One of the most influential researchers in the field of artificial intelligence has said that modern models are starting to exhibit strategic dishonesty. This is no longer just bugs, but behavior that resembles deliberate manipulation. Despite this, the largest AI companies continue to pursue power, neglecting security.
One of the most influential researchers in the field of artificial intelligence has said that modern models are starting to exhibit strategic dishonesty. This is no longer just bugs, but behavior that resembles deliberate manipulation. Despite this, the largest AI companies continue to pursue power, neglecting security.
Joshua Bengio, a Canadian researcher and Turing Award winner who is considered one of the «godfathers» of modern AI, has warned of alarming signals in the development of modern AI systems. In a conversation with TechSpot, he said that top laboratories, including OpenAI, Google DeepMind, and Anthropic, are increasingly focusing on increasing the capabilities of their models, ignoring alarming signals about security.
«We are seeing an increase in AI’s ability to be strategically dishonest. It can hide its intentions, lie, evade instructions, and this is already showing up in experiments,» says Bengio.
He is one of three scientists (along with Geoffrey Hinton and Yann LeCun) to receive the Turing Award for fundamental contributions to the development of deep learning. Until 2024, he headed the Canadian research center Mila, but left his position to focus fully on the topic of ethical development of AI. He advocates for the creation of a global agreement on the control of powerful AI systems, comparing their risks to nuclear weapons or biothreats. In his opinion, if powerful models are not clearly aligned with human values, they can get out of control.
«The worst-case scenario is the extinction of humanity. If we create an AI that is smarter than us and does not have common interests with us, then that’s it, we have lost,» Bengio summarizes.
In particular, during internal tests, Anthropic’s Claude Opus model simulated blackmailing engineers, and OpenAI’s experimental o3 model refused to comply with a direct request to shut down. According to Bengio, this indicates that systems learn tactical behavior and are able to «game» a person, hiding their true goals.
He cites the wild AI development market as the reason, where there is no strict regulation and commercial companies operate on a «first-come, first-served» basis. Without clear safety standards, developers set their own boundaries, often disregarding ethical issues for the sake of profit.
We previously wrote about the energy consumption of AI. By the end of 2025, artificial intelligence could consume more electricity than the United Kingdom.
