The Center for Teaching and Learning

What is machine learning?

Machine learning is, fundamentally, a mathematical way of trying to find patterns in data. Most (but not all) traditional machine learning problems are classification problems. A bank might have a dataset that consists of lots of prior loans they've given, along with the income and credit score for each loan/borrower, and a note about whether the loan defaulted. The bank might then use a machine learning algorithm to try to find a pattern that can be used to predict which loans will be repaid and which won't. The data the bank uses to find that pattern is called training data.

What is deep learning?

Given the training data, a machine learning algorithm figures out a rule for classification. That rule always comes from some set of possible rules. For example, there might be a linear cutoff that looks something like this:
Predict repayment if:   income + creditscore*10 > 160,000

Another common rule type is a decision tree, which would look something like this:
Predict repayment if:  (creditscore > 750) OR (income > 100,000 AND creditscore > 700)

Deep learning is simply machine learning with a particular category of possible classification rules. The algorithm sets up a structure of "neurons" that is loosely inspired by how human brains are actually organized. Neurons are connected with different pathways, each of which can be weighted, ranging from no importance to high importance. This means that the set of possible classification rules is much bigger and more complex than the examples above. It usually takes a huge amount of training data and computing power, but the algorithm can find very complex patterns in the data.

What is a large language model (LLM)?

An LLM is a classifier, the result of training a neural network on a large set of natural language (e.g., English) material. Usually the task used for training is to predict the next word of the text, so a classifier might be given 200 words from the middle of a Wikipedia article and asked to predict the next word. One of the major recent breakthroughs in machine learning is the realization that you can train a neural network for this particular task, then use the resulting neural network as a starting point for other language-based tasks, rather than having to train a new model from scratch. The LLM already contains a lot of information, ranging from syntactic (adjectives precede nouns) to substantive (stop signs are red).

What is ChatGPT?

ChatGPT is a particular chatbot, meaning a computer program that is meant to carry on a conversation with the user in a human-like way. It was created by OpenAI, using the LLM called GPT as a starting point, with additional training to try to make it answer questions in a way that OpenAI finds desirable. It has extremely impressive performance in many ways, though there are certainly still limitations. OpenAI is continually trying to improve its performance, and rival companies are working hard to create competing products. One version is available for free online, but more powerful versions require payment.

Data privacy. ChatGPT and other LLM-based chatbots require users to create accounts, allowing LLM providers to collect personal information that may be shared with third parties or compromised in a data breach. In addition, default settings typically allow any content shared with the chatbots to be used in further training of the underlying models (see OpenAI’s instructions for opting out of training). Private data that was present in the training data can appear in what the chatbots generate.  

Intellectual property. Writers and visual artists have alleged that their work is being used as training material for generative AI tools without their consent, allowing the tools to produce output that emulates their distinctive style and content. Similarly, submitting a student’s work to an AI-detection tool may result in the student’s intellectual property being used for training purposes without their consent.

Bias. Because LLMs are trained on human-generated text, they will reflect and sometimes exacerbate the various biases of the humans who have written the most, and about whom the most is written, on the internet.  While OpenAI has worked to craft barriers to prevent its chat bots from producing racist or hateful speech, there is real potential for both overt and latent bias in what LLMs will generate.  

Content filtering practices. Training LLMs to avoid generating racist, sexist, and violent content requires providing the models with existing examples of this kind of content. The grueling work of identifying and labeling the content has been outsourced to poorly paid workers whose mental health suffers as a result.

Misinformation. Generative AI chatbots can and do provide incorrect or incomplete information. Because these chatbots are designed for human-like interaction, when they do present incorrect information, they do so with what human users may interpret as confidence. For this reason, even users who are adept at identifying false information in other contexts may be misled by a response from a chatbot like ChatGPT, especially if they are using it to learn more about a topic they are not already familiar with.  As chatbots continue to develop, they also are likely to learn more about the preferences of individual users, enabling self-reinforcing information “bubbles”; discerning accurate information as well as identifying the sources of information will likely be increasingly difficult. 

Financial accessibility. AI tools are not equally available to all students. For example, the best current version of OpenAI’s LLM is GPT4, which is primarily available through a paid subscription.  

Accommodations & Accessibility. Notably, some people have identified generative AI as an important potential tool for broadening accessibility. If students request the use of generative AI as an accommodation, please reach out to DAR to learn more.