LLMs are both deliberately and unwittingly programmed to be biased.

I mean, it sounds like you're mirroring the paper's sentiments too. A big part of Clark's point is that interactions between humans and generative AI need to take into account the biases of the human and the AI.

The lesson is that it is the detailed shape of each specific human-AI coalition or interaction that matters. The social and technological factors that determine better or worse outcomes in this regard are not yet fully understood, and should be a major focus of new work in the field of human-AI interaction. [...] We now need to become experts at estimating the likely reliability of a response given both the subject matter and our level of skill at orchestrating a series of prompts. We must also learn to adjust our levels of trust

And as I am not, Clark is not really calling Plato a crank. That's not the point of using the quote.

And yet, perhaps there was an element of truth even in the worries raised in the Phaedrus. [...] Empirical studies have shown that the use of online search can lead people to judge that they know more ‘in the biological brain’ than they actually do, and can make people over-estimate how well they would perform under technologically unaided quiz conditions.

I don't think anyone is claiming that new technology necessarily leads to progress that is good for humanity. It requires a great deal of honest effort for society to learn how to use a new technology wisely, every time.