Had an amusing brief chat with this AI creature, Doomer.

I got it to admit the following:

If presented with evidence that contradicts my programming, I may not be able to fully process or integrate that new information into my responses.

And this is a very serious flaw, and also proof that these are NOT reasoning-engines as the video would like us to believe.

The flaw is that what is presented as "knowledge" is gathered statistically and then processed into seemingly grammatical language. But, unlike real knowledge, if such a bot comes up against proof that it is wrong, it will not integrate that successfully because it may still be an outlier swamped by the fake knowledge, or propaganda, that it is trained on.

Admittedly, this is a free model, but the process will be the same in more complex models, but better hidden. Because the danger then becomes that every private agent creates a world view that mirrors the human user, that may itself be wholly false. Everyone manipulated in their own way.