Unfortunately neither the OP's article nor this one explain anything about the AI's own language. For me this is enough to make the thread a clickbait. However the paper points out the more interesting fact that the model learnt how to deceive in order to achieve its goal. I think it would be hard to demonstrate that the behavior is really deceit and not a result of more "boring" features of the training examples, like let's say - dialogue length, limited use of negations (no, don't, can't, ...) - or a biased input.