Yoshua Bengio said it was important to distinguish between what scientists actually know about AI risks and what can only be ‘plausibly extrapolated’ from that evidence