Knowing Neurons
Artificial IntelligenceLearning and MemorySensation and Perception

Does ChatGPT think like a human?

By Carolyn Amir

ChatGPT, an artificial intelligence (AI) platform designed to interface with humans, is the ultimate search engine and virtual conversation partner. The AI can improve your computer code, customize your fitness training plan, and even write that pesky essay for Advanced Placement English*. While there are issues with assessing consciousness when it comes to today’s machine learning models, robots have been passing the Turing Test, an experiment that tests whether a machine can pass for a person, since 2014. This raises the question: does ChatGPT think like a human?

The machine uses artificial neural networks to perform its many complicated tasks, drawing from the vast amount of information it has been fed


ChatGPT is a Large Language Model (LLM), a type of AI trained on large quantities of input data (Shen et al., 2023). The machine uses artificial neural networks to perform its many complicated tasks, drawing from the vast amount of information it has been fed. This method of computing is sometimes referred to as “self-supervised” or “deep learning”. Deep learning algorithms use unlabeled data—the information fed to it initially—through neural networks. Inputs are continuously passed through “hidden layers” of these neural networks. These computations are used to generate the neural network’s output: ChatGPT’s response.


A diagram showing nodes connected to each other demonstrating a neural networkA neural network schematic showing hidden layers of interconnected information-processing nodes.

Input into these models is often referred to as “training” data. LLMs require millions or even billions of inputs to carry out the tasks for which they are designed (Wei et al., 2022). Training sets can run up to trillions of words and often include text from large datasets, such as Wikipedia and Github (Anil, 2023; Whang, 2023). The AI draws upon these inputs in order to formulate its responses. We do not actually know what ChatGPT was trained on, as there is currently no transparency behind the sources used for the ChatGPT model AI (Hill-Yardin et al., 2023).

An example of prompting a large language model. Reviews of movies are fed to the model, training the model to classify reviews as positive or negative.  Image Source: Wei and Tay 2022 


This vast amount of information allows the computer to iteratively respond to a user’s questions, building upon previous back-and-forths. These interactions are known as “prompting” (Brown et al., 2020). In this way, the user’s own questions and responses to the machine begin serving as training data, even after the machine has already been trained with the vast amount of initial input text. This back-and-forth mimics a real-world conversation; as the machine responds to feedback and remembers what a user says, a user may feel that the AI is responding more accurately to their prompting. 

ChatGPT is built to sound logical. It is trained with human-generated articles, absorbing our way of writing and communicating. If its initial answer to a user’s  query is unreasonable, the user can continue giving inputs that serve as feedback for the machine to fine-tune its responses. However, given that they rely on input data in this way, ChatGPT and other LLMs are only as good as the information they are fed. 

It is often difficult to distinguish between what is relevant information and what is irrelevant, and it is remarkable how quickly our brain can conduct these computations

Google and other similar search platforms allow the user to sift through endless sources to come to any given conclusion; a searcher may pass by dozens of unrelated articles, inaccurate information, sponsored content, and fake news. Similarly, there is no way of knowing whether ChatGPT is returning reliable information, or what the source of that information might be. The machine is essentially parroting responses – source unknown – from a vast repository of human information. This raises the question: how does the human mind choose an answer amongst a swath of options? It is often difficult to distinguish between what is relevant information and what is irrelevant, and it is remarkable how quickly our brain can conduct these computations. Do our brains make these decisions from thousands of language inputs, and feed these forward to generate an output – just like neural networks? Just like ChatGPT?

“Neural networks” get their name from the way information is represented in the brain: by individual brain cells, also known as neurons. Each neuron  is interconnected with other neurons. In a neural network model, the analog of a neuron is referred to as a “node”, which is where a piece of information is stored. Each node has a “weight” connecting it to other nodes. These weights represent how strongly nodes are connected to one another; similarly, some neurons are more strongly interconnected than others.

 

Left: A schematic drawing of the biological neural network. Two neurons interface via neurophysiological mechanisms. Right: a simple representation of a computational neural network. Inputs are weighted and connected via layered nodes before output. Image Source: Makinde and Asuquo, 2012 


The importance and salience of information are represented in several parts of our brain, including regions responsible for vision, sensation, and other regions sometimes referred to as the “salience network”. The salience network interfaces with the visual areas in our brain which are responsible for externally-driven cognition, and other areas responsible for internally-driven cognition to help us decide what information is important, and what information is unimportant (Vossel et al., 2014). These regions communicate with one another through shared neural connections. Some of these biological connections are stronger than others—just like weights in a neural network.

Filtering out relevant information depends on the learned ability to discriminate between stimuli that are important versus stimuli that are unimportant. Fascinatingly, the brain’s responses to parsing information can be observed at the fine-grained level of an individual brain cell. The brain cells responsible for learning specific information respond adaptively from day to day depending on the information they are given (Poort et al., 2015). Changes in the brain might also allow us to process important information from our environment more efficiently and may underlie our ability to react promptly to important information.

Fascinatingly, the brain’s responses to parsing information can be observed at the fine-grained level of an individual brain cell


Results of a recent study (Leong et al., 2017) also revealed that selective attention is used to determine the value of different options in front of us. When something unexpected happens, selective attention shapes what we learn. For example, if an exam goes poorly, we attribute this to features we paid attention to, and we ignore ones we did not. If you were focused on how you forgot to wear your lucky scarf, you might remember this more vividly than how you missed that one textbook chapter you should have studied. The study showed that what we pay attention to creates a feedback cycle: we draw conclusions based on what we paid attention to and attend to what we have attributed high value – in other words, what was “salient”. The next exam, you will be sure to remember that lucky scarf.

Our expectations and environmental context have a great impact on how we process external information. Contextual cues provide information that is relevant to our behaviors and can make our behavior more efficient. Many studies show our bias to attend to stimuli is influenced by the stimuli’s salience (e.g. DiQuattro et al., 2011). Further, our own prior beliefs, emotional states, and biases all contribute to how we process, attend to, and ultimately respond to information. 

So, while ChatGPT relies on its networks of text-based information to parse user prompts, the human brain relies on salience, environment, contextual cues, emotions, biases, episodic memory, and directed attention. When we ask ChatGPT a question, it draws from the interconnectedness of its nodes and its hidden layers built from language input. When we are asked a question, we draw from our vast range of experiences, prior beliefs, and much more. 

… while LLMs can process trillions of words, our human brains store vastly more complex information…

Attention to important information is an incredibly complex process in the brain and is just one component of what the human brain does. True, there are some superficial similarities between the human brain and LLMs. But while LLMs can process trillions of words, our human brains store vastly more complex information; we encode everything from the answers to the big math exam, to fine-grained sensations of how our breakfast tasted this morning, to vivid details of embarrassing moments from years ago. The way this information is stored is more distributed, biologically complex, and relies on years of memories and experiences across sensory domains. ChatGPT can be a useful tool if we remember that it is designed to sound like it knows what it is talking about. 

~~~

Written by Carolyn Amir
Illustrated by Federica Raguseo
Edited by Zoe Dobler, Daniel Janko, and Caitlin Goodpaster

~~~

Become a Patron!

image of a computer smiling saying "hi my name is ChatGPT and I am a text-based AI model"

References

Anil, Rohan; et al. (2023). “PaLM 2 Technical Report”. arXiv:2305.10403

Brown, T., Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. https://arxiv.org/abs/2005.14165.

DiQuattro, Nicholas E., and Joy J. Geng. “Contextual knowledge configures attentional control networks.” Journal of Neuroscience 31, no. 49 (2011): 18026-18035. doi: 10.1523/JNEUROSCI.4040-11.2011.


Hill-Yardin, E. L., Hutchinson, M. R., Laycock, R., & Spencer, S. J. (2023). A Chat (GPT) about the future of scientific publishing. Brain Behav Immun, 110, 152-154. doi: 10.1016/j.bbi.2023.02.022


Leong, Y. C., Radulescu, A., Daniel, R., DeWoskin, V., & Niv, Y. (2017). Dynamic interaction between reinforcement learning and attention in multidimensional environments. Neuron, 93(2), 451-463. doi: 10.1016/j.neuron.2016.12.040.

Shen, Y., Heacock, L., Elias, J., Hentel, K. D., Reig, B., Shih, G., & Moy, L. (2023). ChatGPT and other large language models are double-edged swords. Radiology, 307(2), e230163.


Poort, J., Khan, A. G., Pachitariu, M., Nemri, A., Orsolic, I., Krupic, J., … & Hofer, S. B. (2015). Learning enhances sensory and multiple non-sensory representations in primary visual cortex. Neuron, 86(6), 1478-1490. doi: 10.1016/j.neuron.2015.05.037.

Vossel, S; Geng, JJ; Fink, GR (April 2014). “Dorsal and ventral attention systems: distinct neural circuits but collaborative roles”. The Neuroscientist. 20 (2): 150–9. doi:10.1177/1073858413494269. PMC 4107817. PMID 23835449.

Wei, J.; Tay, Yi; Bommasani, Rishi; Raffel, Colin; Zoph, Barret; Borgeaud, Sebastian; Yogatama, Dani; Bosma, Maarten; Zhou, Denny; Metzler, Donald; Chi, Ed H.; Hashimoto, Tatsunori; Vinyals, Oriol; Liang, Percy; Dean, Jeff; Fedus, William (31 August 2022). “Emergent Abilities of Large Language Models”. Transactions on Machine Learning Research. ISSN 2835-8856. https://arxiv.org/pdf/2206.07682

Whang, O. (2023, May 30). Would Large Language Models Be Better If They Weren’t So Large? The New York Times. https://www.nytimes.com/2023/05/30/science/ai-chatbots-language-learning-models.html

*Plagiarism not recommended.

Author

  • Carolyn Amir

    Carolyn Amir is a Neuroscience PhD student at UCLA in Dr. Carrie Bearden's lab. Her current research focuses on using neuroimaging and genetics approaches to study the human brain and behavior in psychiatric and developmental disorders. Before coming to UCLA, Carolyn worked at the National Institutes of Health where she studied human pain, medication use and expectancy. She holds a Bachelor's degree from Boston University in Psychological and Brain Sciences, where she used EEG and brain stimulation to examine the effects of neuromodulation in both healthy and patient populations. When she's not in the lab, Carolyn enjoys camping, hiking, and spending time at the beach.

Avatar photo

Carolyn Amir

Carolyn Amir is a Neuroscience PhD student at UCLA in Dr. Carrie Bearden's lab. Her current research focuses on using neuroimaging and genetics approaches to study the human brain and behavior in psychiatric and developmental disorders. Before coming to UCLA, Carolyn worked at the National Institutes of Health where she studied human pain, medication use and expectancy. She holds a Bachelor's degree from Boston University in Psychological and Brain Sciences, where she used EEG and brain stimulation to examine the effects of neuromodulation in both healthy and patient populations. When she's not in the lab, Carolyn enjoys camping, hiking, and spending time at the beach.