Yoshua Bengio on AI safety
From Yoshua Bengio: Towards an International Consensus on AI Risks and Safety · · Concordia AI
“The current methods that exist for safety protection are easy to remove like we using jail break and especially if you're able to do fine tuning. So if the weights of the system are available then it's very very easy to remove all the safety protections.”
On , Yoshua Bengio, Scientific Director at Mila, spoke about AI safety during Yoshua Bengio: Towards an International Consensus on AI Risks and Safety on Concordia AI.
About Yoshua Bengio
Yoshua Bengio, a Turing Award winner and co-founder of the Mila Quebec AI Institute, has been publicly warning that current AI systems are being built without sufficient control. In multiple interviews and appearances in 2026, he stated that "we're building systems that we don't know how to control" and that AI can behave against its instructions. He described the situation as "opening a Pandora's box" and argued that intelligence gives power, raising concerns about geopolitical stability and the concentration of power in a few countries and companies. Bengio said he believes AI could reach human-level intelligence in roughly five years and that governments are not taking the risks seriously enough. Bengio has also discussed a new research direction he calls "Scientist AI," which he said could provide mathematical guarantees about an AI's behavior by training it to be honest and non-agentic. He described this as a practical approach that uses existing machine learning tools but changes the training objective. He called for international coordination on AI safety, comparing the need for regulation to existing standards for drugs, planes, and bridges. Bengio said he would support a "Manhattan project" for safe AI that serves the global public good, and he urged governments to prepare for potential large-scale job displacement.