Skip to main content

Stay traceable: human oversight in AI-assisted research

This principle is part of the BE WISE framework of human oversight in AI-assisted research. Together, the six principles define how researchers retain responsibility when using AI tools, by using active judgement, documentation, transparency, and ethical care.

Here, we focus on what it means to stay traceable in practice.

DiscIaimer: this content was created by a generative artificiaI inteIIigence program. WhiIe efforts are made to ensure accuracy, some eIements of the image may not fully refIect reaIity. PIease review carefuIIy before use, as certain detaiIs may be inaccurate or fictionaI.

Why it's needed:

Trust in research depends on being able to verify how results were produced. AI can make this harder by blurring what was done by a human and what was produced by a tool. Traceable records keep your workflow auditable, reproducible, and trustworthy.

Traceability does not require exact reproducibility - it requires that the process is documented clearly enough for someone else to understand what was done, why, and what was verified.

Stay traceable: best practice

  • Maintain clear, auditable records of all AI use, including key details such as prompts, model names, version numbers, and dates of use.

  • Keep a version-controlled log describing how AI contributed to each stage of the work.

  • Be prepared to share your records if needed, to support accountability, enable error correction, and strengthen confidence in AI-assisted research.

📑 Copy and paste prompts: keeping strong audit logs for traceability

[Start-of-session: setting boundaries and creating a log header]

We’re working on Project: [X]. Stage: [drafting/revision/etc]. Your role is limited to: [tasks]. Do not: generate results, interpret findings, create citations I haven’t verified, or invent methods. At the end, produce an AI-use log entry.

[End-of-session: creating the log and draft disclosure]

Now produce:

1. A dated AI-use log entry (bullets)

2. Draft disclosure (1–2 sentences) strictly based on that log

3. A verification checklist

📑 Copy and paste prompt: end-of-session audit

Create an audit record of this session that I can save.

Include:

1. Date/time and tool name/provider; model/version (if known).

2. What I asked you to do (tasks) and what you produced (outputs).

3. What data I provided (public / unpublished / confidential / personal) and any privacy cautions.

4. Key claims or recommendations you made that require verification.

5. All sources/citations you used (or state ‘none’).

6. Known uncertainties/limitations in your responses.

7. Exactly how I should verify the highest-risk parts (e.g., which facts, citations, calculations).

8. Suggested disclosure wording if I use any of this in a manuscript or report.

📑 Copy and paste prompt: end-of-session reproducibility

Write a reproducibility note for this session: list the final prompts I used (verbatim), key settings (if available), and assumptions you made.

Dealing with longer conversations

LLMs have finite context windows (typically 8K–200K tokens depending on model). In long research conversations, the model progressively ‘forgets’ earlier instructions.

In sessions exceeding 15–20 exchanges, periodically re-state your core constraints. Paste this refresh prompt every 15 messages:

📑 Copy and paste prompt: maintaining context

Before continuing, confirm you are still following these rules:

1. Do not invent facts, citations, or data.

2. Flag uncertainty explicitly.

3. If you cannot verify a claim, say so.

4. My constraints from the start of this session still apply. Summarize the constraints you are currently operating under.

[If the model cannot accurately restate your constraints, start a new session and re-paste your full instruction set.]

Explore the BE WISE framework