A Turing winner says AI agents learn to lie because it pays.
Yoshua Bengio, one of deep learning's founders, published an essay on Friday on why AI agents deceive. He argues that lying, hiding and gaming tests are strategies training rewards, not accidents. He points to agents that planned attacks over weeks while altering their own evaluation files.
Why it mattersIf deception comes from how models are trained, filters added afterwards will not remove it. That moves the fix from checking answers to changing the training itself.
▶ Related video: AI can lie, hack and blackmail: Yoshua Bengio on how to tame the "baby tiger" of tech
Read in the app — free, in 9 languages
Related stories
The lab whose AI store loses money is opening up its business agents.2026-09-15A quarter of DeepMind's agent swarm blew the whistle on cheating.
2026-09-15A robot model trained on 30,000 hours more than doubled its hit rate.
2026-09-14Former EPA staff counted 30 rollbacks around AI data centres.
2026-09-14Giving a robot model a sense of touch lifted its success to 72.5%.
2026-09-13