# Advanced Generative AI for Thesis Writing - Who are you + your email? - Caroline Vogt Hansen + mullemiv@gmail.com Mátyás Varga matyi825@gmail.com Magnus Kobborg Henriksen + magnuskknudsen@gmail.com Ulrich Karstoft Have + ukh@cc.au.dk Media and Journalism Studies - What is your field of study? English Studies (elective in political science) - What is your primary use of Generative AI (if you've attended a course before, has your primary used changed)? - What are you expecting to get out of this course? # What have you been using GenAI for? ... Using to give critique: "Pretend you are a religious studies researcher from a university. You are familiar with all the literature and theories in religious studies, and you specialize in ghosts and absorption. You are currently supervising a master's student in writing his thesis, and in this connection, the student has sent you a text passage that he would like you to read through and provide constructive criticism and suggestions for analysis or possible discussion topics. In your conversation, he will also ask if you can come up with arguments for and against his points in order to address criticism and errors in the final repport, okay?" Using it for clarification purposes, specifically political science terms and theories. # Understanding how LLMs work. ## Exercise 1 (group exercise): ### 1.1. What is a misconception that you had about how LLMs work and how do you understand them now? Tokens are the whole word and not single signs. What is good and what is bad information i.e. from Reddit or a professor does not make sense for it. That they were sentient Not understanding terms like "LLM" and "generative AI" Thought it was using online material (thought it was up to date) ### 1.2. What dangers do they pose based on what you've learned about them? A few people biasing and dictating what information GAI can give. People not undertsanding how LLMs work and as a result don't critize the answers they get --> blindly accepting information is never good That this course is not mandatory, but rather some tiny off-corner course offered in Brightspace with very few people attending; especially ctiticizable when AU is now allowing the usage of LLMs in theses and exams, but almost no students have proper knowledge (this course should be obligatory) Dyslexics: On its face, it would seem practical for dyslexics to use LLMs correcting their grammar and spelling. However, this also entails a certain human detachment from language (i.e. the vital tool, we rely on, to communicate with other humans). By outsourcing this to a machine, the dyslexic becomes worse at knowing what they have actually communicated. This is an extreme case, of course, but this could be the same a little bit for everyone else too, in terms of understandability of their text. In a way, this is why LLMs seem to require MORE effort in trying to understand the text, you write, otherwise you might lose critical competencies in building text and language in general. ### 1.3. What potentials do they have based on what you've learned about them? For automating really boring mechanic task like turning a reference list. ... # How to think about the problems ## Exercise 2: ### 2.1. What problems do I want to solve? - The problem with mindfullness: How can people use mindfullness to enter a state of resonance, thus leading a healthier and less stressful lifestyle (on an individual level), which increases their productivity and effectiveness in organizations, without using mindfullness as a commodity that just fuels the acceleration in society even further by overlooking potentially bad working conditions due to the fact that the individuals get better at enduring those? * Car? Should I buy a used old car that might need repairs rather than a new car for triple the price? I will be driving about 20.00 km. I need to understand how many years, if we think that the new car will not need new repairs within the first five years having the car (assuming I do not get into an accident), what would be the best choice. * Should I stop programming myself and let AI do it? I am positively surprised about how well AI can write code for many boring tasks I have on my table. * How should I choose a life partner? Can an AI help me find the best fit based on my personality and values? - Helping with research, finding valuable sources, help with the form/structure in the thesis - Writing programming code and/or finding and correcting mistakes in already existing code M. Download my readings to every class. ... ### 2.2. What are the specific ways in which I want to solve these problems? ... Car: I would like to have two mathematical models that take different things into account, such as starting price, how much it costs to have over the years, and maybe also the difference between having an electric car vs. a car running on gas. Also consider legislation fx in 2030 the politicians in Aarhus wants only electrical cars to drive in the center. M. To have an organised structure with different folders, like Class A, B, C, Lesson 1, 2, 3... Programming: Something to automatically do the work (or do it manually). Structure/research: N/A The problem with mindfullness: Subtract the formal expectations from the process of becoming mindfull, such as going to a course on a specific time, doing specific exercises that consume time or creating any type of goals or milestones that would "make one mindfull", as these might become stress factors themselves, thus elminating the purpose of mindfullness. SOLUTION for the work conditions: The solution should not be found in the mindfullness practice of individuals themselves, but rather in the structural framework of work environments. Unfortunately, this opens up so many more problems, such as the subquestion and -problem: "Should work conditions be negoiated on a political level or by the organizations themselves?" ### 2.3. Are these problems worth solving using AI? Car: I am not good at math, but there are many many variabels... so no idea but maybe it can give a rough sketch. It would also have to local/specific and updated knowledge of legislation (Danish and EU law) ... M. Probably not, but since there are several web pages to download from, it may have potential. The problem with mindfullness: As a quite complex, sociological problem, AI only seems able to solve parts of it, if any at all. Maybe it could be used to break down the problem in many subproblems, perhaps map it out, give suggestions for political implementations or just define terminology from new perspectives. # Using various prompt strategies ## Exercise 3: in groups look back at which problems you want to solve. ### 3.1. Can you solve some of those problems using zero-, one-, few-shot prompting? ... ### 3.2. Think of some prompts that you've used or want to try to solve those problems? ... # Using GenAI to optimize your thesis writing ## Exercise 4 ### Sequential Structure Take the role of a professional editor of academic writing. Based on the description of sequential structure, check the content to see where it deviates from that description: DESCRIPTION Organize your thesis in a logical sequence. Start with an introduction that outlines your research question, followed by a literature review, methodology, results, discussion, and conclusion. Each chapter should build on the previous one to create a coherent narrative. Respond with YES if you understand then I will provide the content. ### 4.1: What are some insights that you've gotten about your writing based on the sequential sequence feedback from the prompt? ... ## Consistency Take the role of a professional editor of academic writing. Based on the description of consistency, check the content to see where it deviates from that description: DESCRIPTION Maintain consistency in terminology, tone, and formatting throughout your thesis. This helps create a professional and readable document. Use the same terms for key concepts and adhere to the same citation style throughout. Respond with YES if you understand then I will provide the content. ### 4.2: What are some insights that you've gotten about your writing based on the consistency feedback from the prompt? It is not clever enough to understand what I want it to do. It changes nothing in the text when I feed it my text. Good at picking out formality and tone, but lacks explanation for WHY it marks thing as wrong. It gives clear results on what is wrong, but this is abtsract and we have no way to check whether it's conclusions are actually correct (for exampple saying "inconsistent uses of semicolons and commas") ## Relevance Take the role of a professional editor of academic writing. Based on the description of relevance, check the content to see where it deviates from that description: DESCRIPTION Ensure that every section and paragraph of your thesis directly contributes to your main research question or hypothesis. Avoid including irrelevant details that do not support your thesis statement or research objectives. Respond with YES if you understand then I will provide the content and the research question. ### 4.3: What are some insights that you've gotten about your writing based on the relevance feedback from the prompt? - Prompting the AI to critique the analytical insights of a random section of my thesis, I got some interesting outlines of areas where my thesis deviates from academic standards in terms of sentence structure, rationale in arguments, repetition of arguments, and it even detected some points, which the AI (from the perspective of a professional editor) thought were important but not outlined enough in my thesis. # Iterative Refinement ## Exercise 5 ### 5.1. Using iterative refinement, take an advanced concept from your field of study and refine it in such a way that you can explain it to someone from your group who isn’t in that field. Using brackets works quite well to prompt the AI to analyze a piece of text that is written outside of the brackets.