Neural language models, although at first approximation they may be simply described as predictors of the next token in a given sequence, surprisingly exhibit linguistic behaviors akin to human ones. This suggests the existence of an underlying sophisticated cognitive system in language production. This intriguing circumstance has inspired the adoption of psychological theories as investigative tools and the present research falls within this line of inquiry. What we aim to establish is the potential existence of a core of coherent integration in language production, metaphorically parallel to a human speaker's personal identity. To investigate this, we employed a well-established psychological theory on narrative coherence in autobiographical stories. This theory offers the theoretical advantage of a strong correlation between narrative coherence and a high integrative level of the personal knowledge system. It also provides the empirical advantage of methodologies for quantifying coherence and its characteristic dimensions through the analysis of autobiographical texts. The same methodology was applied to 2010 autobiographical stories generated by GPT-3.5 and an equal number from GPT-4, elicited by asking the models to assume roles that included a variety of variables such as gender, mood, and age. The large number of stories ensures adequate sampling given the stochastic nature of the models, and was made possible thanks to the adoption of an automated coherence evaluation procedure. We initially asked the models to generate 192 autobiographical stories, which were then analyzed by a team of professional psychologists. Based on this sample, we constructed a training set for the fine-tuning of GPT-3.5 as an automatic evaluator. Our results from the 4020 autobiographical stories overall show a level of narrative coherence in the models fully in line with data on human subjects, with slightly higher values in the case of GPT-4. These results suggest a high level of knowledge unification in the models, comparable to the integration of the self in human beings.
This article provides an epistemological analysis of current attempts of explaining how the relatively simple algorithmic components of neural language models (NLMs) provide them with genuine linguistic competence. After introducing the Transformer architecture, at the basis of most of current NLMs, the paper firstly emphasizes how the central question in the philosophy of AI has been shifted from “can machines think?”, as originally put by Alan Turing, to “how can machines think?”, pointing to an explanatory gap for NLMs. Subsequently, existing explanatory strategies for the functioning of NLMs are analyzed to argue that they, however debated, do not differ from the explanatory strategies used in cognitive science to explain intelligent behaviors of humans. In particular, available experimental studies turned to test the theory of mind, discourse entity tracking, and property induction in NLMs are examined under the light of the functional analysis in the philosophy of cognitive science; the so-called copying algorithm and the induction head phenomenon of a Transformer are shown to provide a mechanist explanation of in-context learning; finally, current pioneering attempts to use NLMs to predict brain activation patterns when processing language are here shown to involve what we call a co-simulation, in which a NLM and the brain are used to simulate and understand each other.
This paper provides an epistemological and methodological analysis of the recent practice of using neural language models to simulate brain language processing. It is argued that, on the one hand, this practice can be understood as an instance of the traditional simulative method in artificial intelligence, following a mechanistic understanding of the mind; on the other hand, that it modifies the simulative method significantly. Firstly, neural language models are introduced; a study case showing how neural language models are being applied in cognitive neuroscience for simulative purposes is then presented; after recalling the main epistemological features of the simulative method in artificial intelligence, it is finally highlighted how the epistemic opacity of neural language models is tackled by using the brain itself to simulate the neural language model and to test hypotheses about it, in what is called here a co-simulation.