Reclaiming AI as a Theoretical Tool for Cognitive Science

Abstract

The idea that human cognition is, or can be understood as, a form of computation is a useful conceptual tool for cognitive science. It was a foundational assumption during the birth of cognitive science as a multidisciplinary field, with Artificial Intelligence (AI) as one of its contributing fields. One conception of AI in this context is as a provider of computational tools (frameworks, concepts, formalisms, models, proofs, simulations, etc.) that support theory building in cognitive science. The contemporary field of AI, however, has taken the theoretical possibility of explaining human cognition as a form of computation to imply the practical feasibility of realising human(-like or -level) cognition in factual computational systems, and the field frames this realisation as a short-term inevitability. Yet, as we formally prove herein, creating systems with human(-like or -level) cognition is intrinsically computationally intractable. This means that any factual AI systems created in the short-run are at best decoys. When we think these systems capture something deep about ourselves and our thinking, we induce distorted and impoverished images of ourselves and our cognition. In other words, AI in current practice is deteriorating our theoretical understanding of cognition rather than advancing and enhancing it. The situation could be remediated by releasing the grip of the currently dominant view on AI and by returning to the idea of AI as a theoretical tool for cognitive science. In reclaiming this older idea of AI, however, it is important not to repeat conceptual mistakes of the past (and present) that brought us to where we are today.

Ways we talk about AI:

  1. AI as "the idea that intelligence can be recreated in artificial systems (Russell & Norvig, 2010)"
  2. "it refers to an artificial system believed to implement some form of intelligence (i.e., ‘an AI’)."
    • discussion whether that's only narrow domain-specific AI or also AGI
  3. You can do AI too, AI as "the research and/or engineering field pursuing the creation of AI systems based on the idea that doing so is possible and desirable. AI as a research field is historically diverse (see Table 1), including AI-as-engineering approaches that focus on different domains of application and draw on different modeling and engineering frameworks, including not only the currently popular machine learning approach, but also approaches involving Bayesian Networks, decision support systems, and robotics."
  4. AI as "the ideology that it is desirable to replace humans (or, specifically women) by artificial systems" (I personally haven't encountered any theory putting forth that women should be replaced by AI, but perhaps there's weird posthuman stuff out there?)
    • Example: replacing participants in psychological experiments with AI (LOL!)
  5. AI as a way "of using computational tools to develop theories of natural cognition."

It's the latter (5) that Van Rooij argues is a lost meaning of AI and that deserves revaluation.

Simon, H.A. (1983). Why should machines learn? Elsevier. Machine learning (pp. 25–37):

Artificial Intelligence has two goals. First, AI is directed toward getting computers to be smart and do smart things so that human beings don’t have to do them. And second, AI (sometimes called cognitive simulation, or information processing psychology) is also directed at using computers to simulate human beings, so that we can find out how humans work. (Simon, 1983, p. 27)

To halt the rerun of history, we think it is vital that the idea of cognition as computation—and therefore the in-principle possibility of realising and/or explaining cognition as a form of computation—is not mistaken for the practical feasibility of replicating human minds in machines (which we prove is not feasible).) (Van Rooij, 2024)

Flying pigs are also possible in principle; possible in principle bakes no bread. (Fodor 2005. Reply to Steven Pinker 'So how does the mind work?', p.27)

Two parts of the overal argument:

  1. Formal proof that the problem of human-like cognition is computationally intractable and that therefore current claims that AGI is inevitable and almost here are unfounded.
  2. Why we need to revalue AI as a theoretical tool for cognitive science (assuming that cognition can be studied as a form of computation) without repeating "the mistakes of the past", which in this case in the assumption that "computationalism implies the practical feasibility of thinking machines/simulations on par with human-level or -like cognition". (I focused less on this part in these notes.)

This mistake was made as follows: the working hypothesis was that human thinking is completely information-processing activity. Hence, if we also believe that these information processes can be explicated and modelled, that it seems possible that we can re-create (aspects of) cognition with computers.

The formal proof mainly debunks claims that we are on a fast track to AGI and/or that AGI is inevitably, because such strong claims (that you often hear nowadays) cannot prove that there is a tractable path to reach it. They cannot possibly prove this, according to Van Rooij et al., because Van Rooij et al. have proven that the problem of human-like cognition is computationally intractable.

"if you think your AI is very human-like, then you are not testing it critically enough (cf. Bowers et al., 2023)."

Being better than the state of the art is not the same as being good.

The prioritization of performance values is so entrenched in the field that generic success terms, such as ‘success’, ‘progress’, or ‘improvement’ are used as synonyms for performance and accuracy [...] However, models are not simply ‘well-performing’ or ‘accurate’ in the abstract but always in relation to and as quantified by some metric on some dataset” (Birhane et al., 2022).

The errors cannot be contained to be small

Right now, by thinking that current AI (Van Rooij calls the paradigm "AI-by-learning") is a step towards human-like cognition, we "are inducing distorted and impoverished images of ourselves and our cognition"; "the AIs do not actually approximate human behaviour. By nevertheless taking the AIs as cognitive models, we—as a field—distort our view of cognition, and it makes our cognitive science theoretically weaker."

I often say that a good way to understand something is by building it. Cf. the Feynman technique.

Although not discrediting this idea in itself, Van Rooij et al argue that

all makeists think that building cognition is sufficient for being able to explain it (b), and some think that this building is also necessary (c). The necessity claim especially reveals the implicit assumption that it is possible to build cognition (a). While some things can be understood by making them, it will not work for human-like or -level cognition, for one because this cannot plausibly be (re)made through engineering.

The methodology endorsed by makeists has been referred to as the synthetic methodology or understanding by design and building (Bisig & Pfeifer, 2008; Pfeifer & Scheier, 2001); also see formal realism (Chirimuuta, 2021). A well-known quote from Feynmann (1988), 'what I cannot create, I do not understand', is often used to support the idea of makeism in AI (e.g. Karpathy et al., 2016) Note that it is especially easy for makeists to fall into map-territory confusion--mistaking their modeling artefacts for cognition itself--due to the view that the made thing could be cognition.

I quite like the sharp criticism of Van Rooij. Ultimately her goal is to show that AI as it's being done now (engineering, ML, claims of AGI) is a "decoy" of cognition and thus inhibits AI as a theoretical tool for cognitive science.

This does not in any way say that AI is not useful, it just counters grand claims about AI being on a fast track to AGI.

Nice paragraph on the "map-territory confusion". Simulation done right and done wrong:

Importantly, this use of simulations is to be distinguished from makeist uses of simulation that confuse the models (explanans) for the thing modelled (explanandum) and/or take the simulation results to directly imply something about ‘how things work’ in the real world (e.g. for real-world brains, cognition, or behaviour). Instead, non-makeist computer simulations are theoretical tools that can demonstrate proof of concept or demonstrate the in-principle (im)possibility of phenomena arising from the theorised constructs and hypothesised mechanisms. Computer simulations support and extend a scientist’s thinking capacity and enable computerised ‘thought experiments’ (Cooper, 2005) to reason through ‘what ifs’ and answer questions like ‘how possibly’.Footnote 22 These simulations—as indeed any models that the cognitive scientist could use—are necessarily abstract and idealised; this is unproblematic, though, as long as the scientist recognises it and takes care to draw only those inferences which are really warranted by the model.