Large Language Models (LLMs), like ChatGPT or Google's Gemini, learn by analyzing huge amounts of text data from various sources. Understanding these sources helps explain why LLMs produce certain responses.

Types of Training Data

Data is collected from publicly accessible websites by automated web crawlers providing diverse content:

Internet
Books and literature
Wikipedia and knowledge bases
Programming code
Images

Examples (select to reveal)

Internet: Models like ChatGPT use text from publicly available web pages gathered by web crawlers. This provides diverse content but includes inaccuracies, biased information, and otherwise harmful or outdated content which models will learn from. In addition, many have criticized the lack of permission involved in scraping websites, while proponents argue that this practice has been employed for years to power tools like Google Search.
Books and literature: LLMs use collections of books, both public domain and proprietary. Book data offers richer linguistic capabilities as these sources provide better represent well-edited, thoughtful language but has sparked controversy over use of copyrighted texts without permission.
Wikipedia and knowledge bases: These platforms provide structured, factually dense, and regularly updated content, helping models develop a baseline of factual accuracy on an extensive range of topics, from detailed biographies to scientific concepts. While valuable for including information about general world knowledge, although they don’t prevent hallucination or learning inaccurate or false information from less reliable sources.
Programming code: Because modern LLMs are increasingly used to assist with programming tasks, they draw on open-source code from platforms like GitHub, enabling code generation and improving grammatical accuracy.
Images: Images may be drawn from available web pages gathered by web crawlers or from databases available to the public, which may have copyright implications.

Encoded Bias

LLMs learn from real-world data, absorbing biases from their sources. Internet text often contains stereotypes or biased perspectives, which models unintentionally reproduce. For example, frequent negative portrayals of a group in training data can lead to the model repeating these biases. Recognizing and mitigating these biases is crucial. Developers and researchers are continually working on strategies to:

Curate data more carefully.
Use bias detection tools.
Implement techniques to reduce biased outputs.

Still, it's important to remember current LLMs contain biases that could impact their responses and take responsibility for potential issues in content generated through your chatbot interactions.

Public vs. Licensed Data

Although data may be publicly accessible, it does not always mean developers have clear permission to use it. This issue has sparked significant ethical and legal debates regarding ownership and licence arrangements.

Public data is easy to access for training LLMs but may lead to escalation of issues around permission and copyright.
Licensed data is well-edited, fact-checked, reliable, and ensures authors receive compensation.

For additional guidance visit University of Toronto Libraries: GenAI and Copyright Considerations

User Interactions

Models like ChatGPT also learn from actual user interactions (prompts, corrections, follow-ups), enabling them to adapt and better handle real-world tasks. However, it's important to be cautious, as user-submitted data could potentially reappear in future responses.

Your Data in LLMs

When you use services like ChatGPT, your interactions can be used to help train and improve the model. This means your shared information, like personal details or assignments, or private information, might shape future responses. For example, if you submit your own essay or original work, the model might generate similar text in the future, potentially sharing your ideas with others.

To keep your data safe and maintain privacy:

Think twice before sharing sensitive or personal details.
Avoid inputting private information like your full name, student ID, passwords, or confidential research data.
Always check how a service handles your data to make informed decisions.

Remember, being cautious about what you share online helps protect your privacy and ensures responsible use of technology.

For additional guidance visit the University of Toronto Information Security Guidelines on use of GenAI.

How Data is Used

During training, LLMs repeatedly see text sequences and attempt to predict the next word based on the context. The model has billions, or even trillions of internal parameters that are adjusted each time the prediction differs from the actual next word. Over time, the model gradually becomes more adept at correctly predicting the next word based on language structure, grammar and more.

This training process helps to explain why these models can provide coherent and contextually relevant responses. However, they don’t “understand” language the way we do – they simply predict words based on highly complex patterns in the data. This is also why models can generate plausible but incorrect information, as their predictions are based on statistical modelling rather than a sense of truth or understanding.

↑ Back to Top