AI and Research Papers

2026 will be a year of AI disruption across all of academia. Let's start by talking about AI is changing how we write research papers. Not the research itself (another post), just about the dissemination thereof.

Technology has changed research distribution since the invention of paper, and the printing press, typewriters, journals and conferences have all had dramatic effects. But we've already seen such dramatic changes just within my research career from authors doing their own formatting (TeX/LaTeX), instant distribution of papers (email/web) and collaborative writing (email/shared repositories/Overleaf).

How AI affects research writing follows a trend not unlike other AI applications. 

The Slop Phase

This is where AI does more harm than good. Useful for finding grammar mistakes and making fancy LaTeX tables, but if you use it for writing it may make stuff up, create citations out of thin air and put in words like "as a large-language model" in the text. Things have gotten much better with the latest models, I'm currently a fan of Gemini 3, but we haven't fully left the slop phase and anything AI written needs close verification.

The Verification Phase

2026 will be the year we go from humans having to look over AI generated content to AI having to look over human-generated content. Already you should give your paper, ideally in LaTeX, to the best model you have access to with a prompt asking for a critical review, or use a specialized service. You don't have to follow what the AI says but you should be open to improving your paper. Google provided an optional tool for those preparing STOC 2026 submission that many found valuable. In 2026 some conferences and journals will start requiring submitted papers to go through an AI review as well as a human review (for now), especially with those overwhelmed by submissions.

At some point, we may require authors, with AI help, have their theorems verified by theorem provers like Lean or Rocq. It will be a while, especially in computational complexity, where it is hard to properly state a theorem in formal logic, let alone its proof. 

The Collaborative Phase

Not just AI critiquing papers, but researchers and AI working together in paper writing. We already see pieces of this, having AI fill in details of arguments, give a first draft of an introduction, help create diagrams, find relevant citations, etc. Right now anything the AI writes requires a careful lookover by the researcher. As these models improve, we'll see more and more generated by the AI and researchers starting to trust the output, maybe more than they should. This will lead to...

The Generation Phase

Here the AI writes the LaTeX from scratch based on proof sketch given by the researchers. As an experiment, I "vibe-coded" a minor research paper and it went better than expected. 

The researcher will go and ask for updates and touch up the LaTeX at the end, but at some point the human won't ever touch the LaTeX file, just a prompt or a markdown file that generates it. 

By the end of the 1980s, researchers just trusted the LaTeX for formatting, fixing only syntax errors and overfull boxes. Most researchers (but not all) stopped fiddling with the look of the paper. At some point, and at higher stakes, we'll do the same with the paper content as well.