Nikhil Mahant: AI generated outputs: beyond the binary

9.1.2025 15:00

Meeting room (zasedačka) of the Institute of Philosophy, CAS, Jilská 1, Prague

Can AI outputs represent objects of the world? Can AI systems perform speech acts like asserting, declaring, and apologizing? Can AI speech be deceptive in the same way that human speech can be? In short, do AI generated outputs have contents?

The popular and academic interest in these questions is immense. Yet the discourse is dominated by an unfortunate binary. At one end are the Blake Lemoines of the world, who maintain that there are no interesting (content-level) differences between outputs produced by (say) a LLM and a human. At the other end are those who take AI systems to be ‘stochastic parrots’: machines that spit out strings of words or collection of pixels that do not have contents.

In this talk, I will provide a third alternative and provide some arguments in its favour. The alternative is that whether/which linguistic outputs are contentful depends on the cognitive demands that an output -- e.g., representation, speech act, act of deception -- makes on the entity producing the output.

The upshot is that there may not be a single, simple answer to the questions posed in the beginning of this abstract. Some generated outputs may represent the world (e.g., pictures, maps), others may not (e.g., caricatures, which require representing in a certain manner i.e., humorously). Some speech acts may be performed by AI systems (e.g., assertions), but not others (e.g., apologies). So goes for deceptive acts. It all depends on the cognitive requirements imposed by the act in question.