SERIES · AI is making papers cheap to produce but they remain costly to referee. Proofs and Prompts is publishing a series of posts from 26 October on the future of publishing, papers and credit. Questions such as how to disclose AI use, who bears the cost of checking, what authorship means and what papers of the future might look like will be addressed. This post is part of the series. Each is written independently and reflects only its authors’ views. All posts in the series
This is a summary of an article written by me with the same title. This summary is presented without references for the sake of brevity. The full version of this article can be found on my personal website.
In August 2026, almost 25% of all mathematics articles uploaded to arXiv disclosed the use of Artificial Intelligence (AI) technologies. Among these, one finds varying levels of detail describing the human-AI-interaction and its impact on the article. Unsurprisingly, one study has found a significant discrepancy between expectations and the reality of AI disclosures in computer science, and many mathematicians likely share this feeling. For this reason, I want to initiate a standardisation of AI disclosures in mathematics. With this goal in mind, I surveyed existing guidelines, recommendations and the work of scholars on research ethics. The outcome of my endeavours is the development of three guiding principles, a taxonomy and a LaTeX package.
Why transparent disclosure is important
Commercial Large Language Models (LLMs) pose many ethical concerns for society and science in general, and for mathematics in particular. While detection tools for AI-generated text remain unreliable, a complete ban on the use of such technologies for mathematics is infeasible. Instead, I argue that academic journals, which are likely to experience a significant increase in submissions, should require maximally transparent AI disclosures going forward. In addition, many existing structures in academia fundamentally depend on the correct attribution of ideas. While it is possible to challenge the status quo, I assume here that it is desirable to uphold this tradition. Even without this assumption, transparent AI use disclosure is beneficial for approximating expended financial and environmental resources, thus providing insight into the sustainability of AI-assisted mathematics. Indeed, “transparency” and “reproducibility” are listed among the most important guiding principles of the UNESCO Recommendations on Open Science. These axioms of responsible research demand a maximally transparent AI disclosure. In fact, under current policies, it is considered academic misconduct to omit disclosure of AI tool use for mathematics research.
The guiding principles
In its simplest form, the human-AI-interaction discussed here may be described as a cycle
Human provides AI input -> AI produces output -> Human processes AI output -> Human provides AI input
which eventually terminates in the production of a mathematics research article. A transparent AI disclosure allows readers to correctly identify the role played by AI technologies throughout the entire process of creating the research manuscript.
Who should be author?
Many others have discussed the question of who should be the author on a (mathematics) research manuscript to which AI contributed. Authorship is primarily an indicator of credit for the progress made in the research article and most scholars on research ethics agree that credit and responsibility are two sides of the same coin. This implies that in order to receive credit, one must assume responsibility. However, as AI tools are not human they cannot be held responsible and may therefore not be authors of (mathematics) research manuscripts.
Principle 1: Only humans are able to assume responsibility for the quality of mathematics research and may therefore assume authorship of a research article in mathematics.
How specific should the disclosure be?
With the goal of achieving a maximally transparent AI disclosure, one must necessarily be as specific as possible. Here, I distinguish between three types of specificity: technological, functional and locational. Technological specificity demands the exact description of the AI tool’s name and any additional attributes used, for example GPT-5.6 Sol with max reasoning. Functional specificity is supposed to describe precisely which tasks of the research process were delegated to AI technologies. I developed the Mathematics AI Delegation Taxonomy (MAIDeT), see below, which is designed to provide a standardised dictionary for this purpose. For example, authors may delegate “Statement generation” or “Novelty assessment” to AI and should disclose this clearly. For locational specificity, consider a mathematics article as consisting of three levels: Indexed environments such as lemmas, definitions or theorems; sections; and the entirety of the manuscript. Locational specificity requires the authors to describe as precisely as possible where, for example in Lemma 3.5, in the manuscript a task was delegated to AI.
Principle 2: AI disclosure should be of the highest possible technological, functional and locational specificity.
Why should AI interactions be documented?
Even with the researchers’ input prompts available, readers are not able to reproduce any reported interactions with AI tools. This makes it of utmost importance to document the entirety of the human-AI-interaction for the sake of transparency and reproducibility. Depending on the technology used, this can be as easy as exporting a chat log as a PDF. The inclusion of this documentation can provide insight into token costs and help illuminate the question of sustainability of AI-assisted mathematics.
Principle 3: Every interaction with AI which has influenced the final presentation or content of a manuscript should be recorded in full and made available.
Visibility of AI-critical positions
Commercially available LLMs raise many ethical concerns both generally as well as for mathematics specifically. Together with the issues of transparency and reproducibility, these issues compile an excessive list of reasons for mathematicians to boycott the use of such technologies. For those holding this stance, I strongly recommend including in their research outputs a statement sharing their conviction, which enhances the visibility of AI-critical positions.
Taxonomy and LaTeX package
In order to implement Principle 2 in practice, I developed MAIDeT and a LaTeX package called mathaidisclose. The taxonomy draws on my own experience as a researcher in abstract algebra and many accounts of “doing mathematics” before and after the advent of LLMs. It breaks up five major themes, representing the non-linear research process of doing mathematics, into minor research tasks one may delegate to AI tools. It is important to emphasise that many mathematicians do not view mathematics as a theorem-production pipeline. While some support the philosophy of theorem-proving-above-all-else, viewing it as the one true way of doing mathematics is a serious misrepresentation of mathematics. The taxonomy can be found below:
Knowledge acquisition:
- Axiomatisation
- Example generation
- Literature search
- Pattern observation
Objective identification:
- Applicability assessment
- Conjecture generation
- Feasibility assessment
- Idea generation
- Novelty assessment
- Objective formulation
- Objective subdivision
- Research gap identification
Proof construction:
- Statement generation
- Proof completion
- Proof generation
- Proof verification
- Proof polishing
Code and computation:
- Algorithm conceptualisation
- Algorithm implementation
- Code generation
- Code optimisation
- Computation
- Computation verification
- Enumeration
- Formal verification
- Process automation
Presentation and communication:
- Context discussion
- Literature review
- Proof exposition
- Reformatting
- Spell-checking
- Text generation
- Translation
- Trend identification
- Visualisation
The LaTeX package mathaidisclose allows users to easily implement locational specificity. For example using the command \discloseAI[model]{task} inside of an indexed environment, such as Definition 3.1, will print the sentence: “The authors disclose the delegation of task to model in Definition 3.1” in the AI disclosure section generated using \printAIdisclosure[]. A similar command \globalAIuse[model]{task} is available to disclose the tasks delegated to AI throughout the entirety of the manuscript. For those disclosures which do not fit into the previous ways of disclosing AI use, the command \customAIdisclosure{text} can be used to print any text in the AI disclosure section. This command should also be used by authors to affirm that they have verified all AI-generated outputs included in the article for correctness and that they assume full responsibility for the contents of the manuscript. If none of the three commands for disclosing AI are used, then \printAIdisclosure[] outputs “There is no use of Artificial Intelligence to disclose.”
Received 10 October 2026.
Add to the discussion