Large language models (LLMs), such as ChatGPT, seem to exhibit human-level intelligence on a wide range of cognitive tasks. This raises the following questions: is such behavior enough to determine the kind of intelligence current AI systems possess? Is this seemingly intelligent behavior enough to attribute to AI the same kind of intelligence we commonly attribute to humans? Pioneer studies (Bubeck et al., 2023) suggest that systems such as ChatGPT start to come close to the generality observed in human intelligence. Such approach has been characterized as using a methodology of “black-box” interpretability. However, “black-box” methodologies are commonly criticized because they ignore the inner functioning of a model, by just analyzing the inputs and outputs of a given system, and because they assume, arguably without good reasons, that the same kinds of tasks that measure human intelligence might also shed some light into LLMs’ intelligence. The argument, inspired by Dretske (1993) and recently used by Grzankoswki (2024), about how to find traces of “real intelligence” in such systems and why “black-box” approaches are potentially misleading has the following form:

1) Intelligence depends on a) mental representations of the correct kind (semantic kind) and b) those same relevant mental representations have to be “used” by the system in order to cause behavior.
2) Black-box approaches to LLMs fail to account for a) and b).
3) Then, we have no reasons to think that there is intelligence in LLMs (at least according to black-box approaches).

The main goal of my paper is to argue against premise 1) by criticizing both a) and b) as conditions for intelligence attribution. I remain slightly neutral about 2) and 3).

As a first step, I problematize the concept of mental representation itself as a scientific or natural kind. Consequently, I also show that there are fruitful candidates for definitions of intelligence that do not appeal to mental representations at all. More specifically, in the same spirit as arguments proposed by Ramsey (2017), I show how there can be a useful “divorce” between cognitive science and mental representation, specifically when it comes to intelligence. This objection also draws on recent objections to mental representations as scientific kinds (Facchin, 2023) or as natural kinds (Allen, 2017). I show that, in the relevant sciences, different definitions of intelligence have been proposed which are compatible with a more deflationary approach to mental representation. This small survey includes popular characterizations in psychology (Sternberg, 2019; Gardner, 2011), neuroscience (Haier, 2023; Duncan, 2020) and AI itself (Russell, 2016; Brooks, 1995). Provisionally, the argument in this first section has the following form:

1) Intelligence requires mental representations.
2) Mental representation is a scientifically problematic concept.
3) Then, intelligence is a scientifically problematic concept. (But actual scientific views do not seem to require mental representations in any case.)

As a second step, I use the distinction between vehicle and content to problematize the causal efficacy of mental content, and argue, borrowing from Egan (2020), that contents and their causal/explanatory role are pragmatically but not metaphysically constrained. Additionally, I argue that, even assuming that representations in LLMs have some structural resemblance to their targets, which makes them good candidates of non-arbitrary content determination, they still suffer of potential causal inefficacy (Nirshberg, 2023) and, therefore, should be considered as mere epiphenomena (Baysan, 2021). Following this line of thought, I argue that even scientific methods like probing LLM representations rely on pragmatic attributions, as it has been argued in neuroscience regarding the use of probes for decoding mental representations (Cao, 2022). Such a caveat might allow us to think of mental representation as an important explanatory concept in AI’s behavior, without engaging with the metaphysical problems that the concept commonly implies, particularly regarding its causal efficacy. Consequently, a useful concept of intelligence should drop the causal requirement of contents without fully trivializing the role of representations in behavioral explanations. This second argument has roughly the following structure:

1) Intelligence requires mental contents to be causal explanations of behavior.
2) Mental contents are not causal explanations of behavior.
3) Then, intelligence is an incoherent concept. (Not even applicable to humans.)

Finally, I present some speculative considerations about a more deflationary concept of intelligence that may have the potential advantage of being scientifically productive and avoid severe metaphysical problems. The general conclusion of my paper is two-sided: it has a negative and a positive conclusion. On the negative side, I claim that, at least until we have a more robust account of mental representation in science and philosophy, we should think of intelligence without necessarily requiring mental representation or its causal powers. On a more positive side, we can evaluate intelligence in terms of more operational criteria, for example, behavioral success, mechanistic complexity and learning conditions, allowing us to take seriously the intelligence that we observe in generative systems. Such a conclusion does not imply that generative systems are categorically intelligent or even more intelligent than humans just because they behave as such, at least with respect to certain cognitive tasks. The mere fact that their mechanistic complexity is limited and not as efficient as the one of the human brain, along with the fact that such systems require vastly more instances and are less flexible in learning when compared to human children, are sufficient reasons to think of them in very relevant respects as less intelligent than humans. Nevertheless, LLMs invite rethinking a concept that, despite being vague, may be more scientifically productive by excluding unjustified anthropocentric requirements, and may have better scientific grounds by being operationally applicable to a wide range of systems, from simple and biological to complex and artificial ones.

References
Allen, C. (2017). On (not) defining cognition. Synthese, 194(11), 4233–4249. https://doi.org/10.1007/s11229-017-1454-4
Baysan, U. (2021). Rejecting epiphobia. Synthese, 199(1–2), 2773–2791. https://doi.org/10.1007/s11229-020-02911-w
Brooks, R. A. (1995). Intelligence without reason. In L. Steels & R. A. Brooks (Eds.), The artificial life route to artificial intelligence: Building embodied, situated agents (pp. 25–81). Lawrence Erlbaum Associates, Inc.
Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., Lee, P., Lee, Y. T., Li, Y., Lundberg, S., Nori, H., Palangi, H., Ribeiro, M. T., & Zhang, Y. (2023). Sparks of Artificial General Intelligence: Early experiments with GPT-4. http://arxiv.org/abs/2303.12712
Cao, R. (2022). Putting representations to use. Synthese, 200(2). https://doi.org/10.1007/s11229-022-03522-3
Dretske, F. (1993). Can Intelligence Be Artificial? In An International Journal for Philosophy in the Analytic Tradition (Vol. 71, Issue 2).
Duncan, J., Assem, M., & Shashidhara, S. (2020). Integrated Intelligence from Distributed Brain Activity. In Trends in Cognitive Sciences (Vol. 24, Issue 10, pp. 838–852). Elsevier Ltd. https://doi.org/10.1016/j.tics.2020.06.012
Egan, Frances (2020). A Deflationary Account of Mental Representation. In Joulia Smortchkova, Krzysztof Dołrega & Tobias Schlicht (eds.), What Are Mental Representations? New York, NY, United States of America: Oxford University Press.
Facchin, M. (2023). Why can’t we say what cognition is (at least for the time being). Philosophy and the Mind Sciences, 4. https://doi.org/10.33735/phimisci.2023.9664
Gardner, H. (2011). Frames of mind: the theory of multiple intelligences. Basic Books.
Grzankowski, A. (2024). Real sparks of artificial intelligence and the importance of inner interpretability. Inquiry, 1–27. https://doi.org/10.1080/0020174X.2023.2296468
Haier, R. J. (2023). The neuroscience of intelligence. Cambridge University Press.
Li, K., Hopkins, A. K., Bau, D., Viégas, F., Pfister, H., & Wattenberg, M. (2022). Emergent world representations: Exploring a sequence model trained on a synthetic task. arXiv preprint arXiv:2210.13382.
Nirshberg, G. (2023). Structural Resemblance and the Causal Role of Content. Erkenntnis. https://doi.org/10.1007/s10670-023-00699-y
Ramsey, W. (2017). Must cognition be representational? Synthese, 194(11), 4197–4214. https://doi.org/10.1007/s11229-014-0644-6
Russell, S. (2016). Rationality and intelligence: A brief update. Fundamental issues of artificial intelligence, 7-28.
Sternberg, R. J. (2019). The Augmented Theory of Successful Intelligence. In The Cambridge Handbook of Intelligence. Cambridge University Press. https://doi.org/10.1017/9781108770422