Can I use AI for this? The four-question check for responsible AI use

← Back to resources

Assessing which uses of AI are responsible

Shortening a public event announcement calls for a different assessment from copying a confidential manuscript into a public AI tool to prepare a review. What matters is the combination:

Material × Tool × Task × Output

The four-question check and the AI Traffic Light are guidance models developed by Sollberger AI Consulting. They do not replace legal, data protection or institutional review. The rules of your organisation and the relevant procedure remain decisive.

The four-question check

1. Material: What will be entered?
2. Tool: Where will the data go?
3. Task: What should the AI do?
4. Output: How will the result be used?

Material: What will be entered?

Material may be public, internal, confidential, personal or particularly sensitive: for example, meeting documents, grant proposals, job applications, manuscripts, review reports, reviewer comments and unpublished research data.

Licensing, context, purpose, citation and institutional rules remain relevant even for open-access content. The more sensitive or consequential the material is for a procedure, the more carefully its use needs to be assessed.

Tool: Where will the data go?

Check the provider, processing location, storage, use for training, institutional approval, access and deletion rules. A paid personal account does not constitute institutional approval. Nor is an institutional tool automatically suitable for all content. ^[1]

A local system can reduce data transfers to external providers. Questions about access, storage, confidentiality, rights and quality remain.

Task: What should the AI do?

Improving wording, shortening, translating, structuring, drafting alternatives or summarising public information: these assistance tasks may be appropriate when the material, tool and review arrangements are suitable.

Evaluating, ranking, selecting, triaging, writing reviews or providing reasons for decisions has a stronger influence on professional judgement. The Traffic Light model takes a cautious approach: AI provides support; subject-matter assessment, decisions and responsibility are not delegated to it.

Output: How will the result be used?

An internal working aid has different consequences from external communication or material used for selection and decisions. Before reuse, check facts, sources, bias, context, tone, shifts in meaning, unacknowledged uncertainty and inappropriate additions.

Human responsibility requires practical arrangements: clear roles, review, documentation and approval. ^[2]

The four AI Traffic Light outcomes

Green: low risk

Example: Shorten a public event announcement with an appropriate tool and review the suggested text editorially. Green does not constitute legal approval or a guarantee of quality.

Amber: review or approval required

Example: Structure internal meeting notes without personal data, once the tool, data boundaries and responsibility have been clarified.

Red: not recommended in this configuration

Example: Copy confidential reviewer comments into a public AI tool to prepare reasons for a decision.

Grey: clarify first

Example: Use a new AI platform whose storage location, use of data for training and institutional approval are still unclear.

Typical practical examples

Public communication

AI can shorten public texts, make them clearer or prepare versions for different channels. The message, tone and subject-matter accuracy still need editorial review.

Grouping abstracts by topic

For public or approved abstracts, AI can suggest thematic groups. These do not constitute an assessment or final selection. Review the categories, weightings and borderline cases.

Peer review

Confidential manuscripts, review reports and reviewer comments must not be entered into unsuitable tools. The rules of the particular procedure determine which assistance is permitted. For reviews of grant proposals, the SNSF permits certain assistance after reviewers have made their own independent assessment, while requiring confidentiality, data protection and disclosure of AI use. Subject-matter assessment remains the reviewers’ responsibility. ^[3]

Why assistance is not automatically harmless

Assistance tasks can also produce incomplete, biased or inaccurate results. Check omissions, weightings and possible consequences for decisions. ^[4]

The Amber checklist: What needs to be clarified?

• Clarify the tool and data flow: Who operates the tool, where is data processed and which rules apply?
• Minimise data: Enter only what is genuinely necessary and permitted for the task.
• Review the result: Check facts, sources, omissions, bias and tone.
• Assign responsibility: Clarify responsibility, approval and documentation before reuse. Where needed, involve data protection, IT or information security, legal services or the responsible management.

When the AI Traffic Light helps

The Traffic Light provides an initial assessment of typical work situations in universities, research support services, foundations, learned societies, editorial teams, associations and non-profit organisations. It asks only for categories, not actual content. Its result provides guidance, not approval.

Use the AI Traffic Light

The link to prompting

Once the proposed use has been clarified, good prompts help guide the task. For an explanation of how this develops into a shared working process, see AI in research support: From individual prompting to shared AI workflows.

Conclusion

Assessing material, tool, task and output together makes uncertainties specific. Good AI use is task-focused, minimises data, takes account of the tool, can be reviewed, is transparent as appropriate to the context and remains under human responsibility.