I built an AI thesis tool after ChatGPT invented 3 fake citations
Posted on dev.to by Dom, founder of Academly.ai
I finished my Bachelor's degree at 35.
Married, one small kid, full-time job in IT. I still remember sitting at my desk at midnight, staring at a blank page, completely blocked. No idea how to structure my theoretical framework. No idea how to find the red thread in 40 PDFs of research material.
That was before AI tools existed. I struggled through it manually.
Fast forward a few years. I started helping a friend with her Master's thesis. She was using ChatGPT to speed up her literature review. Smart, I thought. Until she submitted a chapter to her supervisor.
Three citations were completely invented.
Not paraphrased incorrectly. Not cited from the wrong page. Invented from scratch. Plausible-sounding authors, realistic DOIs, publications that simply did not exist. Her supervisor flagged all three in the first paragraph.
That was the moment I decided to build something different.
The Problem With AI and Academic Citations
Here's what happens when you ask ChatGPT to help with a theory chapter:
The model knows a lot about Maslach's Burnout Inventory. It knows Mayring's qualitative content analysis. It knows the general landscape of organizational psychology research.
But it doesn't know your sources. It doesn't have access to the specific edition of the textbook your professor assigned. It doesn't know which papers your university library has licensed. It doesn't know what page 99 of that 2009 Wickboldt paper actually says.
So it guesses. And it's very convincing when it guesses.
The technical term is hallucination. For a student submitting a thesis, the practical term is academic misconduct - even when it's unintentional.
What I Built
I spent several months building Academly.ai - an AI thesis writing platform with one core constraint:
It only cites sources you upload.
The workflow looks like this:
- Student uploads their research PDFs to a document library
- Selects which documents to use for a specific chapter
- Selects page ranges (we cap at 15 pages per generation to keep output focused)
- Enters their research question or chapter topic
- AI generates a structured chapter draft
Every citation in the output comes directly from the uploaded PDFs. Every page number is extracted from the actual document. If the source isn't in your library, it doesn't appear in the output.
No hallucinations. Not because the AI is smarter - but because the constraint is built into the architecture.
What Else Is In There
Beyond the core Theory & Background builder, the platform has grown to include:
Qualitative Analysis (Mayring workflow)
This was the most requested feature from early beta users. Philipp Mayring's qualitative content analysis is the dominant methodology for social science theses in German-speaking universities. We built a full workflow: upload interview transcripts, build a codebook, code segments, run cross-case analysis, generate findings. No other AI tool I'm aware of has this.
Topic Explorer
Helps students who don't know where to start. Enter a rough idea, get 3 structured topic suggestions with research questions, methodology recommendations, and real literature references. Then generate a full thesis outline and literature keywords.
Methods & Approach Builder
Structures methodology sections. Helps students justify their research design choices academically - which is usually the chapter they're least confident about.
AI Thesis Supervisor
A chat interface that has access to the student's uploaded documents and generated content. Students can ask questions about their own thesis: "Does my methodology align with my research questions?" "What's missing from my theory chapter?" Available at 2am when their actual supervisor isn't.
Discussion & Conclusion Generator
Takes the student's findings and theoretical framework and generates discussion and conclusion chapters. Connects results back to the original research questions.
The Honest Limitations
I want to be upfront about what the tool doesn't do well yet:
Quantitative analysis is limited. The platform handles qualitative research (Mayring, grounded theory, content analysis) much better than quantitative statistics. We're working on an SPSS/R integration but it's not there yet.
Citation verification is still on the student. We generate citations from uploaded text, but the student needs to verify that the extracted text matches the actual source. PDF extraction isn't perfect, especially for scanned documents.
It doesn't write your thesis for you. This is intentional, but it's also a limitation for people who want a one-click solution. The output is a structured draft that the student is expected to rewrite in their own voice. We're explicit about this in the UI: "Use this as your red thread, not as your final submission."
What I've Learned Building in Public
A few things surprised me during development:
The Mayring feature unlocked a market segment I hadn't planned for. Qualitative researchers in the DACH region (Germany, Austria, Switzerland) were actively looking for this. The qualitative analysis module has higher engagement than any other feature.
Students are more worried about getting caught than about learning. The most common question in feedback sessions was "will my professor know I used this?" - not "will this help me understand my sources better?" Designing for the second question while acknowledging the first was a constant tension.
The "no hallucinations" framing is less useful than I thought. Students don't search for "AI thesis no hallucinations." They search for "how to write my theory chapter" and "Bachelorarbeit schreiben." The technical constraint matters to me as the builder. To students, it just needs to work reliably.
Where It Is Now
Academly launched in beta in May 2026. It's free to try - new accounts get 2,000 Student Points, enough to generate several full chapters and explore all features.
If you're building tools for students or researchers, I'd love to compare notes. The space is interesting - high stakes for users, technically constrained in ways that make it genuinely challenging, and full of people who are genuinely stressed and need reliable help.
If you know students writing their thesis right now, especially in German-speaking universities, I'd appreciate if you pointed them to academly.ai.
And if you've built something in the academic writing space - or tried to and hit walls - I'm very curious what you found.
Dom is the founder of Academly.ai, an AI thesis writing platform for Bachelor and Master students. Built with React, Supabase, and OpenAI. Free beta at academly.ai.
Tags: webdev ai startup opensource