Everywhere we turn, AI is everywhere we turn. A couple years ago, AI felt to me like the new blockchain. Ubiquitously discussed circa 2017 as the solution to potential lapses in credibility and ethics, blockchain found its niche and got on with its work; one hardly hears about it anymore outside its main use case. Initially, the dropped references to AI felt like blockchain next generation –trending, hunting for its perfect use case. But then a shift in the AI conversation – it burst the banks from favorite coffee break sidebar chat to embedded, everywhere. Everything appeared to be the AI use case, everywhere.
When did we teeter (or slosh) into true ubiquity? What were the signals? Maybe it is when my teens started discussing how their classmates were using AIs as a tutor and prompt for studies (ok use case) and as a tool to egregiously cheat (not ok), and that their main requirement in choosing a uni course of study is what will not be replaced, immediately, by AI. The signal might have been when colleagues’ emails (and, frankly, a number of my own) became perfectly stilted, always written in “professional yet warm tone.” Certainly, if one even glances at social media, two or three years ago we would have noticed a tsunami of surreal and crystalline images in our feeds, bizarre and clearly figments , some might consider rivaling anything in Salvador Dali’s imagination. What is much more frightening: now we don’t see it anymore. We don’t know what we are looking at. Riptides sit under the glimmering surface.
It was at some point in all of this, in the last six months, CGIAR’s System Council started asking us which aspect of our Advisory Bodies’ work we would be handing over to AI, be it generative AI or large language models (LLMs) within GenAI – nothing was specific, but the exhortation to explore was clear. It is a legitimate and sincere question. The ODA environment is tight. Efficiencies must be found. The entire world of research is questing for the perfect AI use cases, and answering this demand to adapt (to AI) or risk fading away. On the cusp of what is foreseen to be extensive uptake of agentic AI, we in IAES are identifying our AI use cases. Here is the very human job we need to do: we need to find the “ok” ones. The ones that make us lighter and leaner, but don’t become riptides in our work, that do not weaken our ethics, legitimacy, and rigor. So, this year in our capacity implementing independent evaluation, and supporting the mandates of ISDC and SPIA, we are embarking on a roadmap, with quick tests of AI applications. The roadmap is more a back of napkin sketch – you won’t see it published on our website. Of course, I could easily have asked AI to write it for us, but the hand of the human on the rudder is needed here.
Here is one small sliver of our work, explained in six-month increments. One of ISDC’s mandated roles is the review of pooled funding portfolios to advise System Council. Some funders would say this is the most important thing ISDC does. Everyone who prepares grants in CGIAR knows what is happening in this space. David Adam writes (Nature 645 | 25-09-25) that, “a major funding foundation in [Spain] turned to artificial intelligence (AI) for help with reviewing grant proposals.” Adam writes about the great discomfort and very human dismay among grant writers, a deep sense of unfairness and injustice. Six months later, Rees and Wilson comment in the same journal (Nature 652, 27-04-26), about the use of LLMs in grant preparation, among many and all aspects of science “Evidence is mounting of a surge in the use of generative AI across science” – including in proposals. The authors imagine a not so jolly and highly plausible world in the very near future when there is a net surge in high quality proposals, entirely written by AI agents – rendering for grant screening panels the inability, the impossibility to separate the flotsam and jetsam. Everything will look good, on paper. No worries. Let’s imagine the next Nature commentary, 31-10-26: “Progeny of the Spanish foundation’s AI from 25-09-25– available everywhere to screen every grant.” But revise the publication date: 31-10 is too late. Expect the acceleration that is glaringly apparent in AI evolution: from 6 months to 3, 3 months to 6 weeks, to 3 weeks, to….. every cycle will be faster.
So, what’s a human to do?
We carry out our roadmap in short, incremental tests – this blog series is going to tell you what we learn. We are bold and ready to trial, yet we are humble – our resources don’t allow for perfect testing. Our inputs are staff time (none of us computer engineers), corporate-issued and affordable software and platform services we already own/rent with respective AI plug-ins, and guided by the questions we can reasonably test with resources we can avail – in a field that is changing faster than humans can think. Our tests are about finding the boundary (albeit porous and malleable) between human discernment and the work of AI – and how we move from cautious co-existence with these new technologies to confident and responsible use across our mandate areas.
Look in this series for blogs from colleagues working in IAES and more broadly SPIA and ISDC, on, inter alia:
- bespoke interpretation of evidence for funders (what’s reliable?)
- analysis of qualitative Center Review data (what best supports accurate data coding?)
- reviewing proposals and strategic documents (what’s ethical?)
Alongside such technical use cases, the roadmap entails that we either adopt (when credible and affordable solutions are developed in CGIAR and beyond), or develop, operations use cases – the processes of documentation, procurement, logistics, financial management, planning, reporting, communicating that underpin our professional operations.
The chief inputs: the roadmap, internal conversation, time to develop and run tests, a bit of low-cost/free training, ongoing software as a service and secure AI subscriptions (i.e., to keep sensitive content out of AI training data and secure). The output: identify use cases that allow us in IAES, and the bodies we support, to renegotiate the tradeoff between time spent in deliberative or discernment labor versus routine tasks. We are meanwhile building up our IAES roster with AI subject matter expertise for later and more in-depth deployment. As the roadmap gets underway, we are initially focused on LLMs; as the year and technology progresses, we may be ready to take a deeper dive into what agentic AI can and cannot responsibly and affordably offer, for instance exploring a specific use case to support program evaluations more efficiently exploit quality-assured CGIAR monitoring data, in partnership with CGIAR portfolio performance unit.
Please see our next blog in the series: Thomas Griffin (IAES intern) and Swetha Ramachandran (SPIA Use of Evidence Officer) write about the application of LLMs to support CGIAR funders’ use of advice and evaluative evidence.
This series will include blogs about:
- Opening the series: What’s a Human to Do? Spotting Valid AI Use Cases in CGIAR’s Advisory Bodies
- Accelerating Use of Evidence: Bespoke Messaging for Funders Supported by AI
- AI-enabled Qualitative Data Analysis: Accurately and Quickly Working with Qual Data Sets?
- Using LLMs in Proposal and Portfolio Reviews: Credible and Ethical Applications
