What Effective AI Capability Investment Looks Like in Low Resource Research Settings

By Alejandra Carriero @ 2026-08-20T19:58 (+1)

Notes from three workshops, and the evaluation that should follow

Most efforts to widen access to computational research in the Global South treat hardware as the sole barrier. The solution? Donating machines, providing connectivity, offering compute capacity. Having taught machine learning and computational neuroscience in Malawi and Nigeria, I think the problem has not been properly addressed, and this mismatch matters more now than it did three years ago, because the tools that could close the gap are changing faster than the people who need them the most can find out.

I have taught on a Sussex AI funded outreach course in Malawi in March 2025, and at BioRTC in Yobe, Nigeria in July 2025 and June 2026, funded by the Simons Foundation. Out of those experiences I co-founded ISAID (International Sussex AI Development), an initiative to scale this kind of teaching and to mentor students with the objective of giving them the skills to lead the courses themselves. What follows comes from that work. (For the sake of honesty, as ISAID is working to deliver the kind of intervention I argue for here, I am not a neutral observer of whether this work deserves funding.)

Who this is about

The students I taught do not lack motivation or ability. They are PhD holders, lecturers, undergraduates, master's students, all sharing real biological intuition and innovative research questions. I am a firm believer in teaching intuitively, and the best way to do so is by meeting the students at their level. By talking with them, people coming from all over Nigeria, Cameroon, Kenya, and South Africa, I learned that they mostly share the same issues. Early career researchers earn under $200 a month, fund their own research materials, and work on hand me down computers from the first world, with little to no support from their supervisors. This is not only something that I observed. A small qualitative study of participants at a pan-African training workshop identified a lack of mentorship, a lack of funds, and a lack of research and writing skills as the major challenges of an early research career (Kumwenda et al., 2017). Writing about early career researchers in Africa with a focus on Nigeria, Shinkafi (2020) describes the same pattern from the inside, where most early career researchers receive little training or mentorship on taking up their post, institutional research funds are capped at amounts insufficient to run a laboratory that is not already established, the larger national research fund is not designed for them and requires a seniority they do not yet hold, and laboratory space, chemicals and reagents are frequently unavailable. Even where the larger funds are accessible, the ceiling is around £25k over three years, which is not a budget that supports breakthrough work. The structural consequences are visible at a continental scale, because Africa has roughly 236 researchers per million people, against a global average near 1516, and 4240 per million in Europe and more than 4800 in the United States, and the continent contributes a small share of global scientific output, variously estimated at between roughly 1% and 3% depending on the database and field (Farooq and Nganhane, 2026). A two decade analysis of the continent's neuroscience output found the same pattern from the funding side, with international sources predominating over domestic ones in most African countries (Maina et al., 2021). Students notice this early, and some decide against a research career in the life sciences once the financial reality becomes clear.

Their curricula usually covers mathematics in the first undergraduate year, but rarely programming, and rarely in a way that connects the two, on the widespread and false assumption that people who choose to study life sciences are not interested in more technical subjects. It was inspiring, witnessing so many people choosing to pursue academic research in these conditions, and working so hard to get any understanding of computational skills from zero, as a way to improve the scientific rigour of their work and possibly access opportunities they would otherwise be locked out of. It was very clear that what they lack is not talent or drive. It is money, time, and awareness.

The financial barrier is straightforward. Subscriptions to proprietary models are out of reach on those salaries, and the open weight alternatives that would cost nothing to license require hardware they do not have. The time barrier is less obvious but just as binding, because learning to program independently requires spare capacity, and if you are working to survive while self funding your experiments, you do not have months to spare. The third barrier is the one I rarely see addressed in practice, and personally, I think it is the most important. Many of these students either don't know these tools exist, or they don't know the extent of what they can do, or how to use them for experimental design and analysis. There is mostly a gap in workflow and judgement, as well as one in awareness. Coding remains, in their minds, a large and intimidating obstacle that can only be overcome by being educated abroad. This belief could have been accurate five years ago, but it is becoming less accurate every month, and almost nobody is telling them. This is a barrier also identified by the African Union's own Continental AI Strategy, which lists limited awareness of AI among the workforce as the biggest barrier to adoption in the public and private sectors (African Union, 2024). (That finding concerns workforce adoption rather than research capability, so it supports the shape of the problem rather than my specific case.) The diagnosis exists at a continental level. What has not followed is a corresponding shift in where capability funding actually goes.

The moment I understood this

In a workshop on computational behaviour that I taught alongside Prof Marta Moita, from the Champalimaud Foundation, she showed the students the pose estimation analysis she uses to address her questions, using SLEAP, a deep learning tool that tracks the body parts of animals in video and is now standard in behavioural neuroscience. My job was to show them that training a model and getting estimations was not as hard as it looked, explaining the hyperparameters and trade offs when choosing and training a model. But I couldn't do it, I could not train the model on my laptop without it crashing. And I found myself explaining that if their videos were noisy or their camera moved, which is exactly what happens in real experimental setups, they would need a GPU, or they would need to leave their laptop untouched for days and hope it finished. I know this problem well. I built my PhD experimental platform on a ten year old computer with minimal memory, and designing it to run computer vision and present stimuli simultaneously without exhausting memory shaped every technical decision I made. I recognised what we had just done. We had shown a room of people with real experimental setups how research is done by people with the finances to fund the compute power. It was useless to them.

When we ran the workshop again, I changed the approach. I wrote a lightweight script using OpenCV that measures time spent in regions of interest, processing only what happens inside those regions, which is the same trick that made my own maze viable on weak hardware. It answers a large share of the questions these students actually have about their data, and it can run smoothly on their machines. The thing about the script is that I did not write it by hand. I wrote a prompt, and I showed them that if they know what to ask for, they can get a script tailored to their own experimental question too. By the end, students were using it to answer their own research questions. The understanding that AI could be used to get the students beyond the technical barrier of coding, led to another member of the team, Juan Castiñeiras (also from the Champalimaud Foundation) to deliver a workshop on how to use LLMs to address scientific questions, by showing the students how to prompt to address novel and “intimidating” questions in an academically rigorous way. This then evolved into the faculty developing Google Colab notebooks to aid the students in their projects, by providing a guided scaffold that would allow them to develop the mathematical and statistical intuitions required to understand what to ask and what the output should look like, in order to write prompts to write the code and be able check the validity of the outputs.

Why the window is closing

That experience convinced me of something I haven't seen in the governance conversation. The skill that gates entry to computational research is shifting from writing code, to prompting, to directing agentic systems. Each step lowers the entry bar for exactly the people I taught, because it rewards knowing what question to ask, which they have, over syntax fluency, which they lack.

This is not only my impression from a classroom. In a controlled experiment with 95 professional programmers, the group given an AI coding assistant completed the task 55.8% faster than the control, and the authors report larger gains for participants with less programming experience, though that estimate is imprecise (p = 0.06) and they present it as provisional (Peng et al., 2023). A natural experiment exploiting Italy's temporary ChatGPT ban found productivity effects concentrated among developers with the shortest track record on GitHub (Bonabi et al., 2025). Both point the same way, yet both studies focussed on working developers. Peng's participants averaged six years of professional coding, and Bonabi's novices are still active GitHub contributors. Nobody in either sample starts where most of the students I taught start, which is zero. Extending that gradient down to people who have never programmed is an extrapolation rather than a finding, and it is the extrapolation I would most like to test.

But this only works if people know it is happening. Right now, the shift is invisible from where these students sit, so well resourced institutions adopt agentic workflows, while under resourced ones are still afraid of the terminal. The gap does not stay the same size, it widens, and it widens the fastest at the moment it could most easily have been closed. The mechanism here is relative rather than absolute. Tools are getting easier for everyone, and that on its own should narrow the gap. What widens this gap is the difference in adoption speed. The institutions that already follow this field are restructuring their workflows now, while the ones that do not are still waiting to hear that anything has changed. The capability floor is rising, and the distance from it is rising faster.

The compute divide, and why national schemes do not reach these researchers

There is an established literature on the concentration of computational capacity in machine learning research, generally described as the compute divide (Ahmed and Wahed, 2020), and more recently framed as a division between a Compute North that hosts the compute capacity relevant to AI development and a Compute South whose compute is more relevant to deployment (Lehdonvirta, Wú and Hawkins, 2024). The policy response has largely been national. The United States created the National AI Research Resource and the United Kingdom created the AI Research Resource, both intended in part to reduce the effect of limited compute access on academic research, and both justified in terms of national competitiveness and domestic talent development (Besiroglu et al., 2024; UKRI, 2026). I am not against these schemes, and I would use one if I could. But it's worth highlighting their framing, because eligibility is tied to being based in the country that funds them, which is a reasonable position for a national research budget to take, and that also means that the researchers I taught fall outside them by design rather than by oversight. When the main policy response to the concentration of computational capacity is itself organised around national boundaries, the researchers furthest from that capacity are the least likely to be reached by it. This is why I think the more useful question, for them, is not how to get frontier compute into their hands, but what can be done well with what they already own.

What effective investment would look like

If the barrier is money, time, and awareness, rather than hardware alone, the interventions follow. Educator training is the move I would prioritise. Teaching one lecturer who teaches hundreds of students compounds in a way that a workshop for thirty does not, and it is the piece that outlasts the visit. I do not have a cost per trainer, an estimate of how many students each one reaches in practice, or a comparison against alternative uses of the same money. Generating those numbers is part of what the evaluation I am arguing for would produce. And this is already African Union policy, rather than my suggestion alone. The strategy calls for training of trainers to ensure effective transfer of AI knowledge to students, and for mechanisms to upskill teachers and faculty in AI benefits and risks.

The clearest example I know of is the institution I taught at, BioRTC, founded by Prof Mahmoud Bukar Maina, who trained in the United Kingdom, at the University of Sussex, and in 2021 established a research and training centre in Yobe State with support from the Yobe State Government, Yobe State University and TReND. In 2025 he became the first Nigerian institution-based neuroscientist to win a Wellcome Trust Career Development Award. The centre now trains scientists from across the region, and hosted the courses I helped deliver. One researcher supported to train abroad and then supported to build at home has produced a training pipeline that no single visiting workshop could. That pipeline did not appear from nothing either. Prof Maina founded the outreach programme at TReND in Africa, the Sussex based, volunteer run charity founded in 2011, that has run low cost, open hardware biomedical training courses across the continent ever since. The compounding I am talking about is visible in that chain, where an organisation trains a researcher, the researcher builds an institution, and the institution now trains others.

The second example is one I can speak to directly, because it is one of the courses that has been self-replicating. After the datalogging and machine learning course delivered by me and two colleagues at LUANAR in Malawi in March 2025, we spent a year tutoring and mentoring the students who had shown the ability and the motivation to take the work on themselves. This year the course ran again at LUANAR, delivered locally, entirely by those students rather than by us. That is a single case and I am not presenting it as proof of anything, but it is what the compounding looks like when it works, and it took a year of follow up rather than a week of teaching to produce.

None of this is happening in a vacuum. TReND's CaMinA course, run in partnership with host universities across the continent, teaches computational neuroscience and machine learning specifically. The Deep Learning Indaba has convened Africa's machine learning community since 2017 (Deep Learning Indaba, 2026) and now reaches dozens of countries through its IndabaX satellite events, Masakhane has built a distributed research community around African language processing since 2019, and organisations including Data Science Africa and the Lacuna Fund support training and dataset creation across the continent. This work has been happening, but as the dominant model remains the convening and the workshop, which reaches people who already know these communities exist, almost none of this activity is evaluated in a way that would tell us which parts of it produce durable capability.

Investment in open models that run on low power hardware matters more than investment in frontier access. The relevant question is not whether the best model is available, but whether a good enough model runs on a decade old laptop. Awareness and workflow literacy should be treated as a deliverable in their own right. So not merely tool access, but knowing what these systems can do for experimental design and analysis, and knowing what to ask.

And I would argue against reflexively teaching the heavyweight standard. You do not need to run SLEAP to answer every question. Creative researchers with the right tools can make do, and often the lightweight solution answers the actual question. What does not work is donated hardware without training, one off workshops with no follow through, and access schemes gated by institution or nationality, which leave out precisely these researchers.

The policy architecture for much of this already exists on paper. The African Union adopted its Continental Artificial Intelligence Strategy in July 2024, with AI skills and talent among its focus areas (African Union, 2024), and in April 2025, at the Global AI Summit on Africa in Kigali, the Africa Declaration on Artificial Intelligence was signed by fifty-four signatories, including fifty-two African countries, the African Union and Smart Africa, committing to talent development, open datasets and sovereign computing infrastructure. UNESCO has committed to training more than fifteen thousand civil servants and five thousand judicial personnel (UNESCO, 2025). The problem is not the absence of stated commitments. The strategy calls on the African Union Commission to develop innovative financing mechanisms and mobilise resources to build AI capabilities and skills, which is an acknowledgement that the money is not yet in place. Ambition is not the missing input. Money, and the choice of what to spend it on, are.

What is not yet established

There is a reservation that applies to my own argument, and it is a documented one rather than a private worry. Access without capability building can produce dependence rather than capability. I have watched a student refuse to accept textbook information about action potentials because ChatGPT disagreed. Research on novice programmers using generative AI describes exactly this double edge, finding both real benefits and a gap between stronger and weaker students that can widen rather than close (Prather et al., 2024), and a systematic review of AI assistants in software development reports productivity gains alongside concerns about cognitive offloading(Mohamed et al., 2026). Bonabi et al., which I cite earlier with evidence that these tools help the least experienced most, reads its own result the other way round, with the authors attributing the productivity drop among novices during the ban to possible over-reliance, and noting that they had not recovered four weeks after access was restored. Gains in skill acquisition in that study accrued to intermediate users rather than to novices. These findings are not directly against my argument, but the tension is worth stating plainly. Prather et al. study novice programmers inside structured courses, where instruction is available and generative AI can substitute for the effort that produces understanding. The researchers I taught have no such instruction available at all, so the counterfactual is not weaker engagement with a course, it is no computational work whatsoever. My expectation is that scaffolding helps where nothing else is on offer and stratifies where it competes with teaching. I want to be clear that this is a hypothesis about population differences rather than a finding, and it is the first thing I would like to test. So the case is not for access as an unqualified good. It is for a particular kind, where the tool is scaffolding toward understanding a student could eventually reach alone rather than a substitute for it. How to distinguish the two in practice is, as far as I can tell, still an open question.

Three further questions remain unanswered by the existing evidence, and they are the ones I would want to work on. Whether short term gains persist beyond the workshop. Whether they translate into publications, into careers, and into researchers staying in science rather than leaving for better paid work elsewhere. And whether AI assisted analysis by researchers without foundational computational understanding produces good science or merely fast science.

These are answerable. They need longitudinal follow up of participants, comparison across intervention types, and honest measurement of research output quality rather than participant satisfaction. Very little of that evidence currently exists, which is itself the finding. I should apply this to myself before applying it to anyone else. Of my three courses, only the Malawi one has been followed through to the point where I can say what happened afterwards, and even there I have no measurement beyond what I observed. It is precisely the kind of evidence I am arguing the field has too much of. That is not a reason to discount it, but it is the reason I want to run the follow up study rather than simply assert the conclusion. Capability investment in low resource settings is being made largely on intuition, and intuition, as I discovered in that workshop, is not always a reliable guide.

References

African Union (2024) Continental Artificial Intelligence Strategy: Harnessing AI for Africa's Development and Prosperity. Addis Ababa: African Union Commission. Adopted by the AU Executive Council, 45th Ordinary Session, 18-19 July 2024, Accra. Available at: https://au.int/sites/default/files/documents/44004-doc-EN-_Continental_AI_Strategy_July_2024.pdf

Ahmed, N.M. and Wahed, M. (2020) 'The De-democratization of AI: Deep Learning and the Compute Divide in Artificial Intelligence Research', arXiv:2010.15581.

Besiroglu, T., Bergerson, S.A., Michael, A., Heim, L., Luo, X. and Thompson, N. (2024) 'The Compute Divide in Machine Learning: A Threat to Academic Contribution and Scrutiny?', arXiv:2401.02452.

Bonabi, S., Bana, S., Gurbaxani, V. and Nian, T. (2025) 'Beyond Code: The Multidimensional Impacts of Large Language Models in Software Development', arXiv:2506.22704.

Deep Learning Indaba (2026) Our Mission. Available at: https://deeplearningindaba.com/about/our-mission/

Farooq, H.O.M. and Nganhane, I.D.N. (2026) 'Poor pay is holding back Africa's biodiversity research and reducing its contribution to global science', The Conversation, 19 May.

Kumwenda, S., Niang, E.H.A., Orondo, P.W., William, P., Oyinlola, L., Bongo, G.N. and Chiwona, B. (2017) 'Challenges facing young African scientists in their research careers: a qualitative exploratory study', Malawi Medical Journal, 29(1), pp. 1-4. doi:10.4314/mmj.v29i1.1

Lehdonvirta, V., Wú, B. and Hawkins, Z. (2024) 'Compute North vs. Compute South: The Uneven Possibilities of Compute-based AI Governance Around the Globe', Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), pp. 828-838. doi:10.1609/aies.v7i1.31683.

Maina, M.B., Ahmad, U., Ibrahim, H.A. et al. (2021) 'Two decades of neuroscience publication trends in Africa', Nature Communications, 12, 3429.

Mohamed, A., Assi, M. and Guizani, M. (2026) 'The Impact of LLM-Assistants on Software Developer Productivity: A Systematic Review and Mapping Study', ACM Transactions on Software Engineering and Methodology. doi:10.1145/3809494.arXiv:2507.03156.

Peng, S., Kalliamvakou, E., Cihon, P. and Demirer, M. (2023) 'The Impact of AI on Developer Productivity: Evidence from GitHub Copilot', arXiv:2302.06590.

Prather, J., Reeves, B.N., Leinonen, J., MacNeil, S., Randrianasolo, A.S., Becker, B.A., Kimmel, B., Wright, J. and Briggs, B. (2024) 'The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers', Proceedings of the 2024 ACM Conference on International Computing Education Research, vol. 1, pp. 469-486. doi:10.1145/3632620.3671116

Shinkafi, T.S. (2020) 'Challenges experienced by early career researchers in Africa', Future Science OA, 6(5). doi:10.2144/fsoa-2020-0012.

UKRI (2026) AIRR compute opportunity: AI open access. Swindon: UK Research and Innovation.

UNESCO (2025) AI in Africa: UNESCO unveils new solutions for its development at the G20. Paris: UNESCO.