Living guidelines on the responsible use of generative AI in research (2026)¶
Source
Document: "Living guidelines on the responsible use of generative AI in research. ERA Forum Stakeholders' document. Third version, May 2026". Issuing body: European Commission, Directorate-General for Research and Innovation (Unit E4, AI in Science and Critical Technologies), developed with the European Research Area Forum. Date: May 2026 (third version; first published 2024). Distilled: 2026-09-05.
Summary¶
The guidelines are the European Commission's non-binding recommendations to researchers, research organisations, and research funders on using generative AI. They rest on the European Code of Conduct for Research Integrity and on the trustworthy-AI principles (reliability, honesty, respect, accountability). For this record the weight sits in three places: what a researcher may send to an external AI tool, what an organisation should host itself, and how AI use is disclosed. The guidelines name no provider and no jurisdiction. They set the questions a researcher or institution has to answer before data reaches a tool, and they put the answer on the researcher.
Hooks, with citations¶
1. Researchers stay accountable for AI-assisted output and disclose substantial use. The document: "Researchers are accountable for the integrity of the content they produce generated by or with the support of AI tools" and "AI systems are neither authors nor co-authors" (Recommendations for researchers, 1, p. 7). "Researchers, to be transparent, detail which generative AI tools have been used substantially in their research processes. When generative AI meaningfully shapes results, researchers transparently note its use according to the guidelines of their journal or standards in their discipline in the methods section (or equivalent)" (Recommendations for researchers, 2, p. 8). Relevance: disclosing what an agent did, and where, is the human/agent attribution that BP07 asks for. Audiences: practitioners.
2. Anything uploaded may be used for training; protect unpublished or sensitive work. The document: "Researchers remain mindful that generated or uploaded input (text, data, prompts, images, etc.) could be used for other purposes, such as the training of AI models. Therefore, they protect unpublished or sensitive work (such as their own or others' unpublished work) by taking care not to upload it into an external AI system unless there are assurances that the data will not be re-used, e.g., to train future language models or to the untraceable and unverifiable reuse of data" (Recommendations for researchers, 3, p. 8). Relevance: the Commission's guidance names training on inputs as the default hazard of an external tool and puts the burden of obtaining assurances on the researcher (BP10, bp10-a5). Audiences: practitioners, governance.
3. Third parties' personal data reaches an external tool only with a lawful basis and a clear purpose. The document: "Researchers take care not to provide third parties' personal data to external generative AI systems unless the data subject (individual) has given them their consent and researchers have a clear goal for which the personal data are to be used so compliance with EU data protection rules is ensured" (Recommendations for researchers, 3, p. 8). Footnote 24 adds: "Apart from personal consent there could be other lawful basis that could justify the processing of personal data". Relevance: sending personal data to an external model is a processing step that needs its own lawful basis and purpose; the convenience of a tool supplies neither. Audiences: practitioners, governance.
4. Check who runs the tool, where it runs, and under which privacy options. The document: "They check, for example, their institutional guidelines, the privacy options of the tools, who is managing the tool (public or private institutions, companies, etc.), where the tool is running and implications for any information uploaded. This could range from closed environments, hosting on a third-party infrastructure with guaranteed privacy, to open internet-accessible platforms" (Recommendations for researchers, 3, pp. 8-9). Relevance: the operator, the location, and the terms of a tool decide what happens to uploaded data; BP11 (bp11-a1 to bp11-a3) turns this per-researcher check into an institutional decision and an agent limit. Audiences: practitioners, governance, providers.
5. Pick the tool that fits the task. The document: researchers minimise environmental impact "by evaluating whether and which AI tool is best suited for the intended task and by using the most effective prompting techniques" (Recommendations for researchers, 5, p. 9). Relevance: a Commission-level statement of method fit (BP01, bp1-a1), argued from resource use rather than from accuracy. Audiences: practitioners.
6. Keep generative AI out of peer review and proposal evaluation. The document: researchers "Refrain from using generative AI tools substantially in sensitive activities that could impact other researchers or organisations (for example peer review, evaluation of research proposals, etc)", both to avoid unfair assessment and because "this will safeguard the original unpublished work of fellow researchers from potential exposure or inclusion in an AI model" (Recommendations for researchers, 6, p. 9). Funders are told the same for their own processes: generative AI "cannot take a role in the assessment or evaluation of the scientific content of projects" (Recommendations for research funding organisations, 2, p. 11). Relevance: grounds the line in the BP09 governance tab that confidential manuscripts are not uploaded to external models. Audiences: practitioners, governance.
7. Organisations host tools they govern, and keep track of use. The document: research organisations should, "Whenever possible and necessary, implement locally hosted or cloud-based generative AI tools that they govern themselves. This enables their employees to feed their scientific data into a tool that ensures data protection and confidentiality" (Recommendations for research organisations, 4, p. 10). Footnote 31 allows tools "governed by trustworthy third parties e.g. partner research organizations, the EU or trusted countries". Organisations also "Actively keep track of the evolution and use of generative AI systems within their organisations" and "provide support and guidelines to ensure compliance with ethical and legal requirements (EU data protection rules, protection of intellectual property rights, etc.)" (Recommendations for research organisations, 1 and 2, p. 10). Relevance: the institution, not the individual, is asked to supply a governed route for scientific data (BP02, bp2-a4) and to keep an inventory of use (BP04 governance tab). Audiences: governance, providers.
8. Funders weigh data protection when choosing tools and ask applicants to declare use. The document: when choosing tools, funders "take into account elements concerning quality, transparency, integrity, data protection, confidentiality and respect for intellectual property rights, as well as its environmental impact" (Recommendations for research funding organisations, 2, p. 11), and "Applicants declare if they substantially used generative AI tool(s) to prepare their application" (Recommendations for research funding organisations, 3, p. 11). Relevance: funders are an audience the record does not address directly; the disclosure line matches BP07. Audiences: governance.
9. Other people's AI tools are part of the exposure. The document: "Researchers and organisations remain mindful that third parties may use generative AI tools. For example, in the context of meetings, note-taking, summarising discussions or creating overviews of documents (that could be confidential)" (p. 7). Relevance: confidentiality can be lost through a collaborator's or a reviewer's tool as well as through one's own. Audiences: practitioners, governance.
Mapping to practices¶
| Hook | Supports | In tension with |
|---|---|---|
| 1 | BP07 (bp7-a2) | |
| 2 | BP10 (bp10-a5); BP11 (bp11-a1 to bp11-a3) | |
| 3 | BP11 (bp11-a1 to bp11-a3); context for BP10 | |
| 4 | BP11 (bp11-a1 to bp11-a3); context for BP03 vetting | |
| 5 | BP01 (bp1-a1) | |
| 6 | BP09 governance tab (peer review) | |
| 7 | BP02 (bp2-a4); BP04 governance tab (inventory) | |
| 8 | context for BP07 and BP09 | |
| 9 | context for BP09 |
No contradictions. One difference in framing: the guidelines address generative AI tools operated by a person, and put the checks on that person. The record addresses agents that act on their own, where the researcher may never see what is sent. That difference is why BP04 moves the check from the researcher's habits into the institution's contracts and the agent's configured limits.
Proposed changes to practices¶
Proposals for editor review; applied items are ticked. The EDPB Opinion 28/2024 entry and the reference works added with this entry ground the same practice.
- New practice BP11, "Decide where research data may be sent". Drafted here as two BP04 atoms with tab text and two examples; on editor review it became its own practice, because deciding which data may go to which service applies to people using AI tools as well as to agents. The drafted bp4-a8 became bp11-a3 and bp4-a7 became bp11-a5.
- Add this entry to the
sources:of BP07, BP10, BP01, and BP02; the provenance edges are in place. - Optionally note in the BP09 governance tab that the Commission's guidelines make the same point for funders' proposal evaluation.
Cautions and gaps¶
The guidelines are non-binding recommendations, and say so. They concern generative AI tools used by a person; agents that run code and call services on their own are not discussed, so the egress path described in the Gagneur entry is outside their frame. They name no provider, no jurisdiction, and no contract type; "assurances that the data will not be re-used" is left to the reader to obtain. Footnote 31's "trusted countries" is policy language; the record does not adopt it as a criterion. The document is a living text and the third version supersedes the earlier two; the quotes here are from the May 2026 text, and any quote should carry the version. The environmental-impact material is cited only for its method-fit sentence.