The proposal is free now. The trust is not.

The EU funding debate is one version of a problem every donor, UN agency and INGO now faces.

The funding system is writing rules for a world in which AI is optional. Microsoft is building a world in which it’s part of the document.

Copilot now sits inside Word, Excel, PowerPoint and Outlook for a growing number of Microsoft 365 users. Some Copilot Chat functions come with eligible business subscriptions. The fuller Microsoft 365 Copilot experience still requires an additional licence. But the direction is clear. AI-assisted drafting is moving from a separate tool into the software people already use to write proposals, budgets and reports.

That changes the practical question. It’s no longer whether every applicant has equal access to the best AI. They don’t. Language, connectivity, cost and skill still matter. The more defensible assumption is that capable AI is now cheap and widespread enough that any organisation assessing proposals should expect applicants to be using it.

That applies well beyond EU funding. It applies when a bilateral donor runs a call for proposals, when a UN agency selects implementing partners and when an international NGO chooses national partners. It applies whenever a funder receives more polished applications than its staff can meaningfully examine.

Tomorrow, I’ll join Annika Jaansoo, Andrea Igazi-Eppacher, Gilberto Martinez, Ena Hodzic, Gilles Meijer and Matteo Feruglio for a roundtable on what generative AI is doing to EU proposal evaluation.

The panel brings together programme assessors, evaluators, trainers and researchers. My contribution comes from outside the EU institutions. I’m interested in what happens when the technology changes faster than the rules used to govern it.

The online discussion starts at 10:00 CEST / 09:00 UK time on Wednesday 16 September. You can view the event and register here: https://www.linkedin.com/events/7503011437900976128/

My starting point is simple:

You cannot call something fraud unless you have prohibited it.

In much of the European funding system, the rules don’t prohibit applicants from using generative AI. They place responsibility for the final submission on the applicant. That distinction matters. If an applicant remains responsible for every claim, citation and commitment, AI use isn’t automatically fraud. It’s a question of disclosure, verification and accountability.

The real challenge is therefore not how to police AI out of funding applications. It’s how to set standards for its use in a system where polished drafting is becoming abundant and human assessment remains scarce.

What counts as using AI?

The first problem is definitional. Does using AI mean asking ChatGPT to draft an entire proposal? Does it include improving a paragraph in Word, translating a concept note, summarising a call document, checking a budget narrative against the scoring criteria or rephrasing text for a second-language writer?

If AI is embedded in the software, the boundary becomes even harder to draw. A person may use spellcheck, predictive text, translation, rewriting and Copilot during the same drafting session without thinking of them as separate acts.

A rule that simply says “disclose AI use” therefore leaves important questions unanswered. Which uses require disclosure? At what level of detail? How will the information affect assessment? Will disclosure help evaluators understand the proposal, or merely create a new penalty for honest applicants?

The European Innovation Council’s guidance offers one answer. Applicants may use generative AI, but they remain fully responsible for the proposal’s content. They should disclose which tools they used and how, while ensuring that the proposal reflects their own ideas.

The Horizon Europe application form follows a similar principle. Applicants must exercise care, validate the output, check sources and citations, consider the risk of plagiarism, and remain responsible for everything submitted.

The current Horizon expert briefing goes further. It says that an applicant’s use of generative AI in drafting a proposal must not, by itself, be treated as a reason to penalise the proposal. That is a sensible position, but it creates a second question: if AI-assisted drafting cannot be treated as evidence against an applicant, what evidence should evaluators trust?

A process trail is not an authorship trail

EU funding programmes have criteria, approved methodologies and audit trails. Those controls matter. But a process trail shows how a proposal moved through the system. It does not establish who developed the underlying analysis, whether the organisation understands it, or whether it can deliver what the text promises.

The named applicant still owns the proposal at submission. That provides a line of formal accountability, but it doesn’t prove that the applicant can explain the intervention logic, defend the assumptions, manage the risks or implement the work.

This gap existed before generative AI. Consultants and professional bid writers have long helped organisations turn incomplete ideas into convincing applications. AI changes the scale, speed and price of that assistance. It can produce a competent first draft in minutes, map language against call criteria, suggest indicators and reproduce the vocabulary of participation, localisation, sustainability and value for money.

That doesn’t mean the proposal is false. It means the prose is carrying less information than it once did. Good writing remains useful, but it has become a weaker signal of organisational capability.

The rules are already asymmetric

The rules for applicants and evaluators are not the same. Applicants may use generative AI under conditions of disclosure, validation and responsibility.

Evaluators face tighter constraints. The Horizon Europe expert briefing says they must not delegate proposal evaluation to AI. They may use it only for supporting tasks, such as finding background information, while protecting confidentiality and personal data. They must take precautions against hallucinations and bias, and document their use so it can be provided to the contracting authority on request.

The briefing also warns that putting proposal information into an online generative AI tool may itself breach the confidentiality requirements of the expert contract.

The logic is understandable. The applicant is presenting its own material. The evaluator is handling somebody else’s confidential proposal and exercising delegated public authority. But the asymmetry creates an operational problem. Applicants can use AI to produce more proposals, faster. Evaluators cannot simply use the same tools to absorb the additional volume. The burden grows on one side while the permitted capacity to manage it remains constrained on the other.

That is not an argument for allowing AI to make funding decisions. It is an argument for redesigning the process around the new imbalance.

Microsoft is changing the baseline

The debate often imagines that applicants make a deliberate choice to visit a generative AI platform and ask it to draft a proposal. Increasingly, AI arrives through tools they already use.

In April 2026, Microsoft announced the wider availability of Agent Mode in Word, Excel and PowerPoint. It can help users create and revise documents, analyse spreadsheets and build presentations within the applications themselves.

Microsoft subsequently reported sharp increases in the use of Copilot across Word, Excel, PowerPoint and Outlook over the comparison periods it measured. Those are Microsoft’s own product figures, not independent evidence of humanitarian or NGO adoption. But they show how quickly assisted work is becoming normal inside office software.

This has two consequences for funding systems:

• AI use will become harder to separate from ordinary document production.

• Organisations with mature systems, strong digital skills and good source material will gain more than organisations that merely have access to the same button.

The tool may be widely available. The ability to use it well is not evenly distributed.

This argument has been building

In Shadow AI in humanitarian work, I argued that staff were already using general-purpose AI for sensitive professional work, often without clear organisational rules.

In Most AI training in the humanitarian sector is teaching the wrong thing, I argued that prompt techniques are less important than judgement, verification and the ability to recognise when a plausible answer is not a reliable one.

In the Somalia market-assessment experiment, we let AI produce a complete analytical draft. Producing the text was easy. Establishing which claims deserved trust took most of the work.

In Seven propositions for the next eighteen months, I argued that the organisations which benefit most from AI will be those with strong evidence, clear workflows and accountable human judgement.

Proposal assessment brings those arguments together:

The text is getting cheaper. Trust is not.

When familiarity becomes evidence

Evaluators have always had to distinguish between the quality of a proposal and the capacity of the organisation submitting it. AI makes that distinction harder because it can reproduce the language of institutional competence. A small organisation can now present an application with the tone, structure and terminology of an experienced international bidder.

That could be positive. Applicants should not lose funding merely because they lack a professional English-language proposal team. Used well, AI can reduce some disadvantages related to language and drafting experience. But it can also reward familiarity with donor expectations without demonstrating the ability to deliver.

This matters particularly when UN agencies and international NGOs select national partners. A familiar organisation may already have approved systems, established relationships and a track record in the required formats. A first-time applicant may offer stronger local knowledge but present it less convincingly.

AI can narrow the presentation gap. It cannot tell the assessor whether the organisation has safeguarding systems that work, access to the communities it names, the capacity to manage cash or the authority to make decisions locally.

If assessors rely too heavily on proposal language, they risk funding the organisation that best reproduces the expected vocabulary. If they react by becoming more conservative, they may retreat towards existing partners because familiarity feels safer. Both outcomes can work against localisation.

Detection can’t verify capability

Automated systems may identify copied language, duplicated applications or suspicious similarity. The Erasmus+ Programme Guide describes automatic screening for similarities, possible double funding and plagiarism. Applications may be referred for further examination, and rejection is possible where the concerns are not adequately explained.

Those controls address real risks. They do not answer the central capability question. A detector cannot reliably establish who developed the idea. It cannot determine whether the organisation understands the proposal or can implement it. Nor can it distinguish consistently between responsible assistance, extensive rewriting and content generated without genuine organisational ownership.

Reliable governance starts with evidence: can the claims be supported, and can the named human explain and own the judgement?

That shifts attention from authorship theatre to verification. Can the applicant explain its causal assumptions? Can it show where the needs data came from? Can it defend the budget choices? Can the proposed team describe how implementation would work? Can references, partnerships and previous results be checked?

Those are useful questions whether or not AI was involved.

Volume is part of the governance problem

Generative AI reduces the cost of producing an application. It does not reduce the cost of evaluating one properly. That changes the economics of open calls.

More organisations can apply. Existing applicants can submit to more opportunities. Intermediaries can generate applications at greater scale. Assessors then face a larger volume of polished, criterion-aligned text.

An overloaded process is likely to rely more heavily on shortcuts. That may mean keyword scoring, previous relationships, organisational reputation or rapid judgements about whether an application “sounds credible”.

Research by James Wilsdon, Richard Jones and colleagues found that funders already make different choices about assessment models, reviewer numbers and the effort devoted to selecting proposals. AI-generated volume will intensify those choices.

The relevant denominator is not simply the number of awards. It is the amount of credible human review available.

A scheme may claim to be open because any eligible organisation can submit. But if the number of applications rises while reviewer capacity remains fixed, the meaningful access question becomes whether each proposal receives enough attention to be judged fairly.

What would a transparent funding system publish?

Most funding systems publish awards. Far fewer publish enough of the application journey to show where access narrows.

A more transparent system would report:

• How many organisations began and completed eligibility processes

• How long applicants waited for eligibility decisions

• How much review time was available per submission

• How many awards went to first-time recipients

• How many reviewer-hours were available relative to application volume

• How applications were escalated when similarity, plagiarism or unsupported claims were detected

This would help distinguish formal openness from practical accessibility. It would also expose whether AI-generated volume is creating a review bottleneck, and whether that bottleneck is pushing funding back towards known organisations.

Five questions the rules now need to answer

Funding bodies do not need a universal philosophical position on artificial intelligence. They need operational rules. I would start with five questions.

1. What uses must applicants disclose?

The rule should distinguish between translation, editing, research assistance, analytical support and substantial generation of proposal content.

2. What must remain demonstrably human?

The accountable applicant should be able to explain the intervention logic, evidence, risk decisions, budget assumptions and implementation commitments.

3. What evidence matters more than the prose?

Assessment should place greater weight on verifiable experience, references, systems, team capacity and the ability to answer grounded follow-up questions.

4. How will increased volume be managed?

If AI makes applications easier to produce, funders need to monitor submission numbers, reviewer-hours and review time per application.

5. What may evaluators do with AI?

The rules should define permitted support tasks, approved environments, confidentiality requirements, documentation and responsibility for the final judgement.

Without answers to those questions, the system will drift into an unstable compromise. Applicants will use AI because it is available. Assessors will suspect them because it is available. Both sides will operate without a common definition of acceptable use.

The question for tomorrow

Tomorrow I want to move the discussion beyond whether AI is good or bad for funding. Applicants and assessors are already using it, under rules that vary by programme.

The practical questions are harder:

• What should applicants disclose?

• What evidence should evaluators trust?

• How do we prevent increased volume from reducing meaningful access?

• How do we avoid turning familiarity into the hidden standard?

• Who has the authority to set and enforce the rules?

I’d like to be wrong about where this leads. But if funding bodies leave those questions unanswered, the likely response will be more suspicion, more procedural checking and greater reliance on organisations the system already knows.

The proposal is becoming almost free. The trust required to fund it is not.

I’ll be discussing these questions at the online roundtable on Wednesday 16 September 2026 at 10:00 CEST and 09:00 UK time.

Join or follow the event on LinkedIn.

---

AI-use disclosure: I used ChatGPT and Microsoft Copilot to help structure the argument, test counterarguments, check the prose and locate source material. I reviewed the cited sources and made the final editorial and analytical decisions. MarketImpact’s approach to AI-assisted work is explained atHow we use AI. Corrections are welcome.