Can AI create knowledge? A counter to Noam Chomsky et al.
“When you change the way you look at things, the things you look at change” ~ Max Planck
Why are humans different from animals? Though both species exhibit the ability for reasoning and retaining knowledge, it is the ability to think about abstract imaginary concepts that distinguishes humans from other known living beings. While dolphins and chimpanzees can learn how to solve puzzles or draw autobiographical portraits of themselves, they have a really hard time thinking about the future or intangible concepts like creativity, which catalyzed a sense and aspiration for human superiority for several millennia.
With the rise of machine learning, nonetheless, scholars and humans as a whole have had to reevaluate their place in the stratum of intelligence and species superiority. As of 2023, Artificial Intelligence (AI) systems have already surpassed average human performance in at least 6 key domains including image recognition, language understanding, and, with the emergence of Large Language Models, creative writing (Giattino 2023), leading scholars to debate the extent to which AI can really produce knowledge to the degree or greater to that of human intelligence.
Linguists like Chomsky postulate that AI systems can’t generate knowledge because true knowledge requires being able to provide explanations through deduction whereas AI can only provide descriptions through induction. Mathematicians like Cella or Computer Scientists like Vaswani claim that AI with the right fine-tuning can learn generalizable patterns that result in a new understanding of a knowledge space, it’s just a matter of having sufficient amounts of data and compute. This led me to ponder, to what extent can AI generate knowledge?
In this paper, I explore the ways in which knowledge can be generated through induction and deduction in the space of creative computing, focusing on 1) knowing that or descriptive knowledge, 2) knowing how or analytical knowledge, and 3) second order knowledge or explanatory knowledge. I argue that AI can generate knowledge, but today’s AI technologies are mostly limited to knowing that or descriptive knowledge, but perform less well for analytical and explanatory knowledge.
Defining knowledge
Let me start by defining knowledge. Steup and Neta (2020) states that knowledge is the cognitive success of acquiring an understanding about entities, objects, ideas, facts, or practical skills. Knowledge has different meanings depending on the language. In French, it translates into “connaissance”, connoting a higher emphasis on memorizable knowledge, while in Latin it translates into “cognitio” or “scientia”, which connotes an emphasis on the thinking processes for understanding something. Thus, epistemologists instead focus on demarcating the meaning of knowledge by breaking it down into different types of knowledge in which the key distinguishing variable is the source of knowledge. For the purposes of analyzing knowledge in the space of AI and creative computing, I focused on a spectrum that outlines knowledge on the dimension of complexity of critical thinking, going from descriptive to analytical to explanatory knowledge, lowest to highest level of complexity respectively.
AI can produce descriptive knowledge
AI can perform well for knowing that tasks and create descriptive knowledge. Chomsky et al (2023) criticized that ChatGPT and similar generative AI programs can’t produce knowledge because they only return the most probable outcome given a task and training dataset.
I argue that this argument is incorrect because, firstly, it inadequately defines what counts as knowledge. Steup and Neta (2020) explain that knowing how to do something is a different kind of knowledge from knowing a set of facts. Even though ChatGPT doesn’t know how to actually swim, for instance, it can know that swimming involves propelling oneself through strokes to create a force of buoyancy to float and move in water. Language models map out the probabilities that a given token follows another token to form words and sentences, and in this way language models are able to form representations about what is likely to be correct or known by humans (Wei et al. 2023; Vaswani et al. 2017). This goes to show how generative AI models such as language models or markov chains can create descriptive knowledge, assuming they are trained on a diverse enough dataset in which the probabilistic distribution of human-made labels correctly corresponds to the physical world.
Secondly, Chomsky’s claim makes an implicit expectation that AI systems have to get the true answer on the first try regardless of the available training data to be attributed with the quality of producing knowledge, while hypocritically humans are exempt from this expectation. Descriptive knowledge is always dependent on the available information. Aristotelian physics that claimed that matter was made up of earth, air, fire, water and aether, was considered descriptive knowledge up until Galileo Galilei and Isaac Newton in the 17th century. Generative AI and most machine learning systems are really good at constructing representations that discern what are the statistically true patterns in their data, and by doing so they produce descriptive knowledge, which offers a unique understanding of the world they perceive.
AI can superficially create analytical knowledge
Large-Scale Deep Learning performs imperfectly for knowing how tasks and can generate analytical knowledge to a limited extent. Chomsky et al’s (2023) second point of criticism was that AI systems can’t produce insightful knowledge because they are unable to rationally conjecture highly improbable but correct theories due to their tendency to stick to what the majority of the data point to. I agree and disagree with this claim.
I agree because, as Steup and Neta (2020) contend, knowledge of facts and conjectures, which is more complex than descriptive knowledge, can only be produced by being able to identify what is important and what isn’t in order to make abstract connections that are nonexistent in the training data. As machine learning is highly reliant on inductive logic, which computes similarity or variation on either the raw data or the slightly transformed data representations, it has a really hard time trying to make connections on dimensions that do not exist in the training data, suggesting that deduction may be a necessary component for analytical knowledge.
Fernandez and Vico (2013) found that the AI methods implemented for music composition usually only provided meaningful compositions when recreating pieces similar to the original training sound files, but more importantly meaningful results tend to be the exception rather than the expectation, showing how AI systems perform poorly for fields that deal more in the metaphysical space such as music. Cella (2020) explains that even powerful methods for abstract representations like variational autoencoders often failed because they were stuck clustering descriptive knowledge like rhythm or harmony, but couldn’t adequately capture more complex concepts like timbre or style because those are labels that are not present in the training data.
I also disagree, nonetheless, because there have also been examples of AI systems that perform better at knowing how tasks and analytical knowledge. Vaswani et al (2017) found that integrating an Attention mechanism to neural networks enabled models to analyze complex tasks better by forcing the model to first evaluate what context is relevant and what is irrelevant for the task at hand before making predictions. In this way, transformers are better adapted to find improbable theories that offer unique explanations, which is why Large Language Models can be prompted to solve problems of how a given object might move when exerted with a vector of forces and return accurate predictions with explanations for why it’s so. Even for simpler methods like non-negative matrix factorization (NMF), AI systems are able to break apart complex sounds into human-like labeled parts like high-pitched sound and percussion by imposing a non-negativity constraint on the reconstruction loss function, which approximates the way that humans reason to produce analytical knowledge (Lee and Seung, 1999).
Despite this, I argue that Large-Scale Deep Learning and NMF are limited to producing superficial analytical knowledge, because they lack the ability to draw abstract representations outside of the training data set (Gatys et al. 2016; Wei et al. 2023). One way to overcome this is by designing systems that can transfer knowledge from one domain to another, which would emulate one prominent way in which humans produce analytical knowledge, and we’ll have to see the extent to which multimodal AI manages to conceptualize at different levels of analysis to produce analytical knowledge.
AI can’t generate explanatory knowledge
The AI systems of 2023 perform poorly for second order knowledge tasks and can’t produce explanatory knowledge. Chomsky et al’s (2023) third point of criticism was that AI systems can’t produce explanatory knowledge because their source of knowledge is unidimensional, dependent on the way the data is measured, and are thus incapable to form multidimensional knowledge frameworks which are essential to distinguishing between truths that are possible from truths that are impossible.
At the second order of knowledge, epistemologists and philosophers are concerned with how we know what we know, and whether a different combination of sources of knowledge would alter what we perceive as true and false.
How would an AI of 2023 try to understand whether a card facing down is an ace of spades? It would compute the probability of how many ace of spades are in a deck of cards and maybe see a history of how often it appears in a game, and return roughly 154. But quantum physics states that a card exists in multiple states, so before flipping it, the card is all possible options at once and none at the same time, and is therefore always and never the ace of spades. At the same time, a deterministic philosopher like Schopenhauer states that the card is a manifestation of the will, where the will is a metaphysical force that predetermines reality even before they occur, and so the card will be an ace of spades if it’s destined to manifest one. Steup and Neta (2020) explain these are mere interpretations of the truth and only become explanatory knowledge until one is able to offer an adequate justification in the form of, given p sources of knowledge q is justifiably true because of how p sources logically connect to q, where sources of knowledge include reason, sense perception, emotion, faith, intuition, imagination, memory, and language.
The issue with AI, especially models that follow inductive logic like machine learning (Fernandez and Vico, 2013), is that it can’t provide the chain of thought to justify why one interpretation is valid, as it can’t compare and contrast against ideas that are not measured. AI simply replicates the biases of the unidimensional data, which is why it can produce descriptive knowledge and sometimes analytical knowledge, but I don’t see how AI can evaluate sources of knowledge and offer justifications to produce explanatory knowledge. Given that even humans have a hard time producing explanatory knowledge as not everyone can understand and justify that something is true and false at the same time, it seems to me that AI will not be able to produce explanatory knowledge until it reaches a comparable level of consciousness.
Chomsky et al is correct and incorrect about AI
In conclusion, AI systems can produce descriptive knowledge by mapping out representations in their training data and providing their own unique statistical understanding of the projected similarities and variance. AI can also produce analytical knowledge though only superficially by being able to use abstract concepts to predict outcomes, but it’s limited as it only generates abstract concepts that closely stem from training data.
Nonetheless, I agree with Chomsky et al (2023) in that for knowledge that requires higher levels of critical thinking, AI is yet not able and probably will not be able to produce explanatory knowledge. AI can only make unidimensional interpretations as it can only use sources of knowledge that can be neatly gathered in the form of raw data, and thus it’s incapable of offering a justification to move from interpretation to demonstrate a deeper understanding for why something is true according to a particular source of knowledge but false according to a different source of knowledge, which is necessary to produce explanatory knowledge.
An area of future research is examining whether the production of knowledge graphs with new labels that can’t be found in the training data counts as analytical knowledge or is a form of disguised descriptive knowledge. Ultimately, knowledge is justification and a whole lot of belief.
Acknowledgements
Thank you very much for reading all this way! I wrote this paper for my class taught by Professor Carmine-Emanuele Cella where I learned about deep learning and how to apply it to analyze and create music. It was fantastic and I hope my paper got you thinking deeper about AI.
If you want to stay up-to-date with stuff I write, feel free to follow me:
https://twitter.com/RyuseiBest
https://www.linkedin.com/in/ryusei-best-hayashi/
Bibliography
Cella, Carmine-Emanuele. “Music Information Retrieval and contemporary classical music: A successful failure.” Transactions of the International Society for Music Information Retrieval, vol. 3, no. 1, 1 Sept. 2020, pp. 126–136, https://doi.org/10.5334/tismir.55. Accessed 10 December 2023.
Cella, Carmine-Emanuele. “Orchidea: A comprehensive framework for target-based computer-assisted dynamic orchestration.” Journal of New Music Research, vol. 51, no. 1, 21 Dec. 2022, pp. 40–68, https://doi.org/10.1080/09298215.2022.2150650. Accessed 10 December 2023.
Chomsky, Noam, et al. “Noam Chomsky: The False Promise of ChatGPT.” The New York Times, 8 Mar. 2023. Accessed 10 December 2023.
Colton, Simon. “Creativity Versus the Perception of Creativity in Computational Systems.” AAAI Spring Symposium: Creative Intelligent Systems, 2008.
Fernandez, Jose David, and Francisco Vico. “AI methods in Algorithmic composition: A comprehensive survey.” Journal of Artificial Intelligence Research, vol. 48, 17 Feb. 2013, pp. 513–582, https://doi.org/10.1613/jair.3908. Accessed 10 December 2023.
Gatys, Leon A., et al. “Image style transfer using Convolutional Neural Networks.” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, https://doi.org/10.1109/cvpr.2016.265. Accessed 10 December 2023.
Giattino, Charlie, et al. “Artificial Intelligence.” Our World in Data, 11 Oct. 2023, ourworldindata.org/artificial-intelligence#:~:text=few%20key%20countries-,AI%20systems%20perform%20better%20than%20humans%20in%20language%20and%20image,two%20decades%20of%20AI%20development. Accessed 10 December 2023.
Lee, Daniel D., and H. Sebastian Seung. “Learning the parts of objects by non-negative matrix factorization.” Nature, vol. 401, no. 6755, 21 Oct. 1999, pp. 788–791, https://doi.org/10.1038/44565. Accessed 10 December 2023.
López de Mántaras, Ramón. “Artificial Intelligence and the Arts: Toward Computational Creativity.” BBVA OpenMind, Artificial Intelligence Research Institute, 2016, www.bbvaopenmind.com/en/articles/artificial-intelligence-and-the-arts-toward-computational-creativity/. Accessed 10 December 2023.
Mallat, S.G., and Zhifeng Zhang. “Matching pursuits with time-frequency dictionaries.” IEEE Transactions on Signal Processing, vol. 41, no. 12, Dec. 1993, pp. 3397–3415, https://doi.org/10.1109/78.258082. Accessed 10 December 2023.
Rhodes, Mel. “An Analysis of Creativity.” The Phi Delta Kappan, vol. 42, no. 7, 1961, pp. 305–10. JSTOR, http://www.jstor.org/stable/20342603. Accessed 10 December 2023.
Steup, Matthias, and Ram Neta. “Epistemology.” Stanford Encyclopedia of Philosophy, Stanford University, 11 Apr. 2020, plato.stanford.edu/entries/epistemology/. Accessed 10 December 2023.
Vaswani, Ashish, et al. “Attention Is All You Need.” 31st Conference on Neural Information Processing Systems, 2017, https://doi.org/10.48550/arXiv.1706.03762. Accessed 10 December 2023.
Wei, Jerry, et al. “Larger language models do in-context learning differently.” Google Research, 8 Mar. 2023, https://doi.org/https://doi.org/10.48550/arXiv.2303.03846. Accessed 10 December 2023.
