Part One: ‘Houston, we have a problem’

Hello humans. This two-part blog is for those who, for want of a better alternative or lacking the attention span for Substack, are still prepared to wade through the machine-generated slop on LinkedIn in search of interesting content and contacts.

It sets out the latest evidence and research on an existential threat to the operating model of the international development sector: the rapid rise in the use of AI in procurement and grant-giving, and the reality of funders being swamped by AI-generated applications that are impossible to evaluate.

Part One looks at the problems the sector confronts: legacy systems and approaches overwhelmed by huge volumes of applications, the homogeneity of those applications, and the stark reality that locally led grantees and those with real front-line experience are at a disadvantage to those with AI smarts.

As legacy systems for grant-making collapse, donors are increasingly making awards to people they already know, to those who respond quickest, or by lottery. These are short cuts. They come at the expense of quality and choice, and they must be countered with new and innovative ways of assessing capability and determining awards.

As Chairman of MetricsLed, I write with some experience. We have seen 2026 application numbers for a major aid-funded global scholarship programme hit 133,000. Each candidate is required to submit four essays of 300–400 words, leaving evaluators with close to half a million documents to review and score. With a little over 1,500 places on offer, completed submissions were up 20,000 on last year.

We also run large grant-making programmes for UK government departments such as the FCDO and Defra, and for donors in countries such as Somalia and Syria. Here too the volume, homogeneity and polished quality of submissions present major challenges, and we have spent the last three years working with partners to pioneer new approaches to verification and capability assessment.

Next week’s blog is a call to action and considers the solutions to these problems, arguing that donors, fund managers and grant-givers cannot stand still. They must actively encourage, invest in and demand new approaches to grant-giving, application and procurement processes to head off the existential threat that the untrammelled use of AI represents.

The call to action is an urgent one. Identifying and appointing the best partner, delivery agent or grantee for a given role is at the core of effective, efficient and responsive development and humanitarian action. But that task can no longer be discharged primarily by reading what applicants write about themselves. It requires a deliberate shift towards what can be known about them from sources other than their own proposals.

Drivers of change

The world of aid, development and humanitarian action is going through a period of intense change, driven by a small number of new, significant and highly disruptive dynamics:

• Budget: unprecedented, swingeing cuts to official development assistance (ODA) by almost all major state donors.
• Administration: the closure of stand-alone aid ministries and cuts to headquarters management and administrative headcount and capacity.
• Delivery modalities: the pendulum is swinging away from directly awarded, top-down, large-scale technical assistance programming towards dispersed, bottom-up, small-scale, locally led delivery through grant-making instruments, funds and facilities.

These dynamics, interlinked and self-reinforcing, are forcing intense change on a sector that was until recently characterised by a clear sense of purpose and confident, steady growth.

With political headwinds howling against aid as a concept, and budgets under threat, a new landscape for programme design and procurement is emerging.

In this landscape, decisions on the appointment of delivery agents are increasingly decentralised to fund managers and funding instruments located in the Global South. It is characterised by larger numbers of smaller awards, and populated by many smaller, newer and less formally incorporated actors, all of whom can now generate high volumes of polished, responsive but homogeneous proposals that are difficult to evaluate.

Legacy approaches and the limits of narrative assessment

For decades the dominant instrument of partner selection in international development has been the written proposal. It has served as both application and primary evidence base: a document in which an organisation describes its understanding of the problem, its proposed approach, its institutional capacity and its expected results.

In a world of relatively few, relatively large awards assessed by relatively well-resourced central teams, this model was imperfect but workable. Reviewers could still form judgements about authenticity, organisational reality and comparative advantage through a combination of close reading, clarification questions and occasional site visits or reference checks.

That equilibrium no longer holds. The simultaneous shift towards higher volumes of smaller grants, the localisation of fund-management decisions and the near-universal availability of large language models have broken the conditions under which narrative assessment could function.

An applicant can now produce, in minutes, a polished, context-sensitive, results-oriented proposal that meets every formal criterion of a call for proposals. The text can be tailored to the donor’s or fund manager’s preferred language of localisation, gender equality, climate resilience or value for money. It can cite recent literature, reference the correct global frameworks and demonstrate apparent familiarity with the operating environment. The difference between a proposal written by a team with deep local knowledge and one assembled by a consultant using an AI model is increasingly difficult to detect from the document alone.

The challenge of detection

Polished AI output cannot reliably be distinguished from original work.

• Blinded expert reviewers. Scientists identified only 68 per cent of ChatGPT-written abstracts, all of which passed plagiarism checks, and wrongly labelled 14 per cent of genuine abstracts as AI-generated. [1]
• Detectors do not work. OpenAI withdrew its own detector in 2023 after it identified just 26 per cent of AI-written text. [2] A test of fourteen detection tools, including Turnitin, found them ‘neither accurate nor reliable’. [3]
• Foundations cannot tell. In a 2024 survey of 527 US foundations, 57 per cent did not know whether they had received AI-generated applications. [4]
• Machines assessing machines. In 2023 some Australian Research Council assessor reports were found to contain ChatGPT-generated text, [5] and up to 17 per cent of peer-review text at leading AI conferences has been found to be LLM-modified. [6]

The challenge of volume

Although anecdotal evidence abounds, there is little public detail on how AI has affected application volumes for FCDO and wider ODA spending. It is difficult, for now, to assess the impact of AI on the document-driven, narrative-based ‘analogue’ procurement systems that remain in place.

MetricsLed has a clear view of the volume question from our own support to one of the world’s most prestigious global scholarship programmes, as described above.

There is also hard data from sectors where the question has been studied, such as scientific research, charitable grant-making and philanthropy.

• NIH, United States. In July 2025 the US National Institutes of Health, the world’s largest public funder of biomedical research, found that individual investigators had submitted more than 40 AI-enabled applications in a single round. It now caps applications at six per investigator per year. [7]
• Twelve research funders. A 2026 study of twelve funders, including the Wellcome Trust, the Australian and European Research Councils and the Royal Academy, found that applications rose 57 per cent between 2022 and 2025, forcing ‘largely arbitrary choices’. [8]
• Horizon Europe. Calls in 2025 for grants, fellowships and scholarships drew 80 per cent more proposals than in 2024. Denmark’s science minister spoke of public research foundations being ‘run over by AI applications’. [9]
• European Research Council. Panels that once handled 50–150 proposals now face more than 250. [10]
• UK grant-making trusts. The G C Gibson Charitable Trust closed open applications after receiving more than 600 in one round, saying AI had ‘ruined the process’ [11], precisely the retreat into closed networks that IVAR warns against. [12] New Philanthropy Capital reports that UK funders have seen applications double or triple over the past 24 months, and notes a retreat to known recipients and databases of past applicants at the expense of open calls.

The challenge of homogenisation

The problem is not merely volume. It is homogenisation.

When large numbers of applicants draw on the same underlying models, trained on an overlapping body of successful past proposals, strategy documents and evaluation reports, the resulting applications converge.

• Funders report applications that read the same. In the Charity Digital Skills Report 2026, 34 per cent of the UK trusts and foundations surveyed said they were seeing more repetition and similarity across applications. One funder reported that answers were ‘more often than not the same’, and another feared that all applications would become ‘very high quality’ and impossible to choose between. [13] The Paul Hamlyn Foundation now tells applicants that ‘as use of AI increases, we are more able to identify similarities in language and the points being made in applications’, and that this ‘is likely to disadvantage an application’. The National Lottery Community Fund warns likewise that generic AI content ‘may disadvantage your application’. [14]

Distinctive institutional voices, hard-won operational lessons and real differences in comparative advantage are flattened into a shared register of competent but generic language. Fund managers are left trying to rank a set of documents that look similarly strong and that may bear little reliable relationship to the applicant’s actual ability to deliver.

The challenge of unequal access and the distortion of choice

AI capability is not evenly distributed. Larger international organisations, well-resourced intermediaries and applicants based in well-connected hubs can afford better models, more sophisticated prompting, dedicated proposal teams and, increasingly, fine-tuned systems trained on their own institutional knowledge.

Smaller, newer or more remote organisations, precisely the actors that the UK government’s SME support and localisation agendas seek to reach, face higher barriers. They may use free or lower-tier tools, produce less polished outputs, or simply lack the time and expertise to compete on narrative quality. The result is a selection process that risks rewarding fluency in AI-assisted proposal writing rather than proven delivery capacity. A growing body of evidence supports this.

Smaller and local organisations are disadvantaged.

• The narrative form was already exclusionary. Peace Direct’s 2021 global consultation on decolonising aid found that working-level English dominates the sector ‘to the near exclusion of any other’ language, that local actors without English, French or Spanish may be ‘excluded from the research process’, and that funding rewards ‘gaming the system like submitting perfect proposals’, a skill generative AI now puts within reach of anyone. [15]
• The UK Charity Digital Skills Report 2026, drawing on a record sample of 807 charities, warns of ‘deepening inequalities’ unless funders support AI adoption among smaller and marginalised organisations, and records funders’ own concern that AI-era processes could exclude very small organisations with no digital footprint. [16]
• Detectors penalise non-native English. A study of seven widely used AI-text detectors found that, on average, they wrongly flagged 61 per cent of essays by non-native English speakers as AI-written. [17]
• Local organisations lack AI infrastructure, policies and training. A 2025 global study of some 2,500 humanitarian workers found a built-in structural disadvantage: 80 per cent of local NGOs had no AI policy and 64 per cent provided no training. Although individual use of AI exceeded 91 per cent, the study located the gap not in individual fluency but in organisational capacity to deploy AI systematically.

These distortions compound existing structural advantages. Organisations that already possess strong written records, established relationships with donors and in-house proposal expertise can amplify those assets further with AI.

Organisations whose comparative advantage lies in proximity to communities, informal networks or operational agility in difficult environments may be systematically disadvantaged if they cannot translate that advantage into the expected narrative form. Over time, the portfolio of funded partners narrows towards those best able to generate high-scoring text, not those best placed to achieve results.

The challenge of tinkering: why incremental fixes are not enough

Many fund managers are already experimenting with countermeasures: requiring declarations of AI use, imposing word limits, adding oral presentations, or using AI tools themselves to screen for plagiarism or stylistic anomalies. These measures can raise the cost of pure fabrication and may deter the most blatant misuse. They do not, however, restore the reliability of narrative assessment as the primary basis for selection.

Declarations are difficult to verify and easy to game. Stylistic detection tools lag behind the models they are meant to police and generate both false positives and false negatives. Most importantly, none of these fixes addresses the deeper problem: that the content of a proposal remains a claim about capacity rather than evidence of it. In an environment of constrained administrative capacity, itself a product of the budget and staffing cuts described earlier, fund managers cannot compensate for weak primary evidence by simply spending more time reading better-written documents.

First-generation countermeasures: declarations, reviewer bans and rationing

• Bans wholly AI-generated applications but tells assessors not to speculate on AI use. [18]
• Australian Research Council. Limits assessors’ AI use to grammar and requires institutional certification of authorship. [19]
• UKRI, September 2026. Facing very high application volumes, UKRI will trial AI-assisted assessment, the opposite direction to the one next week’s blog recommends. [20]
• Partial randomisation. The British Academy, the Swiss National Science Foundation and the Volkswagen Foundation now use lotteries among proposals that pass a quality bar. [21]

The nightmare scenario is not science fiction. It is the logical endpoint of current trajectories: applications generated by machines, screened and scored by machines, and awarded on the basis of textual patterns that may have little connection to operational reality. In a sector whose mandate is human development, such an outcome would represent a profound failure of stewardship.