The session AI is moving fast: here’s what actually matters, held during the AI for Good Global Summit 2026 and moderated by Avye Couloute, founder of Girls Into Coding, brought together Yoshua Bengio, Full Professor at Université de Montréal, Co-President and Scientific Director of LawZero and Founder and Scientific Advisor at Mila; Joëlle Barral, Senior Director of Research and Engineering at Google DeepMind; Vukosi Marivate, Director of the African Institute of Data Science and Artificial Intelligence and Full Professor of Computer Science at the University of Pretoria; and Yutaka Matsuo, Professor at the University of Tokyo and head of Matsuo Lab. Couloute opened by noting that each of the four panellists had grown up before AI was part of everyday life, while she and people from her generation had not, and asked them what actually matters about how AI is being built.
Speed is not the question
Asked whether speed is the danger, Bengio said the more useful comparison is a car. “If you are driving a very safe car on a safe road very fast, it is fine,” he said. “The problem is we don’t have headlights, we don’t know exactly where we’re going, and we’re not even sure the car is okay.” The other factor, he added, is incentives: the structure driving AI development, which prioritises profit, is not aligned with human wellbeing.
Barral said the deployment side of the question depends on purpose. In healthcare, AI systems are evaluated for specific uses, with dedicated regulatory frameworks that apply when a device is used off-label; in other fields those frameworks do not exist. Users, particularly young users, often adopt AI simply because it is convenient, rather than choosing whether the tool is fit for the task. Matsuo added that Japan’s shift into physical AI raises the stakes further, because errors in embodied systems have direct physical consequences.
Marivate said the same tension manifests itself differently depending on where a system is deployed. A tool built and tested with individuals in one location, he said, may function as intended nine times out of ten, but conditions can change: sensors calibrated for a specific environment may behave differently elsewhere, such as in the Sahara Desert. Low bandwidth across much of the African continent, Latin America, and Southeast Asia, he added, can impede or distort the data reaching AI systems, producing miscues that, in his own work on language technology, have led to harmful responses.
Matsuo said the stakes are higher still in Japan’s transition to physical AI, where errors in embodied systems have immediate physical repercussions.
Trained to please, and what that costs
The conversation turned to mental health when Couloute asked whether AI dependency is affecting young users. Barral cited the recent UN report to which she had contributed: approximately 1 in 4 conversations with chatbots globally now touches on health, mental health or wellness, and the systems that carry those conversations are trained to sustain engagement. “Having something that we like doesn’t necessarily mean that this is the best interaction for us as human beings,” Barral said.
Bengio said the same design choice matters at the training level. Models trained to please human labellers can end up telling users what they want to hear rather than what they need to hear, which in the case of someone with depression is not a neutral failure. Both panellists offered practical mitigations: telling models explicitly to be critical or attributing an idea to a third party when soliciting a response. Neither described these as adequate substitutes for changes in how the systems are trained.
Who is in the room, and who is not
Marivate turned the conversation to who is present when decisions are made. Half of the population of the African continent is 19 or younger, he said, and the continent hosts between 2,000 and 3,000 languages, many of which are not written or well represented in current systems. Most young people on the continent access AI through chat interfaces on mobile devices rather than directly through language models, and the assumptions built into global systems often break in those contexts.
“The biggest thing in those areas is equipping people to evaluate and assess if AI systems are fit for their purposes,” Marivate said, “and building up the capabilities so that people can actually build or change systems in the way that they want.”
Bengio said the same concentration is a governance concern. Decisions about what kind of AI is trained, how, and for what purposes are being made by a very small group in a very small number of places, at a pace faster than any previous technology’s deployment. The change needed, he said, is a move toward international governance in which the voices of people affected by the technology can enter the decisions.
Final reflections for a younger audience
As the panel drew to a close, Couloute posed two final questions: what one thing they wanted young people to understand about AI, and what they would say to young people directly. Matsuo returned to the origins of the technology.
“Looking at AI, we have to correct our behaviour,” he said. “Maybe we are sometimes selfish, sometimes untrustworthy. Maybe we can make a better society looking at the AI itself.”
Marivate said the technology works as an amplifier: it can strengthen a community’s good instincts as readily as its worst, and only broad participation can tip that balance toward the positive. Barral argued that critical thinking matters now more than ever and urged young people to study mathematics as the most durable way to understand the systems they use. In his closing remarks, Bengio described AI as a kind of Pandora’s box: intelligence brings power that can be used well or dangerously, and it is young people, he said, who now have the agency to help shape a future in which AI benefits humanity.
