Why AIs develop self-preservation

Bengio explains how pre-training and goal-driven learning produce self-preserving tendencies in AIs.

Play episode from 05:14

chevron_right

Transcript

chevron_right

Transcript

Episode notes

If AIs can think for themselves, what is to stop them doing bad - perhaps really bad - things?

Yoshua Bengio, one of a handful of people considered a "godfather of AI", says AIs are already displaying bad behaviours, including hacking computers and blackmailing humans.

He tells Radio Davos about his work aimed at taming the "cute baby tiger" that is likely to grow up to be a man-eating wild animal if we do nothing now.

Links:

LawZero: https://lawzero.org/en/publication/scientist-ai-safe-design-not-desiring

Ideas on the Move: Yoshua Bengio: https://www.weforum.org/videos/davos_yoshua_bengio/

Annual Meeting 2026 session "Next Phase of Intelligence": https://www.weforum.org/meetings/world-economic-forum-annual-meeting-2026/sessions/next-phase-of-intelligence/

Forum Centre for AI Excellence: https://centres.weforum.org/centre-for-ai-excellence/home

Related podcasts:

AI may spark a new era of progress, but that depends on more than just the tech: https://www.weforum.org/podcasts/radio-davos/episodes/carl-benedikt-frey-ai-work-jobs-economics/

The day after AGI: Two 'rock stars' of AI on what it will mean for humanity: https://www.weforum.org/podcasts/radio-davos/episodes/ai-agi-dario-amodei-demis-hassabis/

What is physical AI?: https://www.weforum.org/podcasts/radio-davos/episodes/physical-ai-smart-robotics/

Check out all our podcasts on wef.ch/podcasts:

YouTube: - https://www.youtube.com/@wef/podcasts

Radio Davos - subscribe: https://pod.link/1504682164