March 23, 2026
The Mozart Project is an interdisciplinary research project conducted jointly by autism specialists and computer scientists. It is led by Laurent Mottron, an internationally renowned psychiatrist and researcher in the Department of Psychiatry at the University of Montreal, in collaboration with Sylvie Ratté, professor and researcher in the Department of Software Engineering and Information Technology at ÉTS.
This project is based on a fundamental observation: human beings are naturally “wired” to learn languages. In neurotypical children, this learning occurs primarily through imitating parents and social interactions. But for many nonverbal autistic children, the path to language is different.
YouTube, a gateway to language
The work of Laurent Mottron and his students, David Gagnon and Alexia Ostrolenk, has highlighted a striking phenomenon: nonverbal autistic children spontaneously acquire elements of language from online videos, particularly on YouTube. In fact, researchers found that some children used words or expressions that did not come from the language spoken at home.
These children show a special attraction to cartoon-style videos, often focused on letters, numbers, vehicles, animals, or geometric shapes, and frequently accompanied by sounds and songs. Some progress gradually, but not consistently, toward more complex linguistic content: combinations of sounds, words, and sentences, followed by structured sequences.
This observation led to a key question: can we leverage these spontaneous interests to promote the emergence of spoken language, rather than letting YouTube's algorithm alone guide recommendations?
Focusing on children's interests
The main idea behind the Mozart project is to indirectly influence video recommendations in order to promote natural learning of the language spoken by parents. For example, if a child shows a keen interest in cars (their main focus of interest), could we offer them videos featuring cars, but gradually incorporating numbers, letters, or words?
Obviously, the goal is not to make children dependent on YouTube, but rather to harness their natural interest in digital content to gently guide them toward verbal communication.
The children taking part in the project each have an identical tablet, in a very restricted digital environment: only the YouTube app is accessible, via a premium family account with pre-filtered content. Children use the tablet for about an hour a day, and parents receive support in managing any reactions or crises that may arise.
The key role of information technology
This is where the expertise of Sylvie Ratté and her team, PhD students Roya Moeini and Alain Kiemde, comes into play. From a technical standpoint, the project faces several major challenges. First, YouTube does not provide direct access to its list of recommendations. This makes it impossible to modify it in the usual way.
Rather than recreating an environment similar to YouTube, which would have been a monumental task, the research team opted for a roundabout approach. Since the team manages the children's YouTube accounts, they chose to influence recommendations by watching specific types of targeted videos to “feed” the algorithm.
ÉTS Professor Sylvie Ratté
Understanding similarities between videos
Another significant challenge is the notion of similarities between videos, an especially complex concept in computer science. A video can be analyzed from several angles: sound, narrative, objects on screen, colors, characters, rhythm, etc. However, the children observed knew exactly what they liked and intentionally clicked on content that appealed to them.
The researchers are therefore working on collecting playlists automatically and analyzing them using artificial intelligence algorithms. Initial analyses are based on elements such as YouTube categories, titles, and hashtags. Multimodal data are gradually being added: audio transcripts, visual information (shapes, colors, objects), and sound characteristics.
At the same time, screen behavior is also analyzed: clicks, scrolling, fast forwarding, and browsing through the list of recommendations. For some children, the analysis focuses primarily on sound, as the child sometimes continues to explore the recommendations while listening to the video in the background.
Developing a tool for clinical teams and families
By cross-referencing the characteristics of the videos with the behaviors observed on screen, the research team aims to identify attention patterns specific to each child. Preliminary results show that the approach can already make it possible to effectively characterize the content of cartoons and reveal distinct preferences for certain visual, textual, or audio elements.
In the long term, the Mozart project aims to develop a monitoring tool for clinicians to better document progress in language learning, as well as a tool for parents to understand and track their child's interests, while strictly respecting data confidentiality.
This research could also influence content creation teams by helping them design videos that are better suited to the needs and interests of nonverbal children with autism.
In summary, the Mozart project explores a key question: is it possible to design a recommendation system that gently guides children toward language, rather than confining them to YouTube's algorithmic spiral? Ongoing research suggests that the answer may well be yes.
