Using generative AI when writing academic texts within doctoral education

Many advanced text-generating AI tools have been launched over the last couple of years, such as ChatGPT/GPT4/GPT-5, Copilot, Perplexity, Claude and Gemini. In addition, many search engines have implemented AI-summaries before the search results are presented. These AI tools may be helpful, but you need to use them responsibly.

When writing for example a research article, an examination task, the literature review for your half-time report or the comprehensive summary of your doctoral thesis (also called the “kappa”), you may usually use generative AI (GenAI) to assist you, but if you do, you need in most cases describe in a transparent way how you have used GenAI.

In this text, ”AI tools” and ”GenAI tools” mainly refer to Large Language Models (LLMs).

Can you trust GenAI tools to assist you?

First of all, you need to ask yourself: can I trust GenAI tools to assist in my writing? Most of us have probably heard by now that AI tools sometimes provide incorrect information, while at the same time sounding confident. The reason is that GenAI tools used for text creation are built on large language models that produce text based on statistical probability of language. As a consequence, the information can contain factual inaccuracies and biased information and common misconceptions. In addition, it is unclear what data GenAI tools are based upon, which further undermines the credibility of the information they present. Read more on how GenAI tools work.

Some AI tools will provide references, which might make them appear more reliable. However, it is important to be aware that the sources may have been summarized incorrectly and that summaries of scientific articles are often based solely on the article's abstract, since many AI tools cannot access information behind paywalls. It is also often unclear on what basis these sources were selected. Depending on where these AI tools retrieve sources, even questionable journals (often referred to as “predators”) may be included. Additionally, GenAI tools may also summarize the information from sources incorrectly, and the summaries often rely only on the abstracts of articles since many AI tools cannot access information protected by paywalls.

You should not use GenA-tools as sources. The absolute minimum may be to fact-check everything written by AI tools conscientiously and with reliable sources – and add these sources to your text. But please be cautious: if you only fact-check the information provided by generative AI tools, you will only look for information that supports the ideas in your AI-generated text – and you will likely end up with a skewed, biased text.

Moreover, it is clearly stated that both the content and the text of your comprehensive summary should be your own, so this is yet another reason that you cannot rely too much on GenAI tools.

Finally, you need to be aware that GenAI tools may store the data that you feed them. However, doctoral students who are employed by or affiliated with KI may use Microsoft Copilot, which ensures that your data is protected when you are logged in through KI. (Please be aware that Microsoft Copilot is not the same as Microsoft 365 Copilot – the latter requires a paid license).

How can you use generative AI tools?

Despite the limitations of GenAI tools and the requirement that your half-time report, comprehensive summary of your thesis (the “kappa”) or monograph thesis should be written by you, GenAI tools can be used both to find sources and to assist you in the writing process. 

Certain GenAI tools may be useful for searching for literature, such as Asta, Elicit, Scispace, Perplexity, and ResearchRabbit. Unlike traditional databases, these tools can interpret natural language queries, similar to google searches. While using these tools may serve as a useful starting point, it is crucial to recognize these tools’ above mentioned limitations in reproducibility and transparency. To ensure a more comprehensive literature search, you should in many cases complement the AI tool search with a search in traditional databases. Many traditional databases have also started to incorporate AI-powered features. One example is Web of Science Smart Search. Read more on AI and searching. 

GenAI tools may indeed help you with your writing. There is a growing recommendation to use generative AI primarily in the later stages of the writing process, as recent research indicates that creativity and learning may be adversely affected when generative AI is used at the start of your writing process. When you yourself have determined what you would like to write about and have written a first draft generative AI can be used to provide feedback on your writing, or help you improve the structure, language, grammar, and flow of your text. Using AI tools in this way may allow you to write a well-written text that is still your own work.

Since doctoral education is precisely an education you need to constantly reflect on how your use of AI affects your learning. Do you think you learn as much when you use AI as you would have done if you hadn't?  The “kappa” in your compilation thesis or, where applicable, your monograph thesis for your doctoral degree is also crucial in demonstrating that you have achieved the intend learning outcomes for your doctoral education. So, always consider whether your use of AI feels reasonable—and more specifically, whether your use of AI allows you to achieve the learning objectives for a doctoral degree. 

In relation to your learning when using AI, it is good to be aware that research and opinion-forming are ongoing regarding the possible effects on the brain of delegating cognitively demanding parts of writing and other higher executive functions to GenAI tools at the expense of stimulating cognitive processes and critical thinking during the writing process.

Recent research suggests that we tend to overestimate both our learning and our creativity when using AI, so it is important to take extra care to ensure your own learning and thereby gain the maximum benefit from your doctoral education. One way to reflect more deeply on your learning is to think of learning as “friction” (see Against Frictionless AI). This perspective reminds us that the difficulty involved in learning and writing, the “friction” they entail, actually helps us develop deeper understanding and knowledge. Generative AI may disrupt this process by reducing the friction that would otherwise arise during learning and writing. A useful question to consider is: when you use generative AI, do you use it to avoid difficulty and reduce friction in learning and writing, or, on the contrary, do you use AI to create and stimulate productive friction?

Both your learning and any benefits you derive from AI tools are also influenced by how you use AI and how you prompt AI tools (i.e., how you formulate your “question” to the tool). For example, if you formulate a general prompt—where you ask an AI tool to “improve” your text—it may result in various changes, such as changes in word choice and tone, or the addition of new content. These changes may also be based on a large amount of text that is not medical at all. For example, even if you write about “vaccine effectiveness,” a GenAI tool may change this to “vaccine efficacy”. The words "effectiveness" and “efficacy” have similar but distinct meanings in everyday language, but in a medical context, the distinction becomes crucial. When you additionaly receive many changes at once, it can be difficult to take a position on all of them, and it is easy to miss errors, especially if you cut and paste suggestions from AI tools into your own text. Furthermore, it is questionable how much you learn from receiving a rewritten text.

However, if you instead prompt GenAI to suggest adjustments to your text rather than changes, you have the opportunity to both make better decisions and learn more. This is because you must actively consider all suggestions. You can even ask for two or more alternative ways to structure your text, providing additional opportunities to reflect on your own needs and approach. It may also be a good idea to prompt AI tools to provide explanations for the suggestions they make. This allows you to make better choices and also learn more about text and how to make it clear and effective. In addition, you will get more out of AI tools if you prompt them to be specific in their suggestions and explanations, for example by focusing on structure, flow, or correct and formal language. Also, provide as much context as possible to get relevant feedback, such as the context you are writing for.

Declaring your use of AI tools

You also need to be transparent about how generative AI and AI-assisted technologies have been used during the process of writing your text for the half-time report, the comprehensive summary (“kappan”) in the compilation thesis, and the monograph thesis. Many scientific journals also request an AI declaration when a manuscript is submitted to them for review, and course organisers may request an AI declaration from students in connection with the submission of their examination assignments. The KI instructions/template for writing the comprehensive summary and for the monograph thesis require you to add a statement in the text regarding the use of generative AI. The statement must fully disclose your use of such GenAI tools, both in terms of which tools you have used and how you have used them. For example, if you have used AI tools to generate ideas for improving the content of your text, to ask for feedback regarding the structure of the text, or to ask an AI tool to re-write it, you need to declare these specific uses (and, of course, doublecheck everything, ensure that you can be accountable for everything in the text, and that you can explain and justify the content yourself.

However, if you have only used AI tools as for example language and grammar support, you do not have to add such a statement. But what does that mean? Generally, you do not need to disclose that you used tools that automate time-consuming tasks where the end result essentially remains the same. For example, you may use reference management systems such as Endnote, Zotero, or Mendeley to provide your sources in the exact format you need them. Similarly, you may use word processing programs that help you with the spelling, grammar, level of style, and concision of your text. These suggestions are based on grammatical rules and stylistic principles, and these tools will not re-write your text for you. Examples of such programs are Word, the basic feature of Grammarly, and Instatext. Of course, you are still responsible for the output, so make sure that you check it carefully.

If you are uncertain whether or not you should declare your use of AI tools, we suggest that you discuss the matter with your supervisor. Always err on the side of caution; it is safer to declare generative AI use when it may not be needed than to withhold that declaration when it is required.

A few final words 

Finally, we need to remember that advanced AI tools are still relatively new to us and that they now can do things they could not do, up until recently – so we do not yet have all the answers about how to use them responsibly without undermining our own learning. It is important to be humble about that fact, to perhaps be a bit cautious, to communicate with supervisors and peers with open minds, to be as transparent as we can, and to learn together as we move along.

Further exploration 

To explore generative AI and learning in greater depth, see the selection of articles below:

Popular science articles that include several interesting sources for further reading:

Additional resources

If you want to learn more about generative AI tools and how you may use them, you can visit “AI for students”. If you have questions, please contact Karolinska Institutet University Library.

This page was developed by Anna Borgström (Writing Instructor) and Lovisa Liljegren (Librarian) at the Karolinska Institutet University Library, in consultation with the central Director of Studies.