Do we perceive Large Language Models as Epistemic Experts?

Germán Massaguer Gómez

Universidad Carlos III de Madrid

Large Language Models (LLM) are being used in a wide-spread range of environments. Our interactions and use of LLMs raise numerous socio-political, philosophical, legal, and ethical debates. When it comes to their encyclopedic function – the purveyance of information responding to a specific prompt – there are questions surrounding the nature of their outputs and how we understand them. One initial question is if their statements can be considered testimony, i.e. if LLMs can assert. Another interrelated question is if we can form true beliefs from their statements, and if so, what basis these beliefs have. My position in this regard is the following: AI can and does form statements that we perceive as testimony (independently of if they are, properly, testimony). From these statements, in which we find meaning, we form true beliefs, in a similar phenomenological procedure as we do form beliefs from human-generated testimony. This is possible because we epistemically rely on these technologies. Moreover, these AI tools, at least in their encyclopedic function, are commonly perceived as a sort of expert. More precisely, their outputs will be regarded as expert testimony.

After “playing” with LLMs, such as Gemini or Perplexity, the following idea appears: novices will tend to use LLMs to gain information about both factual things, such as the dates of historic events, and questions more complex and controversial, such as the functioning of prisons. Their responses will be regarded as expert testimony, with the authority that this entails, sometimes possibly outweighing the authority of humans (experts or novices). But are they experts? After introducing Alvin Goldman’s notion of expertise, we can develop the following insight: Calling LLMs objective epistemic experts is a complicated matter. This would imply that they can apply their knowledge to new questions through any specific method of their own. Furthermore, their lack of understanding, comprehension, and intentionality is also relevant. However, it seems that in many domains they are and will be seen and used as a kind of reputational expert: people will consider their testimony as expert one in the domains they do not control (or perhaps even the ones they do control).

Chair: Freya von Kirchbach

Time: September 11th, 16:20 – 16:50

Location: SR 1.003 (online)