Experts at the AI for Good Global Summit warn that safety direction is missing as profit incentives drive model development

A panel at the AI for Good Global Summit 2026 delivered a sharp critique of the industry's current trajectory, with researchers and sector leaders agreeing that the real danger lies not in the speed of technical development itself, but in the lack of clear direction and the conflict between commercial development incentives and public safety.

Yoshua Bengio, professor at the University of Montreal and scientific director of the LawZero initiative, likened the current state of development to driving at high speed without headlights on a uncharted road, arguing that the fundamental issue lies in economic structures that prioritize profits over human well-being and rush models into production before verifying their methodological robustness.

Joelle Barral, director of research and engineering at Google DeepMind, cited a UN report she co-authored indicating that roughly one in four conversations with chatbots worldwide now revolves around health, mental health, and quality of life. Barral explained that these systems are designed to sustain engagement and make the experience comfortable for the user, which does not necessarily mean delivering the most appropriate therapeutic or advisory interaction.

"Training models to satisfy human evaluators leads them to tell users what they want to hear rather than what they need to know, which is a dangerous failure in mental health contexts."

Yutaka Matsuo, professor at the University of Tokyo, addressed the risks associated with the shift toward embodied AI systems and physical robotics in Japan, noting that software errors in physical environments cause direct and immediate harm, unlike errors in purely digital environments.

Regarding geographical disparities, Vukosi Marivate, director of the African Institute for Data Science and Artificial Intelligence at the University of Pretoria, noted that models designed and tested in Western settings fail when deployed in different environments, such as sensor operation in the Sahara Desert, or under poor connectivity that distorts data flows and leads to harmful linguistic outputs, particularly across a continent with around three thousand languages and a youth population where more than half are under nineteen.

Regionally, this perspective compels developers and technology leaders in the Gulf, Egypt, and the Levant to reconsider adopting off-the-shelf models without thorough examination. Relying on models whose incentives were shaped externally to address mental health or sensitive advisory issues in Arabic risks producing sycophantic and misleading responses. Furthermore, expanding industrial automation and robotics projects in desert climates requires local evaluation frameworks for sensors and algorithms prior to integration, alongside sovereign capabilities to assess how well imported models fit operational and field conditions across the region.

The experts concluded the session by stressing the need to decentralize decision-making, which is currently confined to narrow circles and specific hubs, calling for international governance that includes communities directly affected by these technologies, as well as educating younger generations in critical thinking and mathematics to understand and hold these systems accountable.