This introductory guide is designed to help students and researchers begin exploring the responsible use of AI and generative AI in academic work. It supports the development of practical skills in using AI tools effectively and ethically.
For information about the university's ChatGPT Edu licence, visit the 'Generative AI at Oxford' page for information about access and training:
🛠️ This resource is a work in progress.
As AI and the University's guidance evolve, we will continue to update this guide. Please note that, due to the rapid pace of AI updates, there may be a delay in this resource reflecting the latest changes.
Artificial intelligence (AI) refers to computer systems that perform tasks typically requiring human intelligence, such as recognising patterns, understanding language, or making decisions. AI already plays a role in many of our everyday interactions, from the recommendations we see on streaming platforms to the spam filters in e-mail.
There are many types of AI, but one of the most talked-about forms is generative AI (GenAI), which can create new content such as text, images, audio, and code based on patterns learned from large datasets. GenAI platforms like ChatGPT produce responses that mimic human-like outputs. However, they may also generate inaccurate, misleading, or biased information, so their use requires critical thinking and careful evaluation.
A large language model (LLM) is a type of machine learning model trained on vast amounts of text data to understand and generate human-like language. These models can support a variety of tasks, such as answering questions, summarising information, and translating languages. LLMs work within a context window - they can "remember" a limited amount of text at one time (including both the user’s input and the model’s response).
OpenAI's GPT-5 and GPT-5 mini are recent examples of LLMs, demonstrating significant leaps in performance compared to previous models.
Further information:
A hallucination is when an AI model generates a response that appears plausible but is factually incorrect or nonsensical. This phenomenon may occur for a variety of reasons, including a lack of sufficient training data, data biases, or wrong assumptions made by the model. For example, ChatGPT might confidently provide a fabricated reference to support one of its claims, but a quick Google search will reveal the source does not exist.
Hallucinations could be an inherent feature of LLMs.[1] Improving training data or model architecture may not entirely remove them from future models, underscoring the need for fact-checking when using GenAI for academic work.
Further information:
Machine learning (ML) is a subset of AI that involves training algorithms on data. These algorithms identify patterns in the material they're given and create models that can then make decisions or predictions on new data without needing explicit instructions. ML allows systems to improve their performance as they process more data; however, the quality of their output depends on what they learn from, which means they can sometimes reflect errors or biases.
Further information:
When interacting with LLMs, a prompt is an input (e.g. a question, instruction, or statement) that is most often typed in to a GenAI platform, such as ChatGPT or Microsoft Copilot. What is returned is a specific response based on that input.
The effectiveness of a prompt impacts the quality of an AI's output. Crafting clear and relevant prompts is known as "prompt engineering", and further guidance can be found on this guide in the section "Prompting Guidance for GenAI".
Further information:
1. Xu, Z., Jain, S., and Kankanhalli, M. (2024). Hallucination is Inevitable: An Innate Limitation of Large Language Models. Pre-print at: https://doi.org/10.48550/arXiv.2401.11817
If you'd like to give feedback about this guide, please contact Bodleian User Education.
For other AI-related queries, please see our "Further Resources and Support" page in this guide to find a suitable contact or resource.