Muvraline launched an Artificial Intelligence to assist the police in legal matters and identification. This Artificial Intelligence could allow better identification of criminals avoiding bias in the construction of lineups. Our goal is to base ourselves on a research process to assess the quality of its approach by comparing the performance of people facing AI.

Linups can be seen as an essential step in an investigation process. It consists in presenting a suspect among other individuals who are generally called distractors.

However, in France, few people are trained to present suspects in an alignment of people, while this method produces the most miscarriages of justice (National Research Council and others, 2015). These errors may be the result of incorrect methods which could bias the outcome of the lineup and, by extension, the outcome of the investigation. These biases are often linked for example to the absence of people available to construct a credible lineup (no individuals of the same age, the same sex or not sharing the same physical characteristics, sometimes going as far as individuals very different skin colors), or to instructions that can influence the choice of the witness (the police officer insists more on one individual than on the others). The suspect is then very salient in this lineup and anyone who has read the description will be able to identify it without having witnessed the situation. The construction of an unbiased lineup is not easy and must be done according to recommendations in order to avoid bias (Wells, Seelau, Rydell, & Luus, 1994). In addition to these elements, the construction of the lineup should be done blind, the person who chooses the distractors is based on the description of the suspect during his previous testimony and therefore has no knowledge of the suspect. In the same logic, the person in charge of presenting the lineup to the witness should neither have knowledge of the suspect nor of the construction of the lineup so as not to influence the choice of the witness. In short, it is easy to create a lineup that can lead to the designation of the suspect, whether consciously or not.

In order to overcome these difficulties, we are currently building an Artificial Intelligence (AI) whose final function would be to build unbiased automatic lineups, automatically selecting photographs of credible descriptors for the display. However, building an AI capable of identifying the elements of a description in a photograph requires knowing how people are described. If we know what are the most mentioned elements. On the other hand, there is no research publication informing us about the content of these descriptions, thus not making it possible to estimate the most common descriptors and their number. Otherwise, there is no certainty that two people give similar descriptions of the same individual since each of them may have a different mental representation of the descriptors. Thus, ambiguous descriptors such as “blond” or “light brown” hair are highly likely to appear in different descriptions of the same individual.

We therefore undertook a first experiment in which we asked French-speaking participants to describe portrait photographs, so as to collect adjectives covering the different variations of each characteristic of the face (hair, eyes, nose, mouth, etc.). Based on the frequency of adjectives we have retained common descriptors. These descriptors were later used to describe a large number of portraits in order to create a database. During this stage, each portrait was observed at least 5 times in order to be able to observe the disparities in the selected distractors. In this way, the AI ​​will then be able to take into account the disparities in possible descriptions among individuals.

The AI ​​trained with this data should then be able to describe new photos by itself and build a lineup with credible descriptors. The first results of this AI seem promising, but it remains to be tested if it produces fair lineups (Tredoux, 1999; Wells & Bradfield, 1999).

Leïla Sadok, Nicolas Rochat


National Research Council and others. (2015). Identifying the culprit: Assessing eyewitness identification. National Academies Press.

Tredoux, C. (1999). Statistical considerations when determining measures of lineup size and lineup bias. Applied Cognitive Psychology, 13(S1), S9–S26.;2-1

Wells, G. L., & Bradfield, A. L. (1999). Measuring the goodness of lineups: parameter estimation, question effects, and limits to the mock witness paradigm. Applied Cognitive Psychology, 13(S1), S27–S39.;2-d

Wells, G. L., Seelau, E. P., Rydell, S. M., & Luus, C. A. (1994). Recommendations for properly conducted lineup identification tasks.