Latest / Crypto Altruists: Real-World Stories of Social & Environmental Impact with Web3 / Episode 269 - Small, Local, and Sovereign: 8 Ways NGOs & Communities Are Using AI for Good Without Compromising Their Values

Transcript

Episode 269 - Small, Local, and Sovereign: 8 Ways NGOs & Communities Are Using AI for Good Without Compromising Their Values · Crypto Altruists: Real-World Stories of Social & Environmental Impact with Web3

• 0:00
  SPEAKER_00: Today, you'll discover an honest look at the three biggest concerns nonprofits have about AI, and that is energy and water use, safety and power concentrating in the hands of big tech, eight ways NGOs and grassroots movements are using AI for humanitarian impact from disaster mapping to maternal health and protecting the rainforest, what sovereign AI looks like in practice, including communities owning their data, training their own models, and running them locally, and a simple checklist any nonprofit can use before bringing an AI tool into their work.
• 0:31
  SPEAKER_00: Welcome to CryptoUltras, where we explore the intersections of Web3 and Social Good to the stories of impact-driven builders and offer actual insights to empower your journey into the world of blockchain for good.
• 0:39
  SPEAKER_00: I'm your host, Drew Simon.
• 0:41
  SPEAKER_00: Now before we dive in, please remember that this is not financial or legal advice, and always do your own research on any of the projects or investment opportunities that we discuss.
• 0:49
  SPEAKER_00: Welcome back to the CryptoUltras podcast for episode 269.
• 0:52
  SPEAKER_00: Thanks so much for being here today.
• 0:54
  SPEAKER_00: This is a solo episode, and it's one I've wanted to make for a while.
• 0:57
  SPEAKER_00: Because lately, whenever I talk with people who work at nonprofits or NGOs or grassroots organizations, the conversation almost always turns to AI.
• 1:05
  SPEAKER_00: And the feeling in the room is usually some mix of curiosity and discomfort or even dread.
• 1:10
  SPEAKER_00: Curiosity is because people can see that these tools might help overstretch teams do more, and discomfort because a lot of what they hear about AI clashes with their values.
• 1:20
  SPEAKER_00: So it's a bit of a tricky situation, and today I want to do two things to kind of unpack this.
• 1:24
  SPEAKER_00: First, I want to take those concerns seriously and give them an honest answer.
• 1:28
  SPEAKER_00: I'm not here to give you a sales pitch about how awesome AI is.
• 1:31
  SPEAKER_00: And second, I want to show you a side of AI that doesn't get enough attention, and that is organizations using it in ways that are small, local, community-owned, and designed with the planet in mind.
• 1:41
  SPEAKER_00: And before we get started, I just wanted to do a special shout out that is totally unrelated to this episode, but still pretty awesome.
• 1:47
  SPEAKER_00: As you might know, I've been working with uh Alliance for Public Health, an incredible organization based out of Ukraine, on um helping implement Web3 tools into their work to help them raise funds and grow their impact.
• 1:58
  SPEAKER_00: Uh, excited to announce that they just launched on Giveth, and so they have a campaign going there now to support blind veterans uh who have been impacted by the war in Ukraine.
• 2:07
  SPEAKER_00: Uh so great organization, great cause.
• 2:09
  SPEAKER_00: I will uh share the link in the show notes if you would like to contribute or kind of spread the word.
• 2:14
  SPEAKER_00: Anything helps uh for this important cause.
• 2:16
  SPEAKER_00: Uh so with that, uh let's start with the hard part of this conversation around AI, and that is the three big concerns.
• 2:25
  SPEAKER_00: So let's start with concern number one, which is probably the biggest, and that is the environment.
• 2:29
  SPEAKER_00: Uh, if you're worried about the environmental cost of AI, you're certainly right to be.
• 2:33
  SPEAKER_00: I definitely am.
• 2:35
  SPEAKER_00: The International Energy Agency projects that electricity consumption from data centers will roughly double from 485 terawatt hours in 2025 to 950 in 2030, which is a lot.
• 2:46
  SPEAKER_00: Uh, this would reach about 3% of global electricity demand.
• 2:50
  SPEAKER_00: And electricity consumption from AI-focused data centers is growing even faster, jumping by 50% in 2025 alone.
• 2:57
  SPEAKER_00: And it's not just energy either.
• 2:58
  SPEAKER_00: These data centers use a ton of water, with large data centers using up to 5 million gallons per day or up to 1.8 billion annually.
• 3:06
  SPEAKER_00: And this is a big problem in a world that's facing more droughts and climate emergencies.
• 3:09
  SPEAKER_00: We do not have an abundance of water, that's for sure, and we need to use it wisely.
• 3:13
  SPEAKER_00: So I'm not going to pretend uh otherwise.
• 3:15
  SPEAKER_00: We need far more renewable energy and sustainable water systems powering this infrastructure, uh, without a doubt.
• 3:20
  SPEAKER_00: We also need real efficiency standards and honest reporting from the companies that are building them.
• 3:25
  SPEAKER_00: So that pressure has to keep coming, including from our sector.
• 3:29
  SPEAKER_00: The second concern is safety.
• 3:30
  SPEAKER_00: And if you've listened to our solo episode on separating hype from hope, you know that I do take AI safety very seriously.
• 3:36
  SPEAKER_00: These are powerful systems, and even many of the people building them are calling for stronger ground rails.
• 3:41
  SPEAKER_00: That said, I often question the motives of those behind large AI companies calling for slowdowns.
• 3:46
  SPEAKER_00: Some would argue that it's them wanting more corporate capture on the AI ecosystem by making it harder for new entrants, or to hype up their models in advance of their upcoming big IPOs.
• 3:55
  SPEAKER_00: Because let's be real, if they really wanted to slow down, they could.
• 3:57
  SPEAKER_00: They don't need to be told by the government to do it, they could just do it on their own.
• 4:01
  SPEAKER_00: But, anyways, um so we might disagree with that, but I do agree that there needs to be better guardrails and slowdown, without a doubt.
• 4:07
  SPEAKER_00: Um, but anyways, talking about uh safety, we got misinformation, deep fakes, bias outputs, the use of AI and weapons, um, and harms we haven't anticipated yet.
• 4:16
  SPEAKER_00: And, you know, these are all risks that very much worry me.
• 4:19
  SPEAKER_00: So for organizations serving vulnerable people, the idea of move fast and break things, which has been kind of a common mantra of Silicon Valley, is not an option.
• 4:27
  SPEAKER_00: And to be frank, I never really liked that statement.
• 4:30
  SPEAKER_00: Um it seems to me kind of like a new form of imperialism, like we know it's best, so it's okay if we break things in service of the greater good.
• 4:36
  SPEAKER_00: That's kind of what I feel when it comes to the AI accelerationist movement as well, is this concept of like, you know, don't worry, we know it's good, so we're just gonna break everything on the way, and there might be risks, but don't worry about it.
• 4:45
  SPEAKER_00: Um I don't like it, and you know, for me, do no harm has to always come first.
• 4:50
  SPEAKER_00: And the third one is the concentration of power, and that is that a handful of enormous companies control the most powerful AI models, and they control the data centers and increasingly the data itself.
• 4:59
  SPEAKER_00: When a nonprofit feeds sensitive information about the communities it serves into one of those systems, you have to ask where that goes, who benefits from it and who's accountable.
• 5:07
  SPEAKER_00: As our guests from Open Frontier reminded us a few episodes back, technology is not destiny.
• 5:12
  SPEAKER_00: What matters is who shapes it, who benefits from it, and what rules govern it.
• 5:16
  SPEAKER_00: And right now, too much of the shaping is happening in very few hands.
• 5:20
  SPEAKER_00: So all three concerns are very legitimate, and we absolutely need to be doing more about each one of them.
• 5:25
  SPEAKER_00: I don't have an answer for you here as to how, and I wish I did.
• 5:28
  SPEAKER_00: Uh, this is something I try to wrap my head around daily and spend a lot of time thinking about because these concerns very much worry me.
• 5:35
  SPEAKER_00: So it is something that needs to be top of mind for every industry using AI, including the impact sector, as their building solutions, you know, to support their work.
• 5:42
  SPEAKER_00: Um so now that we've had that discussion, I do want to shift the conversation a bit by saying that AI isn't one thing.
• 5:48
  SPEAKER_00: And this is how I often open up conversations with folks in the nonprofit sector when talking about AI and talking about these concerns.
• 5:56
  SPEAKER_00: Because there's a world of difference between pasting sensitive community data into a giant corporate chatbot and a community training its own small model on its own data that runs on a phone in someone's pocket.
• 6:06
  SPEAKER_00: So the question isn't really whether NGOs should use AI, it's which AI owned by whom?
• 6:10
  SPEAKER_00: Who?
• 6:10
  SPEAKER_00: Whom?
• 6:11
  SPEAKER_00: I don't know, I always mess that up.
• 6:12
  SPEAKER_00: But like I was saying, which AI owned by whom, running where, and at what cost to the planet?
• 6:18
  SPEAKER_00: For each of the eight examples today, I'm gonna run it through three filters.
• 6:21
  SPEAKER_00: And those filters are who owns the data, number one, whose model is it, and what's the footprint.
• 6:26
  SPEAKER_00: And I'll be honest, where an example falls short, because there are no examples that are perfect in this case.
• 6:31
  SPEAKER_00: Um, AI and AI sovereignty is a spectrum, and so we're gonna share a spectrum of examples here that kind of meet this criteria in different ways.
• 6:39
  SPEAKER_00: So, without further ado, let's dive into some case studies.
• 6:44
  SPEAKER_00: And so the first one is an example about one of the most fundamental needs in any disaster, and that's knowing where people are when disaster hits.
• 6:51
  SPEAKER_00: Uh, think about when a flood, earthquake, or conflict hits.
• 6:54
  SPEAKER_00: Uh, those responding need maps.
• 6:56
  SPEAKER_00: They need to know where to allocate resources, where the help is needed.
• 6:59
  SPEAKER_00: So, and to do this, they need to know where the buildings, roads, and waterways are.
• 7:02
  SPEAKER_00: Uh, unfortunately, in many parts of the world, these maps don't exist in enough detail, at least not yet.
• 7:07
  SPEAKER_00: The humanitarian open street map team, known as HOT HOT, has spent years mobilizing volunteers to map these places.
• 7:14
  SPEAKER_00: They built a really cool tool called FAIR, that's F-A-I-R.
• 7:17
  SPEAKER_00: The F stands for Freedom and Free Open Source Software, and the R stands for resilience and responsibility.
• 7:22
  SPEAKER_00: The AI in the middle stands for AI, obviously.
• 7:25
  SPEAKER_00: And what makes FAIR a cool case study is who is in control.
• 7:28
  SPEAKER_00: It's a free open source service that lets mapping QMU members create and train their own AI models for their own region or humanitarian need.
• 7:35
  SPEAKER_00: And there's a really important reason why they built it this way.
• 7:39
  SPEAKER_00: Most of the existing data that could teach an AI to recognize buildings was created in the global north, and it doesn't reflect the local context where HOT's mappers work.
• 7:47
  SPEAKER_00: For example, an AI trained on suburban rooftops in one part of the world struggles with informal settlements in another.
• 7:53
  SPEAKER_00: So HOT put the training into the hands of local communities themselves, which I think is pretty rad.
• 7:58
  SPEAKER_00: And in this case, the results have been great.
• 8:00
  SPEAKER_00: Mappers typically map between 1,000 and 1,500 buildings a day on their own, but in an AI assisted pilot, that capacity nearly doubled to 2,500 to 3,000 buildings per person per day.
• 8:10
  SPEAKER_00: And crucially, and this is very crucial, humans stay in charge.
• 8:13
  SPEAKER_00: Volunteers use the AI suggestions as an aid to map faster and more accurately, with human judgment validating every element rather than relying on automated mass imports.
• 8:22
  SPEAKER_00: So, looking at the scorecard of our first example, the data is open and governed through the OpenStreetMap community check.
• 8:29
  SPEAKER_00: The models are trained by the community check, and these are small, focused computer vision models, not giant ones, so pretty strong match on all of our filters.
• 8:37
  SPEAKER_00: Good work, OpenStreetMap team and community.
• 8:40
  SPEAKER_00: Going on to the next example now comes from the far north of, and I'm sorry in advance, I'm probably not going to say this right.
• 8:45
  SPEAKER_00: There's going to be lots of names I'm not going to say right throughout this episode.
• 8:48
  SPEAKER_00: And that is Eteoroa, New Zealand.
• 8:50
  SPEAKER_00: Again, definitely not right.
• 8:51
  SPEAKER_00: And it might be my favorite case study in this episode.
• 8:54
  SPEAKER_00: So Tehiku Media is a Maori broadcaster dedicated to revitalizing the Mayori language.
• 9:00
  SPEAKER_00: Their inspiration comes from a challenge many indigenous communities face, and that is that big tech companies weren't building good language tools for them.
• 9:06
  SPEAKER_00: And when those companies did collect indigenous data, the communities rarely saw any benefit from it.
• 9:11
  SPEAKER_00: So they built their own.
• 9:12
  SPEAKER_00: Tihiku Media developed a speech recognition system that achieves 92% accuracy for Tireo Mayori, the Mayori language.
• 9:20
  SPEAKER_00: The project is built around the principle of Ketia Kitanga, or guardianship, with Mayori control over how the language data is used and a focus on ensuring value flows back to the community.
• 9:30
  SPEAKER_00: So the way they gathered the data is pretty cool.
• 9:32
  SPEAKER_00: They started with elders, then opened it up to the whole community.
• 9:35
  SPEAKER_00: In just 10 days, more than 2,500 people signed up to read over 200,000 phrases, providing more than 300 hours of labeled speech data.
• 9:42
  SPEAKER_00: And the philosophy behind it is something I think we can all learn from.
• 9:46
  SPEAKER_00: Since data is cared for under the principle of Katia Katanga, which I mentioned before, any benefit derived from the data flows back to its source.
• 9:53
  SPEAKER_00: And this brings me back to my episode with DG Safir on cooperatives, where we discuss the idea of guardianship, not ownership.
• 9:58
  SPEAKER_00: In other words, value that stays with the people who created it.
• 10:01
  SPEAKER_00: In that sense, I think those building AI solutions for local communities can really look to the history of cooperatives and draw a lot from them when designing these, when designing with communities.
• 10:10
  SPEAKER_00: So looking at our scorecard, the data is governed entirely by the community, which is great.
• 10:15
  SPEAKER_00: The model is their own.
• 10:16
  SPEAKER_00: The footprint isn't zero since they trained it on GPUs, but it is small.
• 10:20
  SPEAKER_00: It's a targeted model built for one specific purpose.
• 10:23
  SPEAKER_00: So it's not a general system trying to do everything and using a ton of electricity or energy in the process.
• 10:28
  SPEAKER_00: So not too shabby overall, I'd say.
• 10:30
  SPEAKER_00: Going on now to the third example, um, it's about listening to nature and stopping a crisis before it happens.
• 10:37
  SPEAKER_00: This one is really cool, and it's kind of one that seems simple but is quite complex.
• 10:41
  SPEAKER_00: And to set the context, I, you know, as we know, illegal logging is devastating to Rainforce.
• 10:46
  SPEAKER_00: Or any force for that matter.
• 10:48
  SPEAKER_00: And unfortunately, by the time satellites spot the damage from illegal logging, it's often too late to do anything about it.
• 10:53
  SPEAKER_00: So this is where uh Rainforest Connection comes in, who came up with an awesome but simple idea, and that is to just listen.
• 11:00
  SPEAKER_00: They take old Android smartphones and repackage them into recycled plastic boxes, fitted with an extra microphone, a battery pack, and solar panels.
• 11:07
  SPEAKER_00: These devices, which they call guardians, are hidden high in the forest canopy.
• 11:11
  SPEAKER_00: They capture the sounds of the rainforest and use AI to pick out chainsaws, logging trucks, people, and gunshots, then send real-time alerts to people on the ground.
• 11:19
  SPEAKER_00: And the people on the ground are often indigenous communities protecting their own land, like the Tembe people of northern Brazil.
• 11:25
  SPEAKER_00: What I love here is the footprint story.
• 11:27
  SPEAKER_00: Old phones that might have otherwise become e-waste are being powered by the sun and protecting force that absorbed carbon.
• 11:33
  SPEAKER_00: It's a really cool example of AI working with the planet rather than against it.
• 11:37
  SPEAKER_00: So looking at the scorecard, the models are small and purpose-built, and the hardware is solar-powered and recycled.
• 11:43
  SPEAKER_00: In the earlier garden systems we're talking about here, audio was transmitted to the cloud for processing rather than being analyzed entirely on device, so it's not fully local, but it is a remarkable example of low footprint design that others can draw from.
• 11:56
  SPEAKER_00: So moving on now to our fourth example, and it is about food security and about what AI can do when there's no internet at all.
• 12:03
  SPEAKER_00: So for smallholder farmers, crop disease can wipe out a seasons income pretty quickly, and diagnosing it usually means waiting for an expert who may never come or at the very least might not come in time.
• 12:13
  SPEAKER_00: Plant Village, on the other hand, is working alongside researchers and the UN's food and agricultural organization to build an app called Nuru to change that.
• 12:21
  SPEAKER_00: Nuru is an AI assistant that runs inside a standard smartphone and can accurately diagnose Cassava diseases offline with no internet connection.
• 12:28
  SPEAKER_00: Whereas most AI systems rely on the cloud, this simply doesn't work for smallholder farmers where connectivity can be poor.
• 12:35
  SPEAKER_00: So the project has reached a remarkable scale with the United Nations using the tool in 21 languages across 90 countries to help smallholder farmers tackle crop health threats.
• 12:44
  SPEAKER_00: It also proved itself in a crisis.
• 12:46
  SPEAKER_00: When desert locusts swept across East Africa in 2020, the team's tracking app accounted for 70 to 90% of locust observations in Kenya at the peak of the outbreak.
• 12:54
  SPEAKER_00: And there's a really cool human story I love here as well.
• 12:57
  SPEAKER_00: One farmer used Nuru in a way that nobody expected, and that was to identify her healthy plants so she could propagate clean cuttings.
• 13:03
  SPEAKER_00: And after one season, she increased her revenue by 55% and her yields by 146%.
• 13:09
  SPEAKER_00: So, really cool example, and looking at the scorecard here, it is researcher developed with a small model running right on the phone, which avoids sending every diagnosis to a cloud-based model, thus reducing its footprint.
• 13:20
  SPEAKER_00: The caveat here though is that this was led by a university and a global research consortium.
• 13:24
  SPEAKER_00: So rather than a grassroots group, the community's role is as users more than builders.
• 13:29
  SPEAKER_00: So that's just one small caveat there.
• 13:31
  SPEAKER_00: But overall, really cool case study.
• 13:33
  SPEAKER_00: Moving on to our fifth example, which comes from Kenya, where Jakaranda Health works to end preventable deaths of mothers and newborns.
• 13:40
  SPEAKER_00: Jakaranda runs a digital health service that supports new and expectant mothers by text message.
• 13:45
  SPEAKER_00: And they wanted AI that could respond to mothers in their own languages with nuance and care.
• 13:49
  SPEAKER_00: Off-the-shelf models weren't good enough in African languages, and they raised serious privacy concerns for health data as well.
• 13:55
  SPEAKER_00: So Jakaranda built their own, called Uliza Lama, which means Ask Lama, and it now supports Swahili, Hasa, Yoruba, Mosa, and Zulu.
• 14:03
  SPEAKER_00: Again, sorry for mispronouncing those.
• 14:06
  SPEAKER_00: And because it's publicly available, organizations can run it on their own servers, which means they keep complete control over their data, which is obviously critical for sensitive health information, unlike off-the-shelf tools such as ChatGPT.
• 14:18
  SPEAKER_00: So looking at the scorecard, the data stays in-house, which is a big win for privacy.
• 14:21
  SPEAKER_00: Check, but here's the caveat.
• 14:23
  SPEAKER_00: And it's why I included this example because I wanted to kind of show a spectrum here.
• 14:27
  SPEAKER_00: Ulisa Llama builds on Meta's Llama models.
• 14:29
  SPEAKER_00: So it still relies on a big tech foundation, and then it gets adapted and self-hosted.
• 14:34
  SPEAKER_00: So that's a trade-off for sure.
• 14:35
  SPEAKER_00: And it's a great example of how, like I said, sovereignty is a spectrum.
• 14:38
  SPEAKER_00: Jacaranda chose control over their data, even if the underlying model came from elsewhere.
• 14:42
  SPEAKER_00: And importantly, Jacaranda keeps humans in the loop with responses reviewed and edited before they're sent to mothers.
• 14:48
  SPEAKER_00: So there's a trade-off in kind of the um the outlook on that one.
• 14:52
  SPEAKER_00: Uh really cool project, but obviously um some caveats there to note.
• 14:56
  SPEAKER_00: And moving on now to our sixth example, uh, which takes the idea of small and really runs with it.
• 15:01
  SPEAKER_00: So we're going to talk about Lilapa AI, which is based in South Africa and builds language tools for Africans.
• 15:06
  SPEAKER_00: And they've challenged one of the biggest assumptions in AI that bigger is always better.
• 15:11
  SPEAKER_00: In their words, they champion smaller models as a pathway to equity.
• 15:14
  SPEAKER_00: Their model is called Incuba LM, which is named after the dung beetle, which can move 250 times its own weight, which is cool, I did not know that.
• 15:22
  SPEAKER_00: Kind of like ants, but maybe even more so.
• 15:24
  SPEAKER_00: It's a compact model that can run on regular CPUs, making it viable in resource-constrained environments, including on laptop computers.
• 15:31
  SPEAKER_00: So why is this a big deal?
• 15:32
  SPEAKER_00: Well, because a large portion of smartphone users in Sub-Saharan Africa are on entry-level devices, and internet penetration in Sub-Saharan Africa hovers around 35%.
• 15:41
  SPEAKER_00: That makes lightweight models especially valuable in environments where devices, bandwidth, and computing resources are constrained.
• 15:47
  SPEAKER_00: And here's the part I love.
• 15:48
  SPEAKER_00: They opened it up to a global challenge to make the model even leaner.
• 15:52
  SPEAKER_00: Over 490 data scientists and engineers from 61 countries took part, and they shrink the model by up to 75% while keeping its translation quality strong.
• 16:00
  SPEAKER_00: It was a great case study in open grassroots innovation at its best.
• 16:04
  SPEAKER_00: So looking at the scorecard, an African-built multilingual model designed for resource-constrained environments, which is great.
• 16:10
  SPEAKER_00: It's openly available and light enough to run on a laptop, so it checks a lot of those boxes there.
• 16:14
  SPEAKER_00: So yeah, pretty cool example overall.
• 16:17
  SPEAKER_00: Our seventh example now introduces an idea that sounds technical but is actually quite intuitive when you break it down, and that is federated learning.
• 16:24
  SPEAKER_00: Spoiler alert, we're gonna dive deeper into this in an upcoming episode with the organization we'll discuss in this episode, so stay tuned.
• 16:30
  SPEAKER_00: Anyways, normally to train an AI model, you gather all the data in one central place, but federated learning changes that.
• 16:37
  SPEAKER_00: In the basic federated learning setup, the model travels to the data, learns from it where it lives, and the raw data never leaves home.
• 16:43
  SPEAKER_00: For sensitive information, that's important.
• 16:45
  SPEAKER_00: Flock.io, and that's different than the Flock camera company, of course, is the AI strategic partner to the UNDP's SDG Blockchain Accelerator, and the projects they're supporting use blockchain power decentralized AI and federated learning to drive impact across the developing world.
• 16:59
  SPEAKER_00: And we're talking climate action and finance, healthcare and wildlife conservation, among many other things.
• 17:05
  SPEAKER_00: One example is in Mauritius and Sheyshell's, where patient trust in digital healthcare is fragile.
• 17:10
  SPEAKER_00: So the project uses blockchain and federated learning to enable real-time consent management.
• 17:15
  SPEAKER_00: What this means is it gives patients transparent control over their health data.
• 17:19
  SPEAKER_00: Another example is in India, where a verification system for rice farmers, carbon credits, engages 3,000 farmers across 2,500 hectares, helping them access carbon credit income directly instead of through expensive intermediaries.
• 17:31
  SPEAKER_00: We'll dig into all this much more deeply in an upcoming episode with Tiffany of Flock.io and the blockchain for good alliance, so stay tuned for that.
• 17:38
  SPEAKER_00: So looking at the scorecard, the approach is designed to reduce the need to use centralized sensitive data, which is a strong match for data sovereignty.
• 17:44
  SPEAKER_00: The caveat here is on footprint because federated learning spreads the computing out, but that doesn't automatically make it lower energy, so something to keep in mind there.
• 17:52
  SPEAKER_00: And now looking to our final example, which takes a bit of a different angle, and I want to be upfront about why.
• 17:58
  SPEAKER_00: Karia is a nonprofit in India that creates the datasets AI needs, especially in Indian languages that big models tend to ignore.
• 18:05
  SPEAKER_00: But they do it in a very different way.
• 18:06
  SPEAKER_00: They pay a minimum of$5 an hour and give workers de facto ownership of the data they create.
• 18:11
  SPEAKER_00: So whenever it's resold, the workers receive the proceeds on top of the wages.
• 18:15
  SPEAKER_00: Korea says it has brought digital work to more than 100,000 people with a minimum wage of$5 an hour.
• 18:21
  SPEAKER_00: Now I had trouble finding a definitive source on how that compares to the average hourly wage in India, but the sources I did found suggest it was around$2 an hour on average, with others suggesting that this is heavily skewed and that the medium is a better indicator and is much lower, around 60 cents to a dollar.
• 18:37
  SPEAKER_00: So hard to narrow in an exact number there, but you know, obviously$5 an hour is better than each of those examples.
• 18:44
  SPEAKER_00: Now the caveats, and there are some real ones here.
• 18:47
  SPEAKER_00: The big one is that Korea sells its data to big tech companies and other clients at market rates.
• 18:52
  SPEAKER_00: So this isn't an example of avoiding big AI.
• 18:54
  SPEAKER_00: And it's been reported that royalty payments have so far been relatively limited, and that commercial firms have pushed back because they'd rather buy data outright than track ongoing claims.
• 19:02
  SPEAKER_00: So why did I include this example?
• 19:05
  SPEAKER_00: Well, because it asks a question the whole industry needs to face.
• 19:08
  SPEAKER_00: If these models are going to be built on the knowledge and voices of people around the world, shouldn't those people share in the value?
• 19:14
  SPEAKER_00: It's a difficult question, and one I don't know the answer to, but I thought I'd share the example here regardless, because it's an interesting question.
• 19:21
  SPEAKER_00: So, as a positive, uh looking at the scorecard, it is worker-owned data, which is a powerful model, but the data ultimately feeds other people's models.
• 19:28
  SPEAKER_00: It's a different kind of sovereignty, and it's a question worth sitting with and thinking more about.
• 19:33
  SPEAKER_00: So I'd like to take a step back now and kind of bring all this together and find what these eight examples have in common.
• 19:41
  SPEAKER_00: And I've come up with three kind of main takeaways.
• 19:44
  SPEAKER_00: The first is that for many tasks, small can be enough.
• 19:47
  SPEAKER_00: For most humanitarian tasks, you don't need a model that can write poetry and code and answer every question on earth.
• 19:52
  SPEAKER_00: What you do need is a model that can spot a building, recognize it as easily, for understand a mother's text in Swahili.
• 19:58
  SPEAKER_00: Purpose-built models like Nuru, Incuba LM, and FAIR can require far less compute and can sometimes run offline and maybe much more practical in resource-constrained environments.
• 20:08
  SPEAKER_00: Second, local data can mean better context.
• 20:11
  SPEAKER_00: When communities help shape the data and the model, the resulting systems can better reflect the context where they're actually used.
• 20:17
  SPEAKER_00: And this is important in ensuring kind of equal access and use across the world.
• 20:23
  SPEAKER_00: And third, sovereignty is certainly a spectrum.
• 20:26
  SPEAKER_00: Jaccaranda built on a big tech foundation to keep control of their data.
• 20:29
  SPEAKER_00: Korea works within the system to redistribute its value.
• 20:32
  SPEAKER_00: In all these examples, there's no perfect option, uh, just better and worse choices, made honestly.
• 20:38
  SPEAKER_00: And it reminds me that sovereignty isn't one technical feature.
• 20:41
  SPEAKER_00: It can mean control over data, control over the model, control over where computation happens, or control over who benefits from the value created.
• 20:48
  SPEAKER_00: So that said, here's a simple checklist.
• 20:51
  SPEAKER_00: Five questions any nonprofit can ask before adopting an AI tool.
• 20:55
  SPEAKER_00: The first is, where does our data go and who can use it afterwards?
• 20:58
  SPEAKER_00: The second is, could a smaller purpose-built tool do the job?
• 21:02
  SPEAKER_00: Three, can it run offline or on devices our communities already have?
• 21:06
  SPEAKER_00: Four, are the communities we serve involved in building or shaping it?
• 21:10
  SPEAKER_00: And five, is the benefit worth the footprint?
• 21:13
  SPEAKER_00: If you can answer those five questions clearly, you're putting some structure around AI adoption that many nonprofits still don't have, so great work.
• 21:19
  SPEAKER_00: Good on you.
• 21:21
  SPEAKER_00: And honestly, being ahead of the sector isn't as high a bar as you might think.
• 21:24
  SPEAKER_00: A 2026 benchmark study of 346 nonprofits found that 92% use AI in some form, but only 7% reported major improvements in organizational capability.
• 21:34
  SPEAKER_00: So why might this be the case?
• 21:36
  SPEAKER_00: It could be a number of reasons.
• 21:37
  SPEAKER_00: It could be uh folks not having a clear policy or procedures to guide AI use.
• 21:41
  SPEAKER_00: It could be using models that aren't a good fit for the work that they're doing, it could be the higher-ups uh informing the staff that they need to use more AI tools, even if you know they're maybe not the best fit.
• 21:51
  SPEAKER_00: Uh, there are many different reasons why this could be the case.
• 21:54
  SPEAKER_00: Um, so the challenge for most organizations isn't necessarily access, it's using it deliberately, safely, and in a way that actually is beneficial to the organization as opposed to just being another tool.
• 22:04
  SPEAKER_00: Uh so that's something that I've been doing a lot of work on at Crypto Altras actually, and I want to take a moment to share uh some of the stuff we do that can help nonprofits that are interested in exploring this area a bit more.
• 22:14
  SPEAKER_00: So we do work hands-on with nonprofits to build safe and responsible AI policies and procedures.
• 22:18
  SPEAKER_00: So your team has clear guidance on what to use, how to use it, and how to protect the people you serve.
• 22:23
  SPEAKER_00: We help you identify realistic tools that will help your team, not just add another subscription, another login, or another administrative burden to an already stretched staff.
• 22:32
  SPEAKER_00: And sometimes that means me telling you the right answer isn't AI at all.
• 22:36
  SPEAKER_00: And we support you in implementing those solutions so they're woven into how your team works and help you maximize your impact rather than sitting unused after the first week.
• 22:44
  SPEAKER_00: If any of that sounds useful, head to cryptoaltruist.com slash consulting and you can book a free consultation call.
• 22:49
  SPEAKER_00: And if you'd rather start learning on your own, our free Web3 Impact Toolkit is also available at cryptoaltruist.com slash toolkit, and we are actively expanding it with new resources on AI as well.
• 22:59
  SPEAKER_00: Both of the links for these things are in the show notes, so check that out.
• 23:03
  SPEAKER_00: And that aside, let me leave you with this.
• 23:05
  SPEAKER_00: The concerns about AI are real, and we should keep pushing on every one of them.
• 23:09
  SPEAKER_00: The energy, the water, the safety risks, and the power concentrating in too few hands.
• 23:14
  SPEAKER_00: But the communities we heard about today show that there's another path.
• 23:18
  SPEAKER_00: It could look like a map drawn by people who live there, it could look like a language model governed by the people who speak it, a phone in a tree powered by the sun listening for chainsaws.
• 23:27
  SPEAKER_00: Technology is not destiny, and these organizations prove that AI doesn't have to be something done to communities.
• 23:32
  SPEAKER_00: It can, if done right, be something built by them.
• 23:36
  SPEAKER_00: And so that's what I want to leave you with.
• 23:38
  SPEAKER_00: Thank you so much for listening in.
• 23:39
  SPEAKER_00: I appreciate you, and we'll see you in the next episode.
• 23:43
  SPEAKER_00: And that wraps up today's episode of CryptoUltrist.
• 23:46
  SPEAKER_00: Thank you so much for tuning in.
• 23:47
  SPEAKER_00: I hope you enjoyed the conversation as much as I did.
• 23:49
  SPEAKER_00: If you're inspired to explore more stories at the intersection of Web3 and Social Impact, be sure to visit us at CryptoUltruist.com for articles, resources, and more.
• 23:56
  SPEAKER_00: And if you love what you heard, it would mean the world if you could rate review and subscribe to the podcast.
• 24:00
  SPEAKER_00: Your support is huge in helping us reach more listeners like you and keeping the platform going strong.
• 24:04
  SPEAKER_00: Thanks again for joining us, and I look forward to having you back for the next episode.
• 24:07
  SPEAKER_00: Until then, let's keep showing the world the good of crypto.