AI-assisted Horizon proposals: rebuild the evidence trail before submission

Generative AI can help a Horizon Europe team organise notes, compare drafts, surface inconsistencies and accelerate first-pass writing. It can also create polished statements whose evidence, ownership or legal status has never been checked. The proposal risk lies less in the tool label than in allowing fluency to replace verification.

The current European Commission RIA/IA application form gives applicants explicit responsibilities when generative AI is used in proposal preparation. It calls for careful consideration, validation of content for accuracy and appropriateness, compliance with intellectual-property rules, disclosure of the tools used and how they were used, double-checking of sources and citations, plagiarism checks and acknowledgement of limitations such as bias, errors and knowledge gaps.

The Commission's updated 2026 living guidelines for generative AI in research reinforce accountability, transparency and responsibility. They are non-binding guidance, but they identify practical control areas that proposal teams should not ignore: privacy, confidentiality, intellectual property, information management and human responsibility for outputs.

The useful operating principle is simple: an AI output is a draft or analytical aid, not a source. Every material claim must travel back to an identifiable authority, dataset, partner input or documented assumption before it enters the submission.

PRINCEPS recommends a seven-field AI contribution ledger for each material drafting or analytical use:

1. Task — what the tool was asked to do, such as structuring, summarising, translating, editing, comparing or generating options.

2. Tool and configuration — the provider, product or model where known, date of use and any material settings or retrieval mode.

3. Input class — whether the input contained public material, consortium-confidential information, personal data, unpublished research or protected intellectual property.

4. Output location — the section, table, annex, calculation or working paper influenced by the tool.

5. Verification evidence — the primary source, dataset, calculation, partner confirmation or expert review used to test every material statement.

6. Human decision owner — the named person who accepted, changed or rejected the output and remains accountable for the final text.

7. Disclosure treatment — how the use will be described in the proposal or internal record under the applicable template and organisational policy.

The ledger should sit alongside a claim register. A claim register records the exact proposition, source, date, relevant page or section, owner, confidence and where the claim appears in the proposal. If a claim cannot be traced, it should be removed, qualified as an assumption or sent back for evidence.

Confidentiality controls should be applied before any prompt is sent. Teams need a clear rule for what may enter a third-party system, which contractual and institutional terms apply, where data may be processed, whether prompts can be retained or used for service improvement and how access is controlled. Sanitising names does not necessarily remove commercial, personal or research sensitivity.

Intellectual-property control requires more than a plagiarism check. Reviewers should test whether generated text reproduces protected expression, whether a proposed method or figure has a traceable origin, whether partner background is properly identified and whether citations actually support the statement attributed to them.

AI-assisted quantitative work needs its own audit trail. Preserve the input data, units, formula or method, assumptions, intermediate calculation and independent recomputation. A plausible total is not evidence that the arithmetic, exchange rate, person-month allocation or impact estimate is correct.

African consortium partners should retain authority over descriptions of their institutions, contexts, data, risks and proposed roles. AI should not flatten local evidence into generic regional language or create commitments that a named partner has not reviewed and accepted.

Before submission, run a clean-room review in which a reviewer checks the proposal only against the live topic, controlling template, authoritative sources, partner approvals and validated calculations. The final question is not whether a sentence sounds convincing. It is whether the team can show why it is true, who owns it and what evidence would survive evaluator scrutiny.

This is a PRINCEPS proposal-control method, not legal advice or a substitute for the live call documents, institutional policies, contractual terms or data-protection assessment. Responsible use does not guarantee eligibility, compliance or evaluation success.

Editorial note: This is PRINCEPS analysis for general information. It does not replace the official work programme, topic conditions, submission system, grant rules or professional advice specific to an application.