AI Usage Cards vs Data Statements

Data statements explain the dataset. AI Usage Cards explain the assistance.

Researchers often reach for the same word when they mean different things: disclosure.

A data statement tells readers what kind of dataset you used or created. It helps them judge population fit, language variety, annotation context, bias risks, and claims about generalization.

An AI Usage Card tells readers how you used an AI tool while doing the work. It helps them judge writing assistance, code generation, search support, summarization, image generation, translation, data analysis help, or other machine help that shaped the paper.

You may need one. You may need both.

If your paper uses a language dataset, a data statement can help readers understand the data. If your paper also uses ChatGPT, Claude, Gemini, Copilot, Perplexity, or another AI tool during the project, an AI Usage Card can help readers understand the assistance.

Those records answer different questions. Mixing them creates muddy disclosure.

The short version

Use a data statement when readers need to understand the data itself.

Use an AI Usage Card when readers need to understand how AI helped you work.

Use both when AI touched the dataset process or the manuscript process.

That last case has become common. A team may build an NLP dataset, use an LLM to draft annotation guidelines, ask Copilot to write preprocessing scripts, and ask ChatGPT to polish the final paper. A data statement can describe the dataset. An AI Usage Card can describe the AI assistance.

If you already know you need an [[AI disclosure](/ai-disclosure-in-systematic-reviews-and-meta-analyses/)](/how-to-disclose-ai-use-for-neurips-icml-and-acl-submissions/), you can generate an AI Usage Card before submission and paste the text into your acknowledgments, methods, appendix, or cover letter.

For a wider map of related records, see AI documentation frameworks compared. If your project uses a dataset outside NLP, you may also want [AI Usage Cards vs Datasheets for Datasets](/ai-usage-cards-vs-datasheets/).

What a data statement does

Bender and Friedman introduced data statements for natural language processing in a 2018 TACL paper. They describe a data statement as a way to characterize a dataset so that developers and readers can judge how experimental results may generalize, where software may fit, and what biases a system may inherit from data. (aclanthology.org)

That focus matters.

A data statement does not tell readers whether you used ChatGPT to improve paragraph flow. It tells them who, what, where, and under what conditions the language data represents.

The University of Washington Tech Policy Lab lists schema elements for data statements that include curation rationale, source dataset documentation, language varieties, language user demographics, annotator demographics, linguistic situation, preprocessing, capture quality, limitations, ethical review, distribution, and maintenance. (techpolicylab.uw.edu)

That list feels long because data carries hidden context. "English tweets" tells a reader little. English from whom? Posted when? In what region? Written for whom? Annotated by whom? Filtered how?

A data statement tries to make those choices visible.

What an AI Usage Card does

An AI Usage Card records how a researcher used AI assistance.

It can include the tool name, provider, date, model if known, task, part of the project affected, human review steps, and whether the AI output entered the paper, code, data, figure, or review process.

This format fits the disclosure problem that many researchers now face: journal policies ask them to explain AI use, but submission systems often leave little room for detail.

An AI Usage Card gives you a stable record before you forget what happened.

If you used AI across several parts of a project, do not rely on memory two months later. Keep a running note with a disclosure log template and turn it into a card when you prepare the manuscript.

The card can sit in an appendix. It can also supply text for an acknowledgments section or a journal form. See AI Usage Cards examples and templates if you need wording.

Why researchers confuse the two

Data statements and AI Usage Cards look similar from a distance.

Both support transparency. Both can appear near a methods section. Both help readers judge limits. Both ask researchers to write down decisions that would otherwise disappear.

But they start from different objects.

A data statement starts with the dataset.

An AI Usage Card starts with the tool use.

If you treat an AI Usage Card as a data statement, you may fail to describe the dataset well enough. If you treat a data statement as an AI disclosure, you may hide relevant AI assistance in a paragraph about sources and sampling.

Editors notice that gap. So do reviewers.

The fix is simple: name the object you document.

Side-by-side comparison

Question	Data statement	AI Usage Card
Main object	Dataset	AI assistance
Common fields	Curation rationale, data sources, language variety, demographics, annotators, preprocessing, limitations	Tool, model, date, task, prompt type, project stage, human review, output use
Best fit	NLP datasets and other projects where data population matters	Any paper, thesis, review, grant, or peer review that used AI tools
Reader benefit	Shows what the data represents and where claims may not generalize	Shows what the AI tool did and what the human authors checked
Common location	Dataset documentation, methods, appendix, repository	Acknowledgments, methods, appendix, cover letter, supplementary material
Main risk if missing	Readers overgeneralize from narrow or biased data	Readers cannot assess AI influence, responsibility, or policy compliance

This table hides one hard truth: disclosure takes time when you wait until the end.

Record the details while you work. Future you will thank past you.

When you need a data statement

Use a data statement when your dataset affects the scientific claim.

This applies most clearly in NLP. If you train, test, annotate, translate, transcribe, classify, or evaluate language data, readers need to know what kind of language the dataset contains.

A sentiment model trained on product reviews does not tell us much about clinical notes. A hate speech classifier trained on one platform may fail on another. A speech dataset from adults may not fit child speech. These are not minor details. They shape the claim.

If you publish a new dataset, include a data statement in the dataset paper or repository. If you use an existing dataset, cite its data statement when one exists and add missing context when your paper depends on it.

For field-specific guidance, read AI disclosure for NLP research papers. NLP researchers often need dataset documentation and AI use disclosure in the same submission.

When you need an AI Usage Card

Use an AI Usage Card when AI helped with the research or the writing.

That includes more than text generation. AI use can enter a project through:

manuscript drafting or editing
literature search and summarization
code generation
statistical analysis suggestions
translation
figure generation
image editing
data cleaning
annotation support
peer review support

Journal policies vary, but major publication ethics groups tell authors to disclose AI tool use and keep humans accountable. ICMJE says journals should require authors to disclose AI-assisted technologies used in submitted work and to describe how they used them. ICMJE also says chatbots and other AI tools should not appear as authors because they cannot take responsibility for accuracy, integrity, and originality. (icmje.org)

COPE takes the same line on authorship and adds that authors who use AI tools in writing, images, data collection, or data analysis must disclose how and which tool they used. COPE also states that authors remain responsible for manuscript content, including AI-produced parts. (publicationethics.org)

An AI Usage Card helps you meet that expectation without writing a vague sentence such as "AI was used."

For journal-level guidance, see Do I need to disclose AI usage in my paper? and AI transparency requirements for journal submissions.

When you need both

Many projects need both records.

Imagine a paper that releases a dataset of classroom dialogue and uses an LLM to help draft annotation rules. The data statement should describe the classroom setting, speaker groups, consent limits, transcription rules, language varieties, annotator background, and dataset limits.

The AI Usage Card should describe the LLM use: tool name, date, task, prompts at a high level, what the authors accepted, and how they checked the output.

Or take a computational social science paper that uses a dataset of public comments and asks Copilot to write scraping and cleaning scripts. The data statement can describe collection scope and sampling limits. The AI Usage Card can explain AI-assisted code generation and human validation.

One record cannot do both jobs well.

If your work also involves discipline-specific risks, pair this comparison with AI disclosure for social science research or AI disclosure for qualitative research.

A simple decision rule

Ask two questions before submission.

First: "Does the reader need more detail about the dataset to judge the claim?"

If yes, write or cite a data statement.

Second: "Did an AI tool help produce, revise, analyze, search, translate, code, annotate, or present any part of the work?"

If yes, generate an AI Usage Card.

If you answer yes to both questions, include both. Place them where readers will find them. Do not bury the AI use in a footnote if the tool shaped the work.

Example wording for a manuscript

You can use short wording in the main paper and place the full records in an appendix.

\section*{AI use and dataset documentation}
 
We provide two transparency records for this study. 
Appendix A contains a data statement describing the corpus, including 
curation rationale, language variety, speaker context, annotation process, 
and known limitations.
 
Appendix B contains an AI Usage Card. During manuscript preparation, 
the authors used ChatGPT on 2026-09-18 to improve the clarity of selected 
paragraphs and used GitHub Copilot during preprocessing script drafting. 
The authors reviewed, edited, and tested all AI-assisted text and code.

If your journal prefers acknowledgments, shorten it:

\section*{Acknowledgments}
 
The authors used ChatGPT to improve the clarity of selected manuscript 
paragraphs and GitHub Copilot to draft parts of preprocessing scripts. 
The authors reviewed and edited all AI-assisted outputs. A full AI Usage 
Card appears in the supplementary material. A separate data statement 
describes the corpus and annotation process.

If you work in Overleaf, see How to use AI Usage Cards in Overleaf and the LaTeX tutorial for AI Usage Cards.

What not to put in a data statement

Do not turn a data statement into a general confession box.

Keep the dataset record focused. Describe the dataset, not every software tool in the project. If AI helped you annotate, translate, filter, or label the data, mention that in the data statement because it affects the dataset. Then describe the tool use in the AI Usage Card with more detail.

Do not write "the dataset is diverse" without saying what that means. Name the groups, sources, language varieties, time period, platform, genre, annotation process, and known gaps where you can.

Do not claim that a data statement removes bias. It does not. It helps readers see what the data contains and what it may miss.

What not to put in an AI Usage Card

Do not use an AI Usage Card to decorate a paper with generic ethics language.

Readers need facts. They need to know which tool helped, what it did, where it affected the work, and what you checked.

Avoid vague lines like this:

AI tools were used to support this research.

That sentence creates more questions than it answers.

Write this instead:

The authors used Claude Sonnet 4.5 on 2026-09-21 to suggest revisions 
to the introduction and discussion. The authors accepted some phrasing 
changes after review. The tool did not generate data, results, citations, 
or statistical interpretations.

Better yet, generate a card and attach it.

How AI Usage Cards fit with other documentation

AI Usage Cards sit beside other records. They do not replace them.

A Model Card documents a model. A Datasheet for Datasets documents a dataset. A System Card documents a deployed system. A data statement documents a language dataset in social and linguistic context. An AI Usage Card documents AI assistance in research practice.

If that sounds like a lot, start with the record that matches the object under review.

Model under review? Use model documentation.

Dataset under review? Use data documentation.

AI assistance in the work? Use an AI Usage Card.

For nearby comparisons, see AI Usage Cards vs Model Cards and AI Usage Cards vs System Cards.

A practical workflow

Do not wait until submission week.

Create a short disclosure note when the project starts. Add entries when you use AI tools. If you create or reuse a language dataset, start the data statement when you choose the data, not after you run experiments.

Your working notes might look like this:

\subsection*{Transparency notes}
 
Dataset record:
Corpus: public forum posts collected between 2024-01-01 and 2024-03-31.
Language: English, mostly United States usage based on platform metadata.
Known limits: no private messages, no deleted posts, no age verification.
 
AI use record:
Tool: Microsoft Copilot, used 2026-08-14.
Task: draft regular expressions for preprocessing.
Human check: authors inspected outputs and tested scripts on 200 sampled rows.

Those rough notes can become polished documentation later. They do not need to look elegant. They need to exist.

Final check before submission

Before you submit, ask what a skeptical reviewer would want to know.

If they ask, "Whose language appears in this dataset?" your data statement should answer.

If they ask, "What did the AI tool do?" your AI Usage Card should answer.

If they ask, "Who takes responsibility?" your paper should answer with human author accountability.

That is the heart of good disclosure: readers can see enough to judge the work.

Generate your AI Usage Card at ai-cards.org, then place it in your acknowledgments, methods section, appendix, supplement, or cover letter. If your paper also depends on a language dataset, pair the card with a data statement so readers can understand both the data and the assistance.