Where we are with GenAI and MERL? Back in November 2024, I wrote a post outlining some trends and learning from 2024. Picking up from there, here are some key trends and happenings worth noting now:
The world (or maybe just the US?) is still mostly on fire.
The year 2025 was one more in a long line of bad years for those who care about things like democracy, justice, and the planet. Big Tech and AI are linked with degradation of all three.
Amongst many other things occurring this year at the global level, in the US we saw Big Tech cozying up to the current administration, a move from AI safety goals to “AI domination” goals, and continued efforts to kill AI regulation — including at the State level.
We also saw the creation of the “Department of Government Efficiency (DOGE),” initially headed by tech billionaire Elon Musk, which oversaw the decimation of USAID and huge impacts on the INGO and NGO ecosystem, affecting thousands of people and critical programs and services around the world.
The most recent fallout from this that I have read about is the Kenyan Government selling its citizen health data, just one more example of the new US policies we are seeing that will benefit Big Tech at the expense of others.
By the end of the year, we heard:
- Rumors that OpenAI was priming the US Government for a bailout in case the company stretched itself even further beyond its financial means,
- EU’s plans to walk back some of their data protection and AI regulations due to US pressure, and
- Google’s plans to create data centers in space.
- Pushback on data centers due to noise (and other kinds of) pollution, climate, and anger over energy costs being passed on to local communities is growing in different parts of the world, including Latin America, Asia, and in the US, even among conservative voters.
Many others will write about all this in their recaps of 2025, so I won’t go into any greater depth. It was certainly a year in terms of the AI industry and its seeming merger with the US Government.
AI continues to shift and change at a furious pace.
And artificial intelligence is getting better at some things. Back in July, for example, I was doing key informant interviews for a research project on AI adoption.
- One interviewee said their organization had started using Copilot, but found it really disappointing: “The juice is not worth the squeeze.” On a call with this same person a couple of weeks ago, they expressed excitement about what their internal Copilot instance can now do for document synthesis.
- On a call with a network of organizations in the EU to discuss AI adoption, two organizations shared how they have created custom libraries using their internal documents, and they were quite pleased with how they could use AI for queries now.
- At the recent Global Digital Health Forum, people shared examples of AI for diagnostics, support for community health workers, and family planning education.
As organizations experiment, some are finding real utility and starting to share more about actual AI applications as opposed to potential AI applications. Having said that, there are still issues with GenAI missing data or adding its own data (the so-called hallucinations), which have proven difficult to overcome.
Subscribe Now to MERL Tech Initiative
At MERL Tech Initiative (MTI) we still like to think about AI applications in three main buckets: backend-focused, frontline worker-focused, and community focused, and we are starting to see more realworld examples and learning about all three.
AI still requires internal capacity and good, clean data.
Organizations that are able to effectively use AI for MERL (and other back-office functions) seem to be those that have in-house data science or AI capacity and those who are able to invest time in data hygiene and coding/tagging, and cleaning up their data before layering in AI.
This report on how foundations and NGOs are thinking about and using AI from the Center for Effective Philanthropy gives a great overview of where implementing organizations are with AI adoption.
Grace Lyn Higdon and and I wrote a paper for The Foundation Review’s Special Edition on Foundations and AI that offers a snapshot in time of where Foundations are. (The entire edition is open source! Do check it out.)
We found four main archetypes of AI adoption (see the table below): the Curious, the Doers, the Dreamers, and the Skeptics. When testing these archetypes at a few sessions, they have resonated.
- Many organizations are still at the Curious stage.
- Doers are working systematically, starting to show their work more publicly, and developing systems, tools, and greater internal capacities and governance processes.
- There seem to be fewer Dreamers in the mix, but we have heard ideas such as the creation of founder bots who can govern their organizations in perpetuity, and those who have very high hopes for where AI can take their work.
- Skeptics are out there as well, with serious concerns about where AI is taking us and whether we want to go there.
Perverse incentives drive failures and poor choices.
A recent article by Priyanka Lakpattu Vasudevan lays out how the wider political context, its influence on AI regulation, and “AI theater” contribute to even large enterprises with unlimited resources failing to effectively use AI despite huge investments.
The issues that big companies face are mirrored in the NGO and foundation sectors: in many cases, simple “AI adoption” is the driver without thought for the governance capacity, technical readiness, mission/business alignment, and evaluation systems that are needed to help identify where AI might be useful and what outcomes would demonstrate its utility.
When AI adoption is the goal, and the overarching planning and systems are not there, AI tends to fail.
This behavior is happening partly due to AI hype becoming the basis for a speculative economy, with tech hype serving as “a political instrument that concentrates power in the hands of those who design and promote these futures.
According to the Financial Times, AI companies are responsible for roughly 80 percent of all gains in the stock market.
This means the American economy is becoming dependent on a single speculative sector while the benefits accrue almost entirely to the wealthiest households, who own the large majority of stocks. This creates an illusion of prosperity that masks underlying fragility while amplifying the influence of tech and financial elites.”
Should we be using AI at all?
Even if the technical issues with AI can be resolved, some organizations and individuals are still asking whether we should be using AI. And even when AI tools are potentially useful, many in the social sector remain concerned about how we can uphold ethics, transparency, safeguards, and environmental sustainability if using AI.
Over and over, I am hearing people ask: “If AI causes [insert problem, e.g., climate change], should we be using AI to solve [insert same problem]?” (Quito Tsui and I wrote about this in regard to AI and Democracy Evaluation)
These questions continue to be as relevant as ever, because we still lack a suite of tools and models designed for the social sector, that use data that is not stolen and not biased, that is built within fair labor standards, and that is possible to use in ways that do not exacerbate environmental and climate harms.
Collaborative funding efforts like Humanity AI are attempting to address some of these challenges. As we highlighted in our paper on AI Archetypes of Adoption, Skeptics tended to be organizations with gender and human rights missions, who are tech savvy, and whose reluctance to adopt AI was tied to their values.
Organizations are developing AI policies and guidance.
The number of organizations with AI policies is growing, but writing policy is still challenging because of how quickly AI changes, the many kinds of AI and AI uses organizations are grappling with, the uneven levels of AI capacity within organizations, and other challenges (see what we learned in two events on this topic on our blog: Part 1 and Part 2).
Situating AI policy in broad AI governance efforts, with considerations related to mission attainment, grantee enablement (in the case of funders), organization efficiency and individual use is an even larger yet worthwhile undertaking.
Alberto Ortega Hinojosa from Packard Foundation explained at our recent session at the American Evaluation Association Conference that since AI permeates all parts of an organization, it’s critical to go beyond just a policy and approach AI governance at the organizational level.
We still need sector-level guidance on AI for evaluation.
My biggest takeaway from attending the AEA conference was that Evaluation Associations need to have a more in-depth conversation and establish specific principles and guidance related to the use of AI within Evaluation.
Rather than portray AI as magic, a more critical take on AI within systems of power and inequity is needed.
While some presentations at the AEA did pay thoughtful attention to responsible use of AI and its potential harms alongside its potential benefits, several people approached me at the conference to express concern about the lack of attention to privacy, ethics, bias, consent, and potential harm that they witnessed in some of the AEA sessions on AI.
They had concerns that bad practices were being promoted. A few of us will be meeting in January to talk about how different regional associations might come together to work on association-level principles and guidance.
We saw exciting advances in AI evaluation frameworks.
In addition to growing AI adoption at organizations, in 2025, we saw a good deal of movement on the Evaluation of AI.
The Agency Fund launched their AI Evaluation Playbook, which lays out a 4-stage framework for evaluating community-facing AI such as chatbots. Their framework offers guidance on how to do Model Evaluation, Product Evaluation, and User Evaluation before coming to more socially focused Impact Evaluation. The framework is being tested by multiple organizations as they start evaluating AI investments.
Other interesting tools include ID Insight and the Agency Fund’s Evidential, a platform for running adaptive experiments on community-facing AI. Weval’s platform is a space for running and sharing LLM evaluations.
MTI will be involved in testing some of these frameworks to evaluate and learn more about what is working with AI challenge funds (and whether these emerging AI evaluation frameworks are fit for purpose in a space that changes so quickly).
We have been working with the GSMA Foundation to develop an AI funder toolkit, for example, that looks at considerations for selecting recipients of challenge funds, a MEL framework for assessing implementation, and how to strengthen internal and external technical capacity with relation to AI.
We’ll be sharing what we learn across these various projects with the goal of helping the wider sector continue to learn and share emerging good practices and lessons.
Let’s remember what we already know.
There is so much the sector has already learned about what makes for successful “ICT4D” – as I wrote last year, and we don’t always apply it.
Front-line workers and community-focused AI applications are still limited by infrastructure, inclusion, and accuracy challenges, and when designing AI tools, we seem to be forgetting the importance of factors like phone capacity, device memory, cost of data, availability of electricity, literacy, and the potential of voice in low-resource, low-literacy settings.
At MTI, we will be doing some work in 2026 that focuses on Formative Research and Digital Design for AI programming in hope of bringing that knowledge back and adapting it as needed to emerging AI. This will include a focus on inclusion, gender equity, access and accessibility, data protection, and safety and safeguarding.
Alternatives to Big AI are emerging.
As predicted in my 2024 post, there is more attention being paid to both investment in models in additional languages and in Small Language Models (SLMs) that can run on laptops and mobile phones.
What I feel is missing in the world of SLMs is user-friendly tools and interfaces that would allow individuals and small organizations to interact with them more easily. Many of the interfaces surrounding SLMs right now require a familiarity with coding (or just general comfort looking at poorly designed screens with tons of text on them) akin to Linux and Ubuntu and other open source platforms, which is a major barrier to uptake.
I am hoping that 2026 brings some user-friendly wrappers for SLMs.
Microsoft recently released Fara-7B, a new 7-billion-parameter “agentic” SLM designed for computer use, which allegedly matches the performance of much larger models while running locally on devices and can complete tasks on behalf of users. Several other SLMs are out there, including open source models, and I am hoping to see more of these being tested in the social sector in 2026 (with attention to security and other potential downsides).
Non-English language models are proliferating.
And funders are increasingly investing in them. For example, The Gates Foundation invested $200 billion in AI and Health in 2025 and is funding a number of African language models through the African Next Voices project.
Other projects include those like custom language models for specific projects (e.g., a model that uses Twi and focuses on maternal health), language models aimed at uptake by businesses (banks, schools, hospitals, government services), and benchmarking data sets trained in local languages and contexts that can be used to assess outputs of Big AI models.
Our paper on Made in Africa AI for MERL highlights the importance of expanding language models, and also calls out that tech/AI developers, privacy and tech justice, and MERL sectors are still largely swimming in their own lanes and should be working together on this.
A lot happened in the AI and MERL space in 2025!
2026 promises more of the same as we consolidate learning on responsible AI adoption and on ways to conduct MERL on the use of AI in multiple spheres. The external environment continues to be rapidly changing and potentially volatile, if we believe the warnings that the AI bubble is close to bursting – time will tell.
As more and more AI development happens in different parts of the world, surely the story of AI will shift and change in ways we have not imagined.
