The Four Kinds of AI LaTeX Help, and When Each One Fails

There are four shapes of AI help for LaTeX, and most complaints about "AI being bad at LaTeX" are really about using the wrong one. They fail in different, predictable ways.

1. A chat window with your text pasted in

Good at: explaining an error message, drafting a block of LaTeX from scratch, answering a question about syntax.

Fails at: anything that depends on the rest of your document. It cannot see your preamble, so it hands you a \toprule without knowing whether booktabs is loaded, and the copy you pasted is stale the moment you keep working. Every edit comes back as text you merge by hand, which is where things get lost.

2. Autocomplete in the editor

Good at: finishing the line you were already writing. Environment names, reference keys, the closing half of a construct.

Fails at: anything you have to think about first. It cannot restructure, it cannot check whether the result builds, and on a bad day it confidently completes a citation key that does not exist, which compiles to a question mark you find later.

3. An agent inside the editor

Good at: the work spanning several files or several steps. Reformatting to a journal class, turning prose numbers into a table, splitting one file into chapters and fixing the paths, reading the log and repairing the first real error.

Fails at: knowing what is true. It will restructure your results section beautifully and has no idea whether the numbers are right. The guard is the review step: an edit that arrives as a diff you accept or reject is one you can refuse in a second, and the ones worth refusing are usually obvious the moment you see them side by side.

4. Your assistant, given a compiler

Good at: staying where you already work. If your day is in ChatGPT or Claude, a remote MCP server lets that assistant create the project, write the files, compile, and hand back a PDF, which removes the copy-paste loop entirely.

Fails at: being where the document is. You are working in a chat transcript, not in an editor with a PDF beside the source, so it suits generating and iterating more than it suits writing.

The one property worth checking

Whatever the shape: can it compile, and does it look at the result? That is the difference between a tool that has checked something and one that is confident. LaTeX is unusually well suited to this, because the answer is not a matter of taste. The build succeeded or it did not, the log says which, and a tool that ends its work with a real compile has done the one thing that makes its output trustworthy.

Ours is Lexi, in the editor, and the same tools are reachable from your own assistant over MCP. What you ask decides what you get back, which is covered in the prompts that actually change the document.