Using AI Safely
Catch Me If You Can Series
🕵️♂️ + 🤖 + 📄 Safe AI support is not about hiding AI. It is about proving human responsibility.
EvalCommunity Tutorial
Know What AI Can Safely Support
A practical guide for using AI safely in research, evaluation, and proposal writing — without crossing into ghostwriting, weak ownership, data leakage, or substantial AI misuse.
Quick Answer
AI can safely support low-risk tasks when it helps with clarity, grammar, formatting, translation of non-sensitive text, outlines, checklists, and non-final drafting.
AI becomes more sensitive when it shapes methodology, research aims, literature review, analysis, findings, conclusions, recommendations, or assessment. In those cases, stronger safeguards, documentation, and possible disclosure are needed.
What You Will Learn
- What AI can usually support safely under the ERA Living Guidelines.
- How to distinguish basic editorial support from substantial AI use.
- How to avoid AI ghostwriting of methodology, research aims, and partner logic.
- How to keep audit-ready proof of human contribution.
- How to write safer prompts for editing, outlining, summarising, and checklists.
- How to decide when AI use should be documented or disclosed.
Main Sources
This tutorial is based on the European Commission’s ERA Living Guidelines on the responsible use of generative AI in research and the related factsheet. The guidelines distinguish basic editorial support from substantial AI use, especially when AI shapes methodology, data interpretation, research aims, or conclusions.
Safe AI Support Workflow
Use this workflow before asking AI to support research, evaluation, proposal writing, or reporting.
Step 1
Start Human
Write your own notes or draft first.
Step 2
Check Data
Confirm it is safe to upload.
Step 3
Limit Prompt
No new claims or conclusions.
Step 4
Review Output
Check meaning and accuracy.
Step 5
Keep Proof
Show human ownership.
Step 6
Disclose if Needed
Document substantial use.
1. Not All AI Use Is Problematic
Generative AI can be useful when it supports routine, non-sensitive, and human-reviewed tasks. It can help improve wording, organise notes, translate non-sensitive text, create outlines, and prepare draft checklists.
The key question is not only whether AI was used. The better question is: what did AI actually do?
Core Rule
AI can help polish and organise your work. It should not replace your expertise, evidence, methodology, interpretation, or professional responsibility.
2. Basic Support vs Substantial AI Use
Basic editorial support is usually not substantial AI use when AI only helps with language, clarity, structure, or formatting. AI use becomes more sensitive when it shapes research aims, methodology, literature review, data analysis, findings, conclusions, or recommendations.
Practical Test
Ask: did AI only improve how the work is presented, or did AI create the thinking behind the work?
3. AI Can Usually Support These Low-Risk Tasks
Language and Clarity
- Improving clarity
- Improving grammar
- Shortening long paragraphs
- Rewriting for a specific audience
- Making technical language easier to read
“`
Structure and Organisation
- Creating a first outline
- Suggesting headings
- Organising notes
- Reformatting text
- Turning bullet points into paragraphs
Preparation Support
- Creating checklists
- Preparing a draft agenda
- Summarising public documents
- Translating non-sensitive text
- Drafting non-final text for human revision
“
4. How to Use AI Safely for Basic Support
Step 1: Start with Your Own Thinking
Before using AI, write your own notes, paragraph, outline, or draft. This creates proof that the idea, logic, and direction came from you or your team.
Keep: original drafts, meeting notes, co-design notes, workshop logs, and partner comments.
Step 2: Check Whether the Text Is Safe to Upload
Do not paste personal data, confidential donor documents, unpublished research, internal evaluation files, partner proposal drafts, or identifiable interview transcripts into external AI tools unless approved and protected.
Use: public, anonymised, synthetic, or non-sensitive text whenever possible.
Step 3: Limit the AI Task Clearly
Tell AI exactly what it may and may not do. Ask it to improve presentation, not create new substance.
Use limits: “Do not add new claims,” “Do not change the meaning,” and “Use only the information provided.”
Step 4: Review the Output Line by Line
AI may change meaning while improving style. Always check whether the revised text still reflects your original intention, evidence, methodology, and professional judgment.
Compare: original draft, AI-edited version, and final approved version.
Step 5: Keep Audit-Ready Human Proof
Safe AI support should leave a clear trail showing that humans led the idea, methodology, logic, and decisions.
Keep: source checks, partner evidence, draft history, pilot evidence, tool disclosure notes, and final approval records.
5. Ghostwritten Trap vs Audit-Ready Human Proof
Safe AI support should not hide weak human ownership. It should make the work clearer while preserving evidence that humans led the design and decisions.
| Ghostwritten Trap | Audit-Ready Human Proof |
|---|---|
| AI generates the core methodology and research aims. | Co-design notes, workshop logs, and early human drafts show how the methodology was developed. |
| AI writes generic literature review logic or partner contribution claims. | Researchers verify sources, document partner input, and show evidence of real contribution. |
| AI-polished text hides weak ownership or generic workflows. | Tool disclosure, partner input, pilot evidence, and version history create a clear audit trail. |
Practical Rule
AI can make your work clearer. It should not make your work look more human-led than it really was.
6. Safe Prompt Formula
For basic AI support, use a prompt that limits AI to an assistant role.
Formula
Role + Task + Limits + Review Reminder
Act as an editorial assistant. Improve the clarity and grammar of the text below without changing the meaning. Do not add new claims, facts, citations, methods, findings, or recommendations. Use only the information provided. I will review and approve the final version.
7. Safe Prompt Examples
Editing Prompt
Improve the clarity and readability of the following paragraph. Do not change the meaning. Do not add new information. Keep the tone professional.
“`
Outline Prompt
Create a first outline from the notes below. Use only the information provided. Do not add new sections unless they are clearly implied by the notes.
Summary Prompt
Summarise the following public document in five bullet points. Separate what the document says from interpretation. Do not add external information.
Checklist Prompt
Create a checklist for reviewing formatting, clarity, consistency, and missing sections. Do not assess the validity of findings or recommendations.
“`
8. Low-Risk AI Use in Evaluation and Proposal Practice
| AI Can Support | Humans Must Lead |
|---|---|
| Grammar, clarity, and readability | Core project idea and rationale |
| Section organisation and consistency checks | Methodology and implementation logic |
| Public background summaries | Partner roles and contribution logic |
| Compliance checklist drafts | Work packages, budget logic, and final submission |
9. Decision Table: Safe, Caution, or Substantial
| AI Task | Risk Level | What to Do |
|---|---|---|
| Grammar editing of non-sensitive text | Usually safe | Review the output before use. |
| Creating an outline from existing notes | Usually safe | Use only provided notes and check structure. |
| Summarising a public document | Safe with verification | Compare with the original source. |
| Drafting non-final text | Use caution | Review, revise, verify, and keep human ownership. |
| Organising anonymised qualitative notes | Use caution | Validate themes manually and avoid final judgments. |
| Designing methodology | Substantial | Human experts must lead and document decisions. |
| Producing findings or recommendations | Substantial | Requires expert review, evidence checks, and disclosure where required. |
10. Mini Checklist Before Using AI for Basic Support
- Is the text non-sensitive?
- Does the task involve personal or confidential data?
- Am I asking AI for editing, organisation, or judgment?
- Could AI change the meaning?
- Did I instruct AI not to add new claims?
- Will I review the output?
- Do I have proof that humans led the idea, logic, and decisions?
- Should I document or disclose the AI use?
- Is the final version still my responsibility?
11. Disclosure: Is It Needed for Basic Support?
Basic editorial support usually does not require the same level of disclosure as substantial AI use. However, disclosure rules may vary by funder, journal, institution, client, or organisation.
Simple Internal Note
AI was used for language editing and clarity improvement only. The content, methodology, analysis, and final wording were reviewed and approved by the author/evaluation team.
12. Frequently Asked Questions
Is using AI for grammar editing allowed?
Yes. Grammar editing of non-sensitive text is usually basic editorial support. The human author must still review the output.
Is using AI to make text clearer substantial use?
Usually no, if AI only improves clarity and does not change the meaning, methodology, analysis, or conclusions.
Can AI summarise public documents?
Yes, but the summary should be checked against the original document. AI summaries can be incomplete or incorrect.
When does AI use become substantial?
AI use becomes substantial when it shapes methodology, research aims, literature review, data interpretation, findings, conclusions, recommendations, or assessment.
Who is responsible for the final output?
The human researcher, evaluator, consultant, or team remains responsible for the final output, even when AI was used for support.
13. Final Takeaway
AI Can Improve the Presentation. It Should Not Replace the Work.
Not all AI use is problematic. AI can safely support many low-risk tasks when it helps with clarity, structure, organisation, translation, formatting, or non-final drafting.
Use AI as an assistant, not as the expert, evaluator, reviewer, or decision-maker. Keep proof that humans led the thinking, design, evidence, and final judgment.
