I identify several uses of AI in my academic research, a few of many. I think the use case and measures taken to reduce the sloppiness depend on the ratio of user-provided vs. AI-generated content in the context.

As a necessary side note, please note that during LLM inference, also known as in-context learning, including agentic AI and so-called retrieval-augmented generation (RAG), no model training takes place. As such, the model remains static; what changes is the context. The context is a single thread of content that the LLM can attend to. This context is marked up to delimit user-provided and AI-generated parts. This markup is not absolute and only works as guidance for the model attention. In agentic use, multiple LLM calls are made under the hood, and so the context may differ between calls. The context is influenced by tool calls, which is a special syntax that allows the LLM to send structured queries to prespecified utilities (like web search or file reading), which retrieve outputs that augment the context (hence retrieval-augmented generation).

1. Agentic RAG as a search engine “on steroids”

Send AI conversational queries that would be difficult to reformulate as a combo of key words. The human-provided part is relatively small while the AI-generated part is large; therefore, this is an “AI slop saturated” use case. If use of the AI-generated content is limited to conceptual use, that is, AI generates information leads that the user follows to review the content and make their own judgements (i.e., in the same way they would probably use conventional Google search), most of the slop is discarded. Chatbots may be a nice interface for this, and original raw chats may easily be exported and shared as a data trace.

2. Agentic RAG as a spellchecker “on steroids”

This use case makes sense if the amount of human-provided content is relatively large. After all, I did not need AI to write this post because it is rather short and ad hoc. By contrast, when doing research I have a tendency to write a lot of lengthy, but rough, “field notes.” They pile up together with various raw materials that I collect for traceability. Agentic AI, such as GPT 5 scaffolded into OpenAI’s Codex coding agent, may be instructed to use this rich context and shallowly rewrite a rough paper outline into a coherent academic text. In this process, text-based citations may be replaced with programmatic ones (e.g., BibTeX), which can in turn be used to render a pretty-printed Word document (e.g., using Quarto).

I argue that this use case is valid and preserves academic integrity, and I would like to make the following points:

• “Shallowly rewriting” someone else’s text to preserve the meaning but change the looks would be considered plagiarism; I conclude, therefore, the original authorship is preserved in the rewritten text. Self-plagiarism is a thing; however, this does not apply if the source material is unpublished notes that are ultimately published together with the final manuscript as a data supplement.
• Researchers are not supposed to rewrite statistical packages upon each use; they use prebuilt ones. Likewise, researchers use PowerPoint templates to draw slides or diagrams instead of doing it by hand. I believe that text, now that we have LLMs, can also be conceptualized as a mere wrapper. This realization is already happening in the programming world, where a prominent viewpoint is that the final code that powers software applications may not be as important anymore (“vibe coding” -> “agentic engineering”) because frontier LLMs can construct it quite faithfully from a source freestyle textual specification. The same is true for academic texts. However, line-by-line scrutiny remains very relevant because, unlike software code, an end-to-end test harness may not always be easily devised for academic prose.

The researcher may opt to break down the source rough outline and the AI-rewritten document programmatically into lines and task AI with establishing explicit cross-correspondence until both pools of lines are exhausted; this gives a convenient means of verifying equivalence for the author and a provenance trace for reporting.

Ultimately, the people who claim authorship are responsible for every line of the manuscript. To make this visible, I instruct AI to only work in a dedicated folder (which, named aislop, serves me as a reminder) from which I then manually copy vetted content to the main source file. Psychologically, this helps procedurally establish accountability.

What is easy to lose in this approach is my own personal writing style, yet I believe that this is more of an artistic matter. Like with pictures, photos, or music, people who enjoy the process as art still opt to do it manually; otherwise they are free to use AI to generate it. I think the same goes for text.

« Previous | Next »

---

Written by Pavel Zhelnov on May 13, 2026. Last revised May 25, 2026.