Ask ChatGPT to write a literature review paragraph on cognitive load. The result is usually fluent, convincing, and wrong. Not structurally wrong, but source-wise. The AI will calmly cite authors, years, and journals that never existed. This is the biggest risk of AI in academic writing, and ironically the least discussed.
The danger is not the sentence
When students use AI to write, two things usually happen. First, the text is good. Second, the citations are fabricated. The problem is that good text makes fabricated citations look credible. The supervisor reads it, finds the argument strong, then checks the bibliography and discovers five of ten sources do not exist. That is not a typo. It is how language models work: they predict text that looks correct, not text that is correct.
Our approach: an AI that refuses to invent
Publikasi is built on a simple rule. The AI may only write from references you upload. No references, no writing. When it drafts a chapter, every claim is traceable to a source text you selected. If there is no source, the AI does not write it.
This changes the relationship between writer and tool. You no longer trust the AI and verify afterward. You provide sources first, then the AI writes within those bounds. Verification happens at the start, not the end.
What you lose, what you gain
What you lose: the speed of inventing a full chapter from thin air in ten seconds. What you gain: writing you can defend in your defense. Every paragraph has a trail to the original document. When the examiner asks where a claim comes from, you know the answer.
Scholarly writing is ultimately about trust. Readers trust your writing because every claim is backed by a checkable source. A tool that invents sources breaks that contract. A tool that refuses to invent preserves it.
That is the entire reason we built Publikasi. Not an AI that writes faster, but an AI that writes honestly.