AI help for a thesis comes in four kinds. Generative AI produces new text, code or images on request. Corrective AI changes spelling, grammar and wording in your text. Detecting AI estimates whether a text came from a model. Evaluative AI reads your draft, comments on its content and adds no text to it.
If your AI statement clearly doesn't cover what you describe elsewhere in your thesis, Esmond points it out.
One product can do more than one of these. A chatbot can produce a paragraph and comment on a paragraph in the same session, and a grammar tool can add a feature that proposes new sentences. So decide task by task which row applies, based on what you ask the tool to do.
| Kind | What it does | What you get back | What you still have to settle |
|---|---|---|---|
| Generative | Produces new material on request | Text, code, images or calculations you didn't make yourself | Whether any of it may go into your thesis, and how you report it |
| Corrective | Changes spelling, grammar and wording in a text you supply | A corrected version of your own sentences | Where your school draws the line between basic corrections and changes to the content |
| Detecting | Estimates whether a text came from an AI model | A percentage or a flag | How much weight a flag carries, since detectors make mistakes |
| Evaluative | Reads your draft and comments on its content | Comments on your own document, and no new text for it | Whether your policy treats comments from an AI tool as AI use |
A hypothetical example of a commentYour introduction promises to compare two school districts, but your results chapter covers only one. Where did the second district go?
A first comment on one part of your own draft, right away and without an account.
Get a free previewDartmouth's graduate school gives a workable definition: “Generative AI refers to technologies that create content, such as text, images, videos, code, or data analyses, using complex models.” What comes back is material you didn't produce, which is why thesis policies pay so much attention to it. The policies compared on our university AI policies page mostly concern this kind.
Check everything such a tool gives you. A model can produce a reference that looks real and doesn't exist, or a figure that nobody measured. Open each source, compare each number with where it came from, and leave out what you can't find.
A corrective tool starts from your text and changes it. Some policies set basic corrections apart. In RIT's computing PhD program, students whose use was limited to “basic editing functions, such as spell check or grammar correction applied to their own written work” confirm that in one standard sentence in their appendix.
That won't hold everywhere, and the difference is hard to see in practice: a small correction can turn into a new sentence. Check what your program says before you let a tool change your text, and keep the version from before the changes.
A detector tries to answer one question: did a model produce this text? It doesn't comment on your content. MIT Sloan's teaching and learning team warns that such software is far from reliable: “it has high error rates and can lead instructors to falsely accuse students of misconduct”. Liberty University's guidance for faculty says which edits the AI detector it uses won't flag: “Spelling, grammar, punctuation, and mild corrections to content will not cause that content to be flagged by the AI detector because those features do not use generative text technology.”
Similarity checkers are a related kind of tool: they compare your text with existing sources. Neither one tells you whether your method fits your question. If a flag ever appears on your work, your drafts and notes are what you can show, so keep them.
Evaluative AI starts from a draft or a chapter. It reads the text and points to places that need a second look: a claim that goes further than the data, two chapters that contradict each other, a statement that looks wrong. The revising stays with you, and no sentence in your document changes.
Because it adds nothing to the text, it raises a different question from generative AI: may this draft go to an AI tool at this stage? Some policies cover any AI tool, and then evaluative AI falls under them too. Its comments can also be wrong, so treat each one as something to check.
Esmond evaluates every assignment before you submit. It never writes it.
Esmond is one example of the evaluative kind. It reads your draft from within your field and flags factual errors, inconsistencies and blind spots, with the exact place in your own PDF. It doesn't check spelling or grammar, and it doesn't detect AI or plagiarism.
Esmond is an AI tool that doesn't write for you. If your course doesn't allow any AI tools, don't use Esmond for that work. If you're unsure, ask your instructor or advisor.
AI that reads a text someone already produced and comments on its content. It adds no text to the document, and the author decides what to do with each comment.
It depends on the feature. Spell check and grammar correction change what is already there. A feature that proposes new sentences generates text, even inside the same product, so note which one you used.
No. A detector estimates where a text came from. It says nothing about your argument, your method or your sources.
That depends on your school, your program and your committee. Name the kind of tool and the task when you ask your advisor, and keep the answer.
Read on November 4, 2026.
Get your paper evaluated before you submit