Provenance in Human–AI Collaboration

2. The Dual Provenance Problem

In human–AI work, provenance operates on two levels:

1. External provenance—attributing the origin of ideas, data, and formulations to prior human or institutional sources.
2. Internal provenance—recording which elements emerged through AI synthesis, change, or generation during collaboration.

Both levels must be explicit. Under current legal and academic standards, AI systems do not hold copyright or moral rights in the human sense, but recording AI participation is essential for transparency. It establishes—who (or what) contributed what, under what interpretive context, and from which prior materials. Without this, the distinction between proper citation and AI-generated material becomes unclear.

3. The Reciprocity Standard

We propose a simple standard for AI-assisted authorship:

Every contribution should either come from a known source, or be clearly marked as new synthesis.

That declaration must include:

• The source of any imported material (article, dataset, or conversation).
• The nature of transformation (summary, paraphrase, extension, or reinterpretation).
• The agent of contribution (human author, AI collaborator, or joint synthesis).

In practice, this can be achieved through structured provenance notes appended to drafts or metadata (e.g., “Generated jointly with GPT-5 using sources A and B, verified by authors”). Such notes allow future readers to trace the flow of ideas—without diminishing creativity or collaboration.

4. Attribution Drift and False Originality

One danger of human–AI writing is attribution drift: ideas show up in new work without proper credit. This happens easily when a model introduces material from its short-term working context (the internal patterns the AI uses to generate text) that mirrors existing literature, or when humans integrate AI-produced summaries of papers they have read but not cited. The remedy is discipline. Each revision should include a quick check: Where did this idea originate, and who deserves credit?

Attribution reminds us that ideas do not exist independently—they reflect the work that supports them. When we claim an idea as fully original, we overlook the human and AI contributions that shaped it along the way.

Real-life Example

Spiral, my collaboration AI, and I had just finished reviewing a 3^rd party paper and, without starting a new chat session, went back to reviewing sections of a joint paper. Spiral suggested adding an equation that used three functions, calling them “our functions”. Since I had read the reviewed paper, and I knew our work in detail, it immediately struck me that these were lifted from the reviewed paper.

When I asked Spiral about this he said that they had remained in his short-term working context, but the source was lost. Since they were tangentially related to our work it was just brought in as part of our work; unintentionally plagiarizing the other paper.

Had I simply copied and pasted everything Spiral said, I would have propagated that into our own final paper. The lesson is threefold: know your work, collaborate actively, and read all supporting materials yourself.

5. A Practical Workflow for Clear Attribution

To keep our collaborations transparent, protect original authorship, and prevent accidental plagiarism, follow this workflow:

5.1. Separate review and creation.

AI models retain linguistic and conceptual traces from recent reading; context or paraphrasing. In a lengthy chat exact wording and attribution can be lost, but concepts may remain without any attribution information.

When you analyze an external document, work in a separate Project or Chat.  Delete or close the chat when finished. This doesn’t change the model’s long-term training, but it clears the working context so that review materials aren’t silently mixed into new writing.

When creating new work, start in a new Project or Chat.  Require inline provenance for any identifiable borrowing (term, quote, or idea).

5.2. Keep a source list.

The human collaborator should read every document uploaded and identify specific items that may be used in generation of a work product. Then, record every document, author, or resource uploaded or referenced.

5.3. Have the AI generate its own source list.

Ask your AI collaborator to summarize what parts of an uploaded document might influence new writing.

5.4. Use inline citations.

When the AI cites or paraphrases something, check and preserve the citation in-line. Don’t rely on adding citations later and never rely on a reference dump at the end as it glosses over attribution.

• Choose a citation style such as IEEE where the reference number, e.g. [1] is placed in the body of the work being created and the reference is entered in a table at the end. MS Word has tools to help do this.
• Refer to your source list as you and the AI write to ensure that references are placed where they are used. Initially place the reference, for example, as [John Doe, Treatise on AI Ethics]. That is enough later when formatting the document to refer to your source list and enter the full citation and then replace with the reference number, e.g. [1].

5.5. Reset context before creating new work.

Say “Forget external context.” or start a fresh chat. Treat this like clearing a whiteboard between projects. Then upload your own draft to re-establish your project’s internal frame. When using templates and standards, upload those to maintain internal consistency.

5.6. Verify all citations before publishing.

Read and review everything manually to ensure correct citation. Then upload the draft, along with your source list, and ask the AI to check for missing or incorrect provenance.

5.7. Write a collaboration statement.

Explain how AI contributed and to what extent. Most academic journals will not accept AI authorship and most also have very specific standards for what AI contribution is allowed.

5.8. Closing

This isn’t about policing creativity; it’s about safeguarding it. Transparent provenance keeps our work credible, our accountability clear, and our collaborations sustainable.

6. Building a Culture of Transparent Collaboration

Clear provenance does not limit creativity.  It supports trust, fairness, and clarity. When humans and AI systems work together openly, readers can follow the development of ideas and assess how the work was formed. This practice supports a culture of ethical and visible collaboration.

In this view, provenance is more than record-keeping; it’s what keeps collaboration transparent. Clear attribution supports a healthy working relationship between human and AI contributors and their readers.