There are many ways in which generative AI can be used to aid your learning. We don’t think you should be afraid to use generative AI when appropriate, the aim of this page is to help you develop a more critical approach (and so learn more effectively).

Before we dive into exploring these different options, the first thing you will need to do is to check the policies that your institution or department has. Some of these polices will be quite open to students using generative AI while other will be far more restrictive. You should in all circumstance follow these policies.

Approaches to using generative AI in your Learning

Given the pace of development and scale of generative AI tools it would be impossible to list all the possible uses of it, and while we will try to give a few concrete examples, these may date quickly. This section will therefore concentrate on how we learn and how the use of AI can impact on the learning process. The aim here is for you to consider your own learning process and reflect on the usage of generative AI. It has the potential to be both useful but also harmful on these learning processes.

Be sceptical

Most of the common generative AI tools available to you are based on large language models. These have been designed to predict the next word in a sequence, optimised to generate plausible sounding output. You’ll notice that the word “understanding” is not found in that sentence. Chatbots are often trained to sound helpful, confident and supportive – yet this can become sycophantic and lead to poor advice. We recommend you to check and challenge generative AI output and understand when it is likely to be helpful and when it is not.

To use or not to use that is the GenAI question

Generative AI has the potential to be used in many different ways when it comes to your studies, but before you start to trying out all of these different applications you should take Jeff Goldblum’s character’s advice in Jurassic Park:

Generative AI can easily be used in a way that is detrimental to your learning. Darvishi et al. (2024) found a risk of students depending on, rather than than learning from generative AI tools and not being able to perform without them. When it comes to learning the process is key. Learning can be difficult and time consuming, but if you try to shortcut this process you can end up not learning anything at all. This is sometimes referred to as cognitive offloading. Ultimately it is important that you are confident that you are learning about the subject you are studying and becoming more knowledgable in the field, not just getting better at crafting prompts for AI tools!

When to use generative AI

When you are confident you can accurately evaluate the AI response

Generative AI can produce an answer on any given topic, and these often sound convincing as the algorithms have been tuned to create plausible output. To determine if the answer is any good requires a certain level of subject matter expertise. If you are using generative AI to find out about fields where you have little prior knowledge, then it can sometimes be difficult to determine whether or not the answers generative AI provides are correct. 

Translating for different audiences

Generative AI is good at taking an existing text and translating it to another language or rewriting it for a different audience (for example, taking an academic article and rewriting it for the general public, or policymakers in government).  This is because the key information is available to the large language model in the original document you provide. Proper translation of many documents (beyond say a washing machine operation manual) requires more than just a dictionary – it needs understandings of the cultural nuances. Remember that most generative AI models were first trained using predominantly English words and so may be much poorer at translating content into other languages, especially if it is difficult to find large volumes of text in that language online.

When the work requires quantity

If the task requires a process to be repeated again and again, then generative AI may be able to assist you. This works best when you use prompt engineering techniques known as “few-shot” – where you provide the AI with a series of examples of what you expect it to do. The completed task still requires a degree of human oversight to check for correctness, but this is often a quicker process than doing it completely manually.

Work where you want a second opinion 

The 24/7 nature of generative AI means that you can get instant feedback on your work instead of waiting for someone else to read your work and give you constructive feedback. This can be extremely useful and powerful; however, if you wish to do this, you should carefully consider how you construct your prompts to achieve your aim of getting useful feedback. Generative AIs can at times be quite sycophantic, and this means that when you ask for feedback, they can sometimes be overly positive. To overcome this, you can include in your prompts for the AI to be critical or argumentative, or plead for it to be brutal and truthful; this should provide you with more critical feedback that, while not as nice to read, may be more useful in improving your work.

Note that generative AI does not understand your work, or your questions in the way a human being does. As such it will not be able to assess work , provide feedback and award scores that match those from your (human) lecturers, because it does not have their knowledge of the subject, insight and true understanding of the marking process.

When not to use AI

When the effort is part of the process

AI can be helpful to learn new topics when it is used in an appropriate manner (e.g. Gregory et al. 2024); however, there is a growing body of evidence that generative AI can be detrimental to the learning process when it is used poorly because it leads to cognitive offloading. This is when an individual delegates too much to AI tools, and it reduces their opportunities to practice and develop their own critical thinking skills.  Critical thinking requires active engagement from the learner to evaluate and accurately analyse information.

A simple example of this would be getting generative AI to summarise an academic article or chapter; while it is likely that the AI will provide a good summary of the text because the learner has not taken the opportunity to actively engage with the text themselves, they have missed a chance to spend time thinking about the text with their own critical faculties.  A longer-term concern of educators is that this over-reliance on generative AI tools will lead to the atrophying of these critical thinking skills. 

Misia Temler, a researcher at the University of Sydney, has written about the risk of using generative AI to do the thinking for you.

When the penalty for error is high

Generative AI makes errors; this fact is not disputed, and sometimes the penalties for those errors, if they are spotted by others, can be incredibly high. There have been plenty of news stories from both academia and the wider world of people using generative AI in appropriate manners that have led to consequences such as academic failure, job loss, and /or severe reputational damage. So one question you should ask yourself when using generative AI for a given task is “ what are the consequences for generative AI getting this wrong and is it worth it?”

Agentic AI

A recent development has been the use of generative AI to perform a series of tasks – so-called agentic AI. Amazon provide a nice explanation of AI Agents. This can either be carried out using code configured in an AI tool (e.g. creating or deploying a ChatGPT Agent), using code running directly on your computer or those built into so-called agentic browsers, such as Perplexity’s Comet or OpenAI’s Operator.

The web is awash with videos showing people using these tools to log into systems, manage their email, book holidays, add out of office messages to their calendar and pick the best hotel. In order to perform these tasks, the browser needs to be able to read your email, purchase things using your credit card, etc. Take a minute to look around. Would you feel happy giving the person next to you your credit card and PIN, your unlocked phone and all your passwords? Then why would you do this to some code?

There are also several cases (case 1, case 2) where security flaws have been found that allow hackers to exploit weaknesses stemming from the fact that the agent is effectively running as a trusted user – as you. Seemingly innocent tasks such as asking an agentic browser to summarise a web page that you are reading can be derailed if the page contains hidden commands, which task the agent with doing something else. Whilst these specific issues are likely to get fixed soon, others will be found and exploited. This technology is still very early in its development and we advise great caution before using it.