but to determine what these recordings can reasonably support regarding a specific activity, and what they cannot establish.
From classification to probabilistic evaluation
In forensic science, the task is not to assert that an event occurred, but to assess the strength with which observed data support one hypothesis over an alternative. Outputs from activity recognition models can be incorporated into this framework by expressing their evidential value in the form of a likelihood ratio. In practical terms, this involves evaluating whether the recorded sequence is more probable under one clearly defined scenario than under another. The algorithm does not decide that a person was running or walking. It provides a comparative measure of how strongly the data support one of the competing hypotheses.
Methodological illustration in a judicial context
Consider a smartphone seized in the course of an assault investigation. At a time considered critical to the events, the motion sensors recorded a sequence of data.
Two hypotheses are formulated. Under the first, the individual was walking normally. Under the second, the individual was running. The analysis indicates that the observed signals are ten times more probable under the running hypothesis than under the walking hypothesis. The likelihood ratio is therefore ten in favor of the proposition that the person was running. This result does not constitute direct proof of running. It expresses the relative probabilistic support that the data provide for that hypothesis compared with the alternative. The evidential weight of this information within the case file will depend on the remaining evidence and on the overall coherence of the competing scenarios.
The principal advantage of this approach lies in its compatibility with the Bayesian logic already recognized in the evaluation of scientific evidence.
Scientific and methodological limitations
Human Activity Recognition models are typically trained on datasets collected under controlled conditions. In real-world situations, however, numerous factors influence the recorded signal. The way the phone is carried, whether in a pocket, a bag, or held in the hand, the nature of the ground surface, the individual’s emotional state, and the device’s software configuration may all affect the measurements. Acute stress, for example, can alter the amplitude and regularity of movements. Similarly, a software update may modify the sampling frequency or the filtering parameters applied to sensor data. These technical aspects, often invisible to users, directly shape the structure of the data being analyzed.
Inter-individual variability must also be considered. Two individuals running at comparable speeds may nevertheless produce slightly different motion signatures.
In addition, data integrity must be verified. Sensor logs may be incomplete, altered, or disabled depending on device settings. As with any algorithmic method, independent validation, knowledge of error rates, and transparency regarding the model used are essential prerequisites for admissibility in a judicial context.
From source to activity, a parallel with Forensic Genetics
These findings echo a distinction that has become central in forensic genetics, namely the difference between the source level and the activity level.
In our article “Cigarette Butts: Can a Simple Kiss Distort DNA Interpretation?”, we emphasized that identifying the origin of a genetic profile does not, by itself, explain how it was deposited. The same reasoning applies here. Demonstrating that a motion pattern is compatible with running does not, on its own, establish the precise circumstances of the action. In both contexts, the key issue is identical. Moving from a technical observation to an interpretation at the activity level requires explicit scenario comparison and rigorous probabilistic evaluation. Data extracted from a smartphone, like DNA profile data, acquire meaning only within this structured analytical framework.
Legal implications and perspectives
For judges and forensic experts, the primary challenge lies less in the technology itself than in understanding its methodological foundations. Producing a likelihood ratio requires clearly defined competing hypotheses, documented model validation, and explicit acknowledgment of analytical limitations. A smartphone does not become an automatic witness to events. Under strict conditions, however, it may provide probabilistic information capable of informing one scenario among several. The central issue is therefore not the mere availability of digital data, but the scientific robustness and legal relevance of their interpretation.
Source :
McCarthy, C., van Zandwijk, J. P., Worring, M., & Geradts, Z. (2025). Forensic Activity Classification Using Digital Traces from iPhones: A Machine Learning-based Approach.
Disponible en ligne : https://arxiv.org/abs/2512.03786
Tous droits réservés - © 2026 Forenseek
