Can AI write your Horizon Europe proposal, without leaking your IP?

Every researcher preparing a proposal now asks the same question: "Can I just use ChatGPT to write my Horizon Europe proposal?" The honest answer has two parts, and both matter.

The first is a scoring problem. A general-purpose chatbot will happily fill forty pages, and evaluators will just as happily mark those pages down, because generic text is precisely what this evaluation punishes.

The second is a confidentiality problem, and for a European consortium it is the more serious of the two. A proposal contains unpublished research, budgets, and intellectual property. Pasting it into a consumer chatbot sends all of that to servers outside the EU.

So the real question is not whether an AI can produce text. It obviously can. The real question is whether you can use AI without producing text that scores badly, and without leaking your consortium's confidential IP outside the EU. This page answers both, as part of our guide on how to write a winning Horizon Europe proposal.

What AI can and cannot do on a proposal

Start with what is genuinely useful, because it is real. Generative AI drafts, restructures, tightens language, summarises source material, and turns rough notes into readable prose faster than any human. If you write in English as a second language, it levels the field. None of that should be dismissed, and none of it needs to be.

What AI cannot do is win the grant for you, and the reason is specific rather than philosophical. A general-purpose chatbot does not know your call topic. It does not know the expected outcomes attached to it, the official 0 to 5 scoring rubric, the tie-break order, or the 2026 changes to the rules. It has no idea what the evaluator sitting across from your proposal is instructed to look for.

Lacking all of that, it does the only thing it can: it fills the page fluently and generically. The prose flows, the vocabulary sounds right, and nothing in it is anchored to the call you are actually answering. Fluent and generic is a comfortable combination for a blog post. For a Horizon Europe proposal, it is a losing one.

Why generic AI text scores badly

Evaluators score a proposal "as submitted and not on its potential", and they strike out assertions that are not backed by evidence. That is the core of the method, and it is exactly where generic AI prose fails, because generic AI prose is assertion-dense by construction.

These are real phrasings that evaluators have flagged on actual proposals: a project "strategically integrating cutting-edge technologies", a consortium "committed to a clear and innovative path", a pathway that assumes "seamless uptake". Each one reads well. Each one says nothing an evaluator can score. There is no mechanism, no evidence, no target, no link to the expected outcomes of the call. An assertion without backing is not a weak point in this system, it is a struck-out point.

Note what the failure mode actually is. It is not "an AI wrote this". Evaluators do not score authorship. The failure mode is "something generic wrote this", and a chatbot with no knowledge of your call writes generic by default. A tired human on a deadline produces the same empty phrases; the chatbot simply produces them faster and in greater volume.

If you want the full picture of what evaluators reward and strike out, read how Horizon Europe proposals are scored. And because unbacked assertion is only one of the recurring failure patterns, see also the typical mistakes in Horizon Europe proposals.

The declaration is not a formality

Horizon Europe does not ban generative AI in proposal preparation, but it attaches obligations to it, and they have teeth. The application form requires you to take full responsibility for any AI-generated content, to be transparent about the tools you used, to be able to provide the sources behind any text an AI generated or rewrote, and to check that text for plagiarism. Failing these obligations can render a proposal ineligible, which is a heavier consequence than most researchers assume when they reach for a chatbot.

Read together, those obligations settle a debate that wastes a great deal of researchers' time. The goal is never to disguise AI use, and never to write in a way that defeats an "AI detector". You are declaring it anyway, and you stay accountable for every line. The goal is to use AI well enough that the text is specific, evidenced and traceable to real sources, which is exactly what a chatbot producing generic prose and fabricated citations does not give you.

One separate point, often confused with the declaration: proposals whose project itself involves AI must meet the AI Act robustness and trustworthiness requirements where applicable. That is about what you are proposing to build, not about how you wrote the document.

The lesson is the same from every angle. Transparency is required, accountability stays with you, so compete on quality, not concealment.

The confidentiality problem

Here is the part of the question that too few people ask, and it is the core of this page.

A Horizon Europe proposal is one of the most confidential documents a research team ever produces. It contains unpublished research strategy, budgets, the composition of your consortium, intellectual property, and exploitation routes that competitors would pay to see. Your partners shared that material with you under an implicit, and often explicit, promise of confidentiality.

Now consider what happens when you paste that document into a consumer chatbot.

• It leaves the EU. Consumer AI tools run on US servers, for example OpenAI's own servers, or Microsoft Azure behind Copilot and SharePoint. Those servers remain exposed to the US CLOUD Act regardless of the stated hosting region.
• It persists. The content is stored in your thread history, or indexed in your organisation's tenant, rather than discarded once the answer is produced.
• It may be used. Unless you are on a specific tier with the right settings, content can be retained and, on some tiers, used to improve the model.

There is a sharper version of this risk for anything you intend to protect. Many providers retain content for "abuse monitoring" even when training is switched off, and disclosing an unpublished invention to a third party can, in some circumstances, count against the novelty that a patent requires. For a consortium whose exploitation route runs through intellectual property, a proposal pasted into a consumer tool is not a private working document, and that is not a hypothetical concern.

To be precise about the enterprise case: Copilot and SharePoint are legitimate enterprise tools, and the comparison here is architectural, not moral. The point is jurisdictional. Whatever the contract says about regions and tenants, the infrastructure underneath is operated by a US company subject to US law.

For a European consortium, that is a GDPR question and a sovereignty question, not a matter of taste. Your partners' unpublished IP, sitting on a server outside EU jurisdiction, is a genuine exposure, and it is an exposure you created with a paste.

The European way to use AI

There is a way to get the drafting help without the exposure, and it is architectural, not behavioural. You do not solve it with a usage policy that asks researchers to be careful. You solve it by building the AI into infrastructure where the leak cannot happen. This is how GrantForge is built:

• EU model, Zero Data Retention. Processing runs on an EU-based language model under Zero Data Retention: calls run stateless, nothing is kept after the answer is produced, and your content is never used to train models.
• EU hosting end to end. The whole stack (application, database, file storage) is hosted in the EU, in Germany, on an EU cloud operator rather than a US hyperscaler (an EU cloud operator, ISO 27001 at the data-center level). EU jurisdiction applies end to end, and there is no CLOUD Act exposure.
• Anonymisation before any non-default provider. Before any non-default AI provider sees an imported document, named entities (organisation names, person names, project codenames, email addresses) are stripped and replaced with neutral tokens. The token-to-name mapping never leaves EU infrastructure, and real names are re-attached locally afterwards.
• Bring your own key. A workspace can bring its own AI key (BYOK) if it prefers its own provider.
• Paperwork for your DPO. A data processing agreement (GDPR Article 28) and a security pack are available on request.

The full detail of this setup is on our security page: how your data stays in the EU.

And here the confidentiality answer meets the scoring answer, because they turn out to be the same answer. The cure for generic AI text is not evasion, it is grounding. An assistant that knows your call topic, the rubric, and your own material writes specific, evaluator-aware text. A general chatbot cannot, because it has none of that context. On this rubric, specificity beats fluency, every time. An AI that is grounded in your call does not read like AI, because it has something concrete to say.

The takeaway

Yes, use AI. It is now part of the job, and Horizon Europe requires you to declare it anyway, so there is nothing to hide and no reason to pretend otherwise.

But use it in a way that satisfies both halves of the real question. The text it helps you produce must be specific and scorable, grounded in your call and the rubric, not fluent filler an evaluator strikes out. And your consortium's confidential IP must stay inside EU jurisdiction while the AI reads it, not sit in a thread history on a US server.

Those two requirements point to the same answer: not a general chatbot, but an assistant grounded in your call and built on EU infrastructure.

See how GrantForge keeps your proposals in the EU

Part of our guide on how to write a winning Horizon Europe proposal. See also how your data stays in the EU.