How AI Research Gets Cited in Real Publications: A Practical Guide

Most researchers learn citation rules for books and journal articles. Then they try to cite a GPT-4 output, a Hugging Face model card, or a landmark neural architecture paper and the whole system breaks down. The formats were never designed for AI artifacts. What follows is inconsistent citations across publications, frustrated editors, and papers that fail peer review on a technicality that could have been avoided.

This guide goes deeper than the usual template lists. It covers the structural reasons AI citations are different, how the three major style guides handle them, what to do with models and datasets specifically, and the mistakes that appear most often in submitted manuscripts.

Why Citing AI Research Is Structurally Different from Citing Other Sources

Traditional citation exists to let readers trace a claim back to its source. With AI research, this breaks down at multiple levels simultaneously. A language model output has no fixed page number, no author in the conventional sense, and no version that stays stable over time. A dataset like ImageNet has gone through multiple releases. A framework like PyTorch receives regular updates that change behavior.

The deeper issue is that AI artifacts exist on a spectrum between tool, dataset, model, and publication. Each position on that spectrum requires a different citation logic. When you flatten all of them into a single format, you lose the information readers actually need to evaluate or reproduce your work.

Here is the core distinction most guides skip entirely:

• Citing an AI paper (such as the original Transformer paper by Vaswani et al.) is straightforward. It is a standard academic publication with authors, a venue, and a date.
• Citing an AI model (GPT-4, Claude 3, Gemini 1.5) is closer to citing a software tool. The "author" is an organization, the version matters enormously, and the artifact is not a document.
• Citing AI-generated content is a different act altogether. It is closer to citing a research assistant's contribution than a source document, and most style guides treat it accordingly.

Conflating these three categories is the single most common error in AI citation practice. Before choosing a format, identify which category you are actually dealing with.

APA, MLA, and Chicago: What Actually Changes for AI

The three dominant academic styles have each issued guidance on AI citations, but their approaches differ in meaningful ways. Understanding the structural logic of each matters more than memorizing templates.

APA 7th Edition

APA treats AI-generated text as a personal communication when the output cannot be retrieved by readers. If you used ChatGPT to help draft a section, current APA guidance suggests citing it in-text with a parenthetical noting the model, version, and date, but not including it in the reference list because the output is not recoverable by another reader.

For AI research papers, APA is standard. The attention mechanism paper, for example, would appear as:

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.

For an AI model cited as a software tool, APA uses its software citation format:

OpenAI. (2023). ChatGPT (Mar 14 version) [Large language model].

MLA 9th Edition

MLA is more flexible about containers and handles AI citations by treating the model as the author or creator, with the generating organization as the publisher. MLA also places significant emphasis on the access date, which matters more for AI tools than for most other source types because outputs can change between sessions.

For AI-generated content, MLA recommends describing the prompt used in the body of your text and including the tool in your Works Cited as a source you consulted, not as an author of a claim.

Chicago 17th Edition

Chicago distinguishes between notes-bibliography and author-date formats, and both handle AI sources similarly: the model is treated as a kind of software, the organization is the responsible party, and a note in the text should describe how the tool was used. Chicago style is particularly common in humanities work where AI-generated content raises additional questions about attribution that go beyond formatting.

Citing AI Models and Datasets: A Practical Step-by-Step

Models and datasets are the two artifact types researchers most frequently get wrong. Here is a practical approach that works across style guides and satisfies most peer reviewers and editors.

Step 1: Identify the exact version. "GPT-4" is not specific enough. If the model has a version identifier, a release date, or a model card on a platform like Hugging Face, use that information. Version matters because behavior changes between releases.

Step 2: Identify the responsible organization. For most commercial models, this is the company. For open-source models, it may be a research group or a university lab. This becomes your "author" in the citation.

Step 3: Locate the canonical reference. Many models have an associated technical report or paper. When one exists, cite that paper rather than the model directly. The GPT-4 technical report, the LLaMA papers, and the BERT paper are all citable academic documents.

Step 4: For datasets, cite the data paper if one exists. ImageNet, COCO, and most major benchmarks have associated publications. Citing the paper is more stable and more informative than citing a download URL.

Step 5: Record access dates for anything hosted online. Model cards, dataset repositories, and documentation pages can change. An access date protects you if a reviewer asks about a discrepancy.

The Mistakes That Appear Most Often in Submitted Manuscripts

After reviewing how citation errors appear in AI research submissions, a few patterns stand out consistently.

• Citing the interface instead of the model. Citing "ChatGPT" when you mean the underlying GPT-4 model, or citing a web application URL when a technical report exists, makes your reference less precise and harder to verify.
• Omitting the version entirely. A citation to "BERT" with no version, checkpoint, or paper reference gives readers almost no useful information. BERT has been fine-tuned and adapted into hundreds of variants.
• Treating AI-generated content as a citable source for factual claims. If you used a language model to generate a summary and then cited that summary as evidence, you have cited a tool's output as a source. Style guides and editors will flag this. AI-generated content can be cited as an example of model behavior; it cannot stand as evidence for an independent claim.
• Using a generic URL without a retrieval date. Linking to a model's homepage or a dataset landing page without noting when you accessed it leaves a gap that reviewers notice.
• Inconsistency within a single paper. Citing one model as a software tool and another as a personal communication in the same manuscript signals that the author has not thought through the logic. Pick an approach and apply it consistently.

A Quick Reference: Which Format to Use When

The table below summarizes the citation approach by artifact type across the three major style guides.

The most important takeaway is not which template to use but which category your source belongs to. Once that is clear, the formatting follows naturally from whichever style guide your publication requires. When in doubt, err toward more information rather than less: version numbers, access dates, and brief descriptions of how a tool was used all help readers and reviewers understand exactly what you are citing and why.