Raison d’exprimer: why should machines want to talk to us?
The problem of defining Artificial Intelligence has always clashed with the difficulties we have in understanding and defining what Intelligence itself is. As a matter of fact, for legislative purposes it is preferred to define what an AI system is. The problem has been debated for centuries, influencing how research has dealt with the problem of simulating the phenomenon in machines. While the hype around AI is high, at this time, given the success of Large Language Models (LLMs), we are still very far from reproducing intelligence, whatever that is. We have, however, created some useful and powerful tools that learn to perform a good amount of tasks without any need to understand them. While there is still a long way to go before we can explore the problem of providing machines with raison d’être, these tools provide a good set of new capabilities that enable us to start investigating the problem of raison d’exprimer, or the capability for machines to use language as a tool to influence a communicative context, rather than as a product in itself.
In the previous post of this series, a linguistic perspective of communication highlighted the importance of using language to do things, or to produce changes in the world. Representing the illocutionary force of the utterance is, therefore, of utmost importance to build a machine that talks because it is trying to accomplish something that goes beyond the mere task of text completion. In particular, modeling illocutionary force attempts to give these an actual reason to communicate, exploring the fundamental aspect of intentionality in communication. From a computational point of view, Large Language Models are trained to react to linguistic stimuli with the most probable sequence of tokens that follows the input. They model, in linguistic terms, the locutionary aspect of communication. Acting, however, is also concerned with illocutionary aspects, or the reasons why linguistic content has been produced, which cannot be represented by a model of language alone.
LLMs, therefore, model surface aspects of human communication: as per their definition, they model language but do not necessarily model communication. In this sense, LLMs speak without saying anything and leverage on human attribution of meaning to linguistic stimuli, producing an illusion very similar to the one described by Braitenberg vehicles (Braitenberg, 1986).
This aspect surfaces in specific kinds of hallucinations LLMs take, like citing inexistent sources when explaining something: they capture the surface aspect of specific kinds of text of presenting supporting data (scientific papers, legal trials, etc…) but do not model the need for these sources to exist nor the reason why they are used. While specific techniques like, for example, Retrieval Augmented Generation, are developed to mitigate similar problems that keep emerging, we live through a rush at making LLMs bigger and bigger. This approach, however, risks to produce a series of slightly more convincing illusions, at each iteration, that may never match actual communication, possibly resulting in disaster, as in Aesop’s fable “The frogs and the Ox”.
In our view, following directions also given by Gary Marcus (Marcus, 2020), an alternative course of action consists in integrating the Natural Language Generation capabilities LLMs provide with capabilities provided by other kinds of models. In particular, models dedicated to decision making and coming from symbolic AI, possibly reinterpreted using modern technology, may be integrated in an architecture designed to match a theory of human communication. Such an architecture would have the fundamental advantage of exhibiting the explainability characteristics of older approaches to AI with the flexibility needed to handle the physical world possessed by machine learning.
A fundamental set of computational tools to implement this approach is found in the context of Probabilistic Graphical Models (PGMs), which have been explicitly developed to support decision making and advanced reasoning capabilities that escape the associative modelling capabilities of machine learning. The relevant work by Judea Pearl in this sense, summarised in (Pearl & MacKenzie, 2018), provides a strong theoretical support in this sense. The first note of interest, from a theoretical point of view, is the role of the concept of doing in the description of the Ladder of causality. Notably, the definition of the second level of the ladder is based on intervention capabilities matching the illocutionary force driving Austin’s theory of speech acts, described in the previous post.
The first cognitive ability, seeing or observation, is the detection of regularities in our environment, and it is shared by many animals as well as early humans before the Cognitive Revolution.
The second ability, doing, stands for predicting the effect(s) of deliberate alterations of the environment, and choosing among these alterations to produce a desired outcome.
Since, in this framework, the associative capabilities of machine learning are confined to the first layer of the ladder, other approaches, working together with machine learning, are needed to develop actual communication capabilities. In the rest of this series, we will discuss the linguistic foundations of our approach for the development of a new computational model of communication and the tools we use to implement it.
References
(Braitenberg, 1986) Braitenberg, Valentino. Vehicles: Experiments in synthetic psychology. MIT press, 1986.
(Marcus, 2020) Marcus, Gary. “The next decade in AI: four steps towards robust artificial intelligence.” arXiv preprint arXiv:2002.06177 (2020).
(Pearl & Mackenzie, 2018) Pearl, Judea, and Dana Mackenzie. The book of why: the new science of cause and effect. Basic books, 2018.
