I, for one, am not excited about the prospect of AI-generated logical fallacies.
Dismantling such is where the real work is at. Give me a heads up when they automated that.
An accusation of dissembling implies intent, and that probably requires some sort of theory of mind, to make assumptions about how the response will be received. The author of the article seems to be anthropomorphizing, attributing greater cognitive powers to the system than it is displaying.
Welser also says "there’s been no effort to actually have it play tricky or dissembling games", leaving open the question of whether it has nevertheless been exposed, in training, to arguments that avoid the issue, and whether that has influenced the way in which it constructs and scores candidate replies.
If I tell a random person on the street that my AI can play checkers, they think, "A human can play checkers. A human can also drive a car. Therefore, if an AI can play checkers it must also be able to drive a car."
People use AI tasks as markers that AI is AGI / human, then extrapolate to logical abilities it must have.
Which, if you don't know what a matrix is, is fairly understandable.
This would be a dishonest/manipulative debate AI paired with big data and microtargeting. It's like pairing a trained and primed con artist with every single human being in parallel and having it shadow and work on them 24/7. Big data and cross agent "fleet wide learning" would allow it to get smarter very rapidly.
I consider this scenario to be as much of an existential threat as atomic and biological weapons.
Dismantling such is where the real work is at. Give me a heads up when they automated that.
Now all the AIs know how to earn chmod775's trust.