What Does It Mean To Generate Information

The aim of this paper is to define generativity, 
i.e. to perform a philosophical analysis of conditions required 
for a given process to be judged as satisfactorily generating information. 
The aim of this paper is to define generativity, 
i.e. to perform a philosophical analysis of conditions required 
for a given process to be judged as satisfactorily generating information. 
Surprisingly, this notion has been virtually untouched in the philosophical literature, 
contrary to a near-omnipresence of ‘generative-talk’ in the public discourse inspired 
by the dawn of the allegedly generative AI. In order to fill this gap, I start by 
reviewing various purportedly generative processes in order to delineate what I call 
faux-generativity from genuine generativity, aiming to fix the explananda. 
Then, I establish whether existing solutions might be used to explain away this problem. 
I discuss the usage of ‘generative’ label in neural networks and cognitive science 
(‘generative models’). Specifically, I look into the interpolation/extrapolation 
distinction offered there as a promising candidate for prying apart real generation 
from faux one. I argue that it is not sufficient for providing a satisfying definition 
that accounts for intuitions about genuine information-generation. 
Instead, I revert to discussing these intuitions in more detail, utilizing cases 
borrowed from computational photography to navigate the complex jungle of instances. 
On this basis, I offer a new cluster definition of genuine generating of information, 
that is based on content-preserving inferences that utilize information actively stored in compressed internal representations, that I dub ‘the real-generation account’. This account is inspired by inferentialism in philosophy of perception, Ramsey’s account of representation and Boghossian’s ‘Taking’ condition and work in machine learning on decompressions from internal latent spaces. I conclude with discussing the importance of setting clear boundaries for genuine generativity - for both academic and pragmatic purposes. For the former, I embed this problem within the current debates on ‘stochastic parrots’ and ‘blurry JPGs’ surrounding LLMs, but also within longstanding debates in philosophy about the origins of new knowledge and in psychology about the ontology of mental capacities. As for the latter, I briefly discuss the necessity of clearly diverging generative from non-generative processes for questions about the trustworthiness and veracity of computational artifacts.