AI Regulations: Trickling up, Pouring Down, or Nowhere to Be Seen?

Summary

Humanitarian Frontiers Podcast: "AI Regulations: Trickling up, Pouring Down, or Nowhere to Be Seen?"

Podcast: Humanitarian Frontiers
Episode: AI Regulations: Trickling up, Pouring Down, or Nowhere to Be Seen?
Date: April 14, 2025
Host: Chris Hoffman, with co-host Nassim Motalebi
Guests: Gabrielle Tran (Institute for Security and Technology) & Richard Maura (Tech Worker Community Africa)

---

Episode Overview

This episode takes a critical look at emerging AI regulations—where they exist, where they don’t, and what this means for the humanitarian sector globally. Host Chris Hoffman, co-host Nassim Motalebi, and expert guests Gabrielle Tran and Richard Maura discuss the realities behind model training, hidden labor, diverging regulatory approaches, ethical grey areas, and how all this impacts humanitarian organizations’ responsibilities and risks when deploying AI.

The conversation moves between technical foundations, global policy gaps, discussions of power and consent, supply chain ethics, and concrete suggestions for NGO practitioners navigating this ever-shifting landscape.

---

Key Discussion Points & Insights

1. What Actually Happens Behind the Scenes of LLMs?

(02:44 - 05:54)

• Richard Maura demystifies model training by explaining the critical but often unseen role of human labor in AI moderation:
  • “AI does not exist without data. This data cannot work on itself... The need to employ individuals who can sit behind the machine and train this data…cannot be underestimated.” (02:44 - 04:53, Richard)
  • Moderators remove harmful or sensitive content; this is why, for instance, a chatbot like ChatGPT "refuses" certain requests—humans have deliberately trained those guardrails.
  • The work of training and moderating LLMs is typically outsourced to low-wage workers in the Global South, raising fundamental ethical questions.

Memorable Quote:
"Moderators are the gatekeeper of the text, so to say. Absolutely." (04:29, Chris & Richard)

---

2. Ethics or Just a Performance? The Labor Supply Chain & Unacknowledged Workforce

(05:54 - 09:46)

• Gabrielle Tran argues that current AI ethics discourse largely serves providers, not society:
  • Focuses on technical dimensions (privacy, fairness, explainability), which benefits companies seeking to sell a responsible image.
  • Ignores deeper social and political consequences—including the exploitative labor behind content moderation and data labeling.
  • Cites investigative work showing AI supply chains often depend on underpaid, poorly protected labor in developing economies, exposing humans to toxic material for model filtering.
  • “If the foundation of even ethical AI depends on unethical labor, can you really call that ethical?” (07:24, Gabrielle)
• Richard adds:
  • Much training data isn't owned or sourced ethically, compounding risks of intellectual property abuse and consent violations.

---

3. Policy Landscape: Europe Leads, Others Patchwork, Global South Lags

(09:46 - 13:06, 16:01 - 20:35, 21:34 - 25:50)

• EU AI Act: The first comprehensive regulation, enforces staff training and lays out risk-based rules and documentation mandates.

• US: No overarching federal law. Mix of voluntary frameworks (“AI Bill of Rights,” NIST RMF), plus patchwork state laws (e.g., biometric privacy in Illinois, hiring rules in New York).
  “The US...emphasizes not stifling innovation... The EU is more willing to impose hard requirements.” (12:23, Gabrielle)

• Global South/Africa:

  • Richard explains that, for many African policymakers, AI isn’t a priority amid economic and political challenges.
  • Gaps in ethical oversight, lack of investment in regulatory frameworks, and a profit-over-impact mindset hinder meaningful governance.
  • “AI in Africa is not a priority. Unfortunately, there are underlying issues... Politicians [are] not about what AI is and what impact it has in society.” (22:13, Richard)

• Practical Implications:

  • NGOs straddling multiple jurisdictions are left with ambiguity.
  • Gabrielle notes potential for global “race to the top,” as major firms may standardize systems to the strictest rules (similar to GDPR’s worldwide influence).
  • However, when compliance costs outweigh potential revenue (especially in smaller markets), companies might simply pull services—“Meta pulled the news from Canada in response to Canada’s [regulation]...they pulled the plug.” (23:53, Gabrielle)

---

4. Who Is Accountable When AI Crosses Borders?

(16:01 - 20:35, 18:46 - 21:33)

• The EU AI Act requires “appointed representatives” for general-purpose models.
  • Humanitarian organizations can (soon) request full documentation and “paper trails” on how models were trained—“Can I have the paper trail of how the model was built?” (18:46, Gabrielle)
  • This could improve transparency, but accountability for downstream humanitarian deployments outside the EU (e.g. EU-based NGO using AI in Kenya) remains murky.
  • Chris: “What happens then? Who’s the responsible party for what’s happening in Kenya if you’re a European NGO?” (20:01, Chris)

---

5. Responsible AI in High-Risk Humanitarian Contexts

(13:06 - 16:01, 26:06 - 29:27, 36:05 - 40:29)

• Nassim raises the stakes:
  • Humanitarian work often avoids AI deployments that haven’t been stress-tested, due to concerns for at-risk populations.
  • “We don’t see that many AI use cases that are actually tailored for affected populations because we don’t want to risk them. We don’t want to risk deploying an AI solution that could have consequences.” (14:54, Nassim)
• Gabrielle explains regulatory mechanisms for accountability:
  • Providers must equip downstream users with documentation and risk guidance.
  • Under the EU Act, there will be channels for NGOs to request audits/insights into models they're using.
• Global divergence vs. convergence:
  • While regulations diverge, market pressures may push companies to adopt strictest standards globally (GDPR precedent).

---

6. Consent, Data Sovereignty, and Power Imbalances

(29:27 - 33:22)

• Data trusts as an (imperfect) solution: a third party holds and manages data for beneficiaries, helping navigate permissions and privacy for vulnerable groups.
• However, Chris and Richard highlight huge deficits in digital maturity and meaningful consent:
  • Workers in the Global South often accept poor conditions without understanding data rights or consequences, driven by economic desperation.
  • “Consent is the last thing individuals, especially the workforce, will be mindful of... They are desperate for work.” (31:15, Richard)

---

7. What Should Humanitarian Organizations Do?

(36:05 - 40:29)

Gabrielle’s Policy Roadmap:

• Start with core ethical principles: fairness, transparency, accountability, human well-being.
• Align policies with the highest available standards (EU, NIST).
• Assign explicit AI governance responsibilities internally—no matter organization size.
• Maintain human accountability for every major AI-driven decision.
• If in doubt, delay deployment and beta-test in controlled environments.
• Invest in AI literacy for practitioners, especially in high-stakes contexts.
• Use simple models where possible, as high-complexity models increase opacity and risk undetected biases.
  • Example: Healthcare AI that misunderstood asthma data due to flawed training assumptions, which could have endangered patients (38:38, Gabrielle).

---

Notable Quotes & Memorable Moments

• “If the foundation of even ethical AI depends on unethical labor, can you really call that ethical?”
  — Gabrielle Tran (07:24)

• “Some of this data... is not even the owners’... That also raises a whole issue about ethical considerations in AI.”
  — Richard Maura (08:36)

• “Accountability in humanitarian AI use means that humans remain answerable for any decision... If a human in your organization feels they can’t answer for an AI-driven decision, I don’t think you should use it.”
  — Gabrielle Tran (36:52)

• “Consent is the last thing individuals... will be mindful of... They are desperate for work.”
  — Richard Maura (31:15)

• “These corporations will comply with stringent regulations only when the cost of non-compliance outweighs the cost of adaptation… In smaller markets, perhaps in the Global South, where compliance could outweigh potential revenue, they might just opt out entirely.”
  — Gabrielle Tran (23:53)

• “AI, or more complex AI, is not an inevitability if it does not enhance these capabilities and real kind of ethical deliberation. It demands asking not just how to ethically implement AI, but whether certain applications of AI should exist at all. It's inherently social, it's political, and it's a moral judgment rather than a technical one.”
  — Gabrielle Tran (42:42)

---

Timestamps for Significant Segments

• 02:44 — Richard Maura explains the basics of LLM moderation and training labor.
• 05:54 — Gabrielle Tran discusses how ethics conversations serve providers and expose labor exploitation.
• 09:46 — Chris & Gabrielle on the fragmented global regulatory landscape.
• 16:01 — Accountability and auditability in new EU regulatory frameworks.
• 21:59 — Richard outlines Kenya’s and Africa’s low prioritization of AI policy.
• 23:53 — Gabrielle draws lessons from Meta's withdrawal from Canada.
• 29:27 — Chris and Gabrielle discuss data trusts, digital maturity, and the challenge of consent in humanitarian contexts.
• 36:05 — Gabrielle’s roadmap for humanitarian AI governance and the asthma patient AI case.
• 42:23 — Closing: Gabrielle on the need to ask “whether” some AI use cases should exist.

---

Closing Messages & Hopes for the Future

• Gabrielle:
  Emphasizes pursuing fairness, transparency, accountability, and human well-being over technical solutionism, arguing that the debate on AI's ethical use must be grounded in real-world consequences rather than mere performance of ethicality.

• Richard:
  Calls for stronger UN and international intervention to ensure member states uphold human and labor rights in AI supply chains, especially in contexts vulnerable to exploitation.

• Chris & Nassim:
  Highlight ongoing need for investment in AI literacy, deliberative organizational governance, and a culture of “principles first,” particularly when the stakes are highest for humanitarian populations.

---

In Summary

This episode of Humanitarian Frontiers lays bare the real-world, high-stakes drama of AI deployment: a vast gulf between shiny tech promises, global regulatory gaps, invisible labor, and the practical struggles faced by humanitarian organizations. The conversation is a must-listen (and a must-read) for practitioners wanting to look beneath the surface of AI tools and make responsible, human-centered choices for the most vulnerable in society.

Transcript

[A]

Welcome to Humanitarian Frontiers in AI, the podcast series where innovation meets impact. In each episode, we dive deep into how artificial intelligence is reshaping the future of humanitarian work. From enhancing crisis response to making a delivery smarter and more effective, AI is opening new doors in the way we support communities in need. In this series, hosts Chris Hoffman and Nassim Motelaby bring you thought leaders from academia and the tech industry discuss not only the vast opportunities AI offers, but also the ethical considerations and risks we all must navigate. Join them on this journey as they explore AI's potential to transform lives and address humanity's most pressing challenges.

[B]

Hello everyone and welcome back. This is our next episode of Humanitarian Frontiers in AI. I'm your co host Chris Hoffman together with Nasim Motalebi. Hi Naseem, how are you?

[C]

Hi Chris, I'm back.

[B]

It's great to have you here today and we've got such a amazing and dare I say, unique panel because we're really going to be diving into some of the greater, higher level policy discussions that are going on, but also the impact of those policies and how it's directly impacting the people that are working around the tech. And so I think it's going to be a really rich conversation. I'm super excited to have our guests here today. We've got Gabrielle Tran, who's from the Institute for Security and Technology as a senior analyst. And also we're here with Richard Maura, who's the co founder of Tech Worker Community Africa, as well as times one of 100 most influential people in AI of 2023 and a number of other things. So Richard's quite a busy guy. So it's great to have you both here and to spend that time with us today. So welcome Richard and Gabrielle.

[D]

Thanks Chris.

[E]

Thanks for having us.

[B]

Yeah, it's going to be great. I want to start maybe Richard with you just because this is a piece of the puzzle that doesn't necessarily get talk too much. Right. We talk about AI, we talk about training models, we talk about moderators, but I don't think anyone quite understands what that means. And so I think for the humanitarian community specifically, being able to kind of understand better where it comes from, why does it speak in the way that it does when we talk to a conversational LLM? Why does it have right answers and wrong answers and then later those answers are corrected if they were wrong, how does that work? And, and really being someone that comes from that space, just, yeah, take a moment and kind of explain to us what moderation is and how moderation is being done for a lot of the LLMs.

[D]

Thanks, Chris. So I'll just go straight on to the answer. AI does not exist without data. This data cannot work on itself. And so the need to recruit individuals who can work on this data, who can train this data to be understandable algorithms that. And I'll try as much as possible to explain this as if I'm explaining it to a nursery or a kindergarten school.

[B]

That's great.

[D]

And so the need to employ an individual who can sit behind the machine and train this data to be algorithms that is understandable to the general public, it comes in very handy. What am I saying? ChatGPT, for instance, and by the way, one of the people who trained ChatGPT to be what it is today. So ChatGPT for it to be ChatGPT right now or for it to be understandable in whatever manner as it is right now, we sat behind the machine to sift off data which is data which is understandable or data which is unfriendly to the human, to the human public or to the human eye. And so, for instance, if you go to ChatGPT and ask the platform, how do I get to a place where I can destroy so and so or kill so and so, for lack of a better word, ChatGPT will not respond or will not give you a response in as far as that query is concerned. Why? Because that information has already been sieved off. It has already been put off that platform by individuals like us. Right.

[B]

Moderators are the gatekeeper.

[D]

Right. Of the text, so to say. Absolutely. Absolutely. And so the need to employ individuals who can train this data comes in very handy. And so the importance of an individual or in the importance of human labor in as far as shaping this data or in as far as training, this data cannot be underestimated. Thank you.

[B]

Absolutely. Sure, Gabriela. Obviously that then goes directly to the policy thing. But Nassim, go ahead. I'd love to hear what you think.

[C]

Thanks, Chris. And thanks, Richard. You speak around training and model training, and we don't really get into the nitty gritty technical topics in this podcast, but maybe, Gabrielle, maybe you could reflect a little bit for our audience. What do we mean when we talk about training a model, especially with LLMs, how can we distinguish training a model, what Richard was speaking to, versus when we're fine tuning or versus when we are using, let's say, a rag architecture, which is really popular these days. I think we talk a lot about the risks of training models and how they can access our Data, what is the realities that we're facing and how does that. In fact, I guess the next question would be our governance and strategies around AI moderation, I guess.

[E]

Yeah, sure. So I guess just to start off what Richard was getting at, like LLMs are just kind of pattern recognition tools, right? It's predicting the next word. It doesn't really understand what's going on. And I think to kind of Richard and to your point, AI ethics typically follows the concerns of AI practitioners rather than questioning AI's role in society. So ethical issues are framed as technical challenges that can be solved through a better design, like you know, better privacy, like you're saying, rather than social or political problems. And that leads to a focus on explainability, fairness, privacy as important as they are. And we can get into that how kind of organizations can operationalize privacy in this huge AI world with all these regulations and stuff. But I think we need to account for, as Richard is saying, these things that we focus on in AI ethics rather than the ethics of AI in part benefits the provider, right? Because that's what people want to see when they buy these services. They want to see their privacy as a priority, they want to see the company is accountable and so on. But the ethics that don't benefit them is typically what the general consumer is unaware of. For example, like Richard is saying, like you said, he took part in many general purpose models. Content moderation is outsourced to countries that labs they might know they can kind of exploit. So we know they go to developing economies and task workers with reviewing and labeling kind of content which could result in catastrophic risk. Or it's like descriptions of violence and other toxic materials to detect, to filter harmful content in their model inputs and their outputs. So that's the kind of clickwork that goes into them. And often this comes. There was a times recent investigation that lots of times these people are underpaid. And this kind of raises the bigger question, which AI ethics is not just about outputs, it's about people behind the machines. And if the foundation of even ethical AI depends on unethical labor, can you really call that ethical? And if AI governance doesn't address that, is it missing the point? So I think to what we're all saying here, it's one thing to see the kind of legislation we're seeing which is absolutely necessary to set the common standards for risk reduction and encourage companies to be thoughtful and accountable. But something I would love to see more from humanitarian organizations and governments, frankly, is a much more broader look at the AI expansive supply Chain.

[B]

Absolutely. Richard, over to you.

[D]

Yeah, and I totally agree with what Gabriel, I totally resonate with Gabriel's sentiments in as far as ethics is concerned. But I also need to raise a fundamental issue in as far as ethical considerations of the whole LLM platforms and LLM models are concerned. Some of this data, Chris, and some of this data that individuals or these organizations outsource, believe it or not, it's not even, they are not even the owners of this data. And so that also raises a whole issue about the whole aspect about ethical considerations in AI because it does not generate from their organizations. You find an organization getting data from website X, getting data from website Y, then compiling them as one data and then pretending that this data comes from their own organization, then they outsource this work to far and wide organizations or far and wide individuals. And that breaks the whole notion about ethical consideration. And so that's what I wanted to point out in as far as AI ethics and governance is concerned. Thank you for sure.

[B]

Gabrielle, you mentioned supply chain. And I don't think people really, they're only seeing the delivery, right? They don't see everything that goes on behind the scenes. And you pointed that out quite well. And we get into the policy discussions, right? We know that the EU is put out an AI Act. It's actually come into, well, phase one has come into force, I think last month, which is around staff having to be trained. So if you are using AI or selling AI, your staff have to be trained on what AI is, et cetera, as kind of the baseline. But then the US is not doing anything around this per se. California tried to push forward with some legislation that was unfortunately shut down, but the US as a whole really doesn't have anything. I'm not sure that Canada does either yet. So really, what are governments, what are governments outside of Europe starting to do some where where the LLMs are or where the companies that are perpetuating the LLMs are sitting, what are they doing?

[E]

Right. So yeah, as you pointed out, the EU has kind of a single comprehensive AI act that applies across industries, sets unified rules in the United States. There isn't a single overarching AI law at the federal level. So what is in the US is a mix of federal agency guidelines, executive actions, state law, all being addressed, piecing together the AI puzzle. So right now in the US federally, there are kind of soft law efforts like the AI Bill of Rights blueprint released by the White House, the NIST AI Risk Management Framework, which provides voluntary guidance on principles like fairness and transparency. We had under the Biden administration the executive order on safe, secure, trustworthy AI, but this was recently rescinded and replaced with removing barriers to American leadership in AI. So these are significant, but again, they're not all encompassing law. They're more like directives, recommendations for best practice. And like you mentioned, Chris, at the state level you have certain laws targeting specific AI issues. For instance, Illinois has biometric information privacy acts and New York rules out rules on AI and hiring. But I guess the resulting divergence here is that the EU focuses on ex ante regulation. I would say, you know, setting rules before harm kind of occurs because their preventative approach is based on risk levels. Whereas the US leads to more ex post specific sectors, specific states and addressing harms through existing laws like anti discrimination consumer protection by the ftc. So culturally I would say, you know, the US kind of emphasizes, you know, not stifling innovation. Regulators say they don't want to overregulate such a competitive technology. And the eu, while they would say they also encourage innovation by pointing to certain exemptions on open source models or open science models, they're more willing to impose kind of those hard requirements. But yeah, navigating multiple AI regulatory regimes is a very complex task, especially for humanitarian groups. Yeah.

[B]

You know, in the US we have an organization called Ohio, we have it called osha, which really just manages the standards for workers to make sure that workers don't hurt themselves. And so having those types of things are super important. But I want to give Nassim, I know that she's got a lot of questions and I always am excited to hear them go for it. Nassim, thanks.

[C]

I really liked what Richard was talking about regarding law and the implementation of the law. Right. And I think that's the main challenge that we're facing because sometimes the regulations and governance strategies don't have an implementation plan. Gabrielle mentioned NIST framework, the rmf and I think that's a, if I'm not wrong, the acronym responsible AI framework, raif. So with that framework I think you have a good checklist of what do you need to implement in the organization in order to reach a level of AI maturity. But also responsible AI practices. But it also takes a lot of investment, it's costly to do this. So it's also about a trade off. Right. How much can businesses spend in terms of AI governance and managing their AI solutions and also training their staff? Should there be responsible for doing that or should we actually hold AI companies responsible for this? Right. So I think that's a Very interesting topic to explore, the responsibilities of AI users versus AI providers. But I also wanted to kind of maybe speak to ethical AI and AI governance in terms of what it means. I hear from Gabrielle and Richard that there is the privacy and security and accountability, but also there is this social aspect, the implications that AI has in our society and sometimes it's not so measurable and we can't really see it directly. So we often go towards accountability and protection of our data privacy and our own workforce. For example, who is using AI, but we don't see its impact in the humanitarian sector. I think that's a very, very important question. And Chris, we've had this conversation before where we don't see that many AI use cases that are actually tailored for affected populations because we don't want to risk them. Right. And then we don't want to risk deploying an AI solution that could have consequences. So my question is generally how do we deploy AI responsibly in a very high risk circumstance in the humanitarian sector?

[D]

Right.

[C]

And should we, or should it consider as one of those flagged high risk use cases or should it be sector specific? So, yeah, what do we think about these social implications versus this privacy security conversation?

[B]

Gabriela, we were together in San Francisco. I don't know if you want to speak a little bit about, you know, that conversation and the conversations that you're going much deeper in than all of us around that social aspect.

[E]

Sure. I think to the first point, maybe we can actually start with the EU AI act and the general purpose AI when we talk about like accountability and who's accountable. So looking ahead in the EUA act, the next big milestone is August 2nd. So that's for foundation models and general purpose AI. So those are AI systems that are capable a range of general tasks like text synthesis, image manipulation and stuff. So notable examples are OpenAI's like GPT3 and GPT4, which are the foundation models that underpin ChatGPT. And because foundation models can be built on top of to develop many different applications for like, Chris, you were saying, like Chatbot companions or even like if you have a finance firm and they have an internal GPT email writer, you know, that is the foundation model, but that makes them very difficult and important to regulate. So this part of the EUA act, the general purpose AI requires information to downstream users, which I think is getting to your point. NASSIM providers need to provide necessary information and documentation to downstream providers who intend to integrate their general purpose AI model to their own systems. And by mandating technical documentation and information sharing. And the act ensures that those who build AI platforms provide enough knowledge to others to use them responsibly. And this information should enable these users to understand the model's capabilities, the model's limitations. And in practice, that could include things like a model's intended use, accuracy metrics, known risks for potential misuse, and guidance on integrating the model responsibly. And I think for humanitarian organizations that are working in the EU or across borders, or they're using different tools and stuff, they should know that providers that are established, maybe in the US or a non EU country, before they place a general purpose AI model on the EU market, they need an appointed representative in the eu. So in terms of accountability, if you're a nonprofit operating or using an EU tool, you do have the right, not quite yet, but soon, to go to this EU representative mandate and request. Can I have the information for downstream users? Can I have the tech, the goal documentation for the model, which should cover its training data. You know, how much was it from a copyrighted source, how much was it from images from a certain data set? You know, how can I improve the transparency and what the knowledge is built on? Essentially, can I have the paper trail of how the model was built?

[C]

That's interesting, but. So if I understood correctly, you're asking for humanitarian organizations to have representatives to be able to kind of ask or be held accountable in the EU AI Advisory Group.

[D]

Right?

[C]

Is that what you're suggesting?

[E]

No, no, sorry. This was for general purpose AI. They should have this representative. So I'm saying humanitarian orgs can go to. They should be able to go to this representative and ask for the kind of paper trail of how the model was built. Right. So this is holding the providers accountable and it's giving organizations that kind of pathway to get the appropriate resources they need to kind of maintain their internal practices.

[C]

Right. Like a centralized system where you can actually get the information you require for auditing.

[D]

Right.

[C]

And doing that. Yeah, that's an interesting system.

[B]

Yeah, I think that's super important for our listeners. Right. I mean, I think that that point. Because right now, what's happening with all of the different people we've interviewed throughout the podcast and people that we've talked to separately, that level of understanding of what you can do or what you should be able to do in the future, but also what you need to keep in the back of your head, because we're talking about accountability of the providers right now. But then there's also the second accountability, which is those that productize them and then use them for services to others. And so then how are these? I guess the big question is how for me is what will be the responsibility of humanitarian organizations that are based in in Europe but implement these tools, for example, in the Global south, let's just say in Kenya, since Richard is here. And so ngox, based in Ireland or based in the Netherlands, implements a chatbot using generative AI in Kenya. What happens then? Who's the responsible party for what's happening in Kenya if you're a European ngo? And I think that's a big question. I don't know if you have the answer to that, Gabriel or Richard, but I think that's an important point that we'll have to suss out over the course of this podcast.

[E]

Yeah, absolutely. I think some AI tools will be developed with only the US market in mind, meaning they don't meet EU requirements. So a humanitarian org might find a great AI tool for say analyzing satellite images for disaster response. But if it doesn't comply with EU standards or even using it in an EU funded project, that could be problematic. But I think the flip side is also true. If AI developers, they might decide to comply globally with the toughest standards to just simplify their operations. So if the EU A act becomes that kind of de facto standard, we might actually see some convergence in practice with companies just globally raising the bar on their AI systems so they can operate in the EU and beyond. So divergence at a policy level, maybe state to state, but possible convergence in practice due to market pressure, frankly. And that's how the gdpr, which is the eu. Yeah, Data Privacy Law, it ended up being influencing data practices worldwide. And I think we might be even seeing the same thing with the EU AI Act. Yeah.

[D]

Wow.

[B]

Yeah. And Richard, what's going on in Kenya right now? I know that I just saw that the government had released conversations around developing a Kenya based LLM. Right. So they're trying to look to see how to do a Swahili specific one. I don't know if they're going to have enough data to fill in there. But what else is is happening and what are you hearing from your colleagues across the continent around what's happening with AI at this policy level discussion?

[D]

The thing about coming up with platforms, and this is the unfortunate part about Africans, most of us, we are as selfish as it's not for the greater good. If I want to create a platform that is meant to benefit people or myself, I will think about profits as opposed to the impact of these platforms on society. But be it as it may, the policy framework in as far as AI is concerned, we have different, a divergent of views ranging from one country to another, ranging from one person to another. And this is simply because AI in Africa is not a priority. Unfortunately, there are underlying issues that and so policymakers might not be considered in as far as when it comes to AI, we have a whole lot of things to think about. We have, for instance, in Kenya we have politicians running around thinking about the next elections and not about what AI is and what impact it has in society. And so unfortunately, to respond to that question, it is not a priority per se in Kenya or in Africa. And that's what the leaders have placed in as far as priority or nitty gritty in their line of action is concerned.

[B]

Right. So there's not much happening. Gabriela?

[E]

Yeah, I think, yeah, to Richard's point, that's exactly the case. I think, as I said before, I think AI providers are likely to find the highest common denominator, like the EU AI act, to adapt their operations to streamline rollouts, or at least where not doing so would result in a loss of service that would cost them greatly. So what I mean by that is we see the case in Canada. Meta pulled the news from Canada in response to Canada's Online News act, which required platforms to compensate the news publishers for content arguing Meta argued this is unsustainable for a business model. So instead of bending the need to Canadian policymakers, lawmakers, they said, you know, it's not profitable for us to do that, they pulled the plug. So that just goes to show that these corporations will comply with stringent regulations only when the cost of non compliance outweighs the cost of adaption. So in case where compliance is too expensive or it undermines profitability, they might just withdraw service entirely, you know, as seen in Meta in Canada. And that suggests for AI governance, global tech firms are likely to standardize operations around the state strictest, most economically significant regulations because it ensures this kind of, you know, market access while avoiding region specific headaches. But in smaller markets, you know, maybe perhaps in the global south, where compliance could outweigh potential revenue, they might just opt out entirely, you know, rather than adjust their deployment strategies.

[C]

You know, I really like what Gabrielle, you just mentioned about the case of Canada. And I think what Richard is I like that we have the perspective of Richard speaking to a specific context and then going and having a global perspective, but thinking about this globe, like the presence of these companies on a global scale. Right. The Case of Canada is a good example where the government is actually proactive in AI regulations, right? AI for humanity, AI for good, and in the benefit for its populations. But when it comes to the global politics, we can't say that for every government, right. There are governments that actually use AI for surveillance, for control, and actually actively work against their own people, if we may call it that. So my question is, can these regulations actually be deployed then globally, and what does that mean for the humanitarians?

[D]

Right.

[C]

I mean, what Richard is mentioning around governments activities or intentions around AI governance, that's one thing. But being complacent versus being actively trying against AI for humanity is a different topic.

[D]

Right.

[C]

And I think that creates a counterbalance for what you're describing, Gabrielle. So how do you think we can create policies that could kind of level this playing field in some ways? Or can we actually, especially as humanitarians?

[E]

That's, I mean, my first instinct to just completely support what you're saying. Right, so the EU AI act, the prohibited AI practices that just came into effect on February 2nd. Right. And one of the practices that are banned are real time biometric identification in public spaces. So like you're saying kind of mass surveillance using, using facial recognition. And that is a firm ban. But of course that has some exceptions that are very important for the humanitarian space. So this act actually makes exceptions for law enforcement related activities, including migration authorities, to use like real time facial recognitions in public spaces or an aid in a suspect's participation in a crime or missing people or like serious cases like threat of terrorism. Terrorism, as long as they have the judge's authorization. So what you're getting, yeah, that's worrying on two fronts. One, the technical faults within the AI system itself, you know, false positives, the racial and ethnic bias. Racial facial recognition systems have been shown to disproportionately misidentify people of color. But also what you're, what you're getting to is kind of of the mission creep. You know, initially we're going to say it's used for emergencies and these technologies could expand to monitor certain types of people in everyday life. And I think the big risk, what you're getting at is that NGOs who maybe work with the government on say, migration issues, they may feel pressured to share this data, right, to put their beneficiaries at risk because of this exemption. So it's very important that we kind of stay on top of that and we stay on top of the privacy laws. And I guess to add to that, I think an interesting A possible kind of solution is the idea of a data trust, right? So a data trust is a legal framework that gives an independent entity the responsibility to hold and manage data on behalf of a group of beneficiaries. So the independent entity, you can kind of think of them as like a board of directors, right? It's meant to act in the best interest of its users so they can define access, use liability to align with agreed upon guidelines and privacy standards, value driven goals. So it's kind of a solution to create fiduciary accountability in situations where giant pools of data are of interest to multiple stakeholders, including the government. So you can consider a nonprofit entity or a third party coalition established to serve as a trustee for user generated data for say, a refugee service. And that trustee body could be composed of experts in fields like AI, ethics, privacy, law, migration, human rights, geopolitics, tech. And trustees would only grant the access to this data under controlled conditions that align with the best interests of their users and maybe the humanitarian mission of their organization. So the idea being that this interdisciplinary expertise ensures the ability to address complex trade offs and kind of prioritize outcomes that serve the collective interest of the user. So not a total solution, but it is a governance structure that could be in place to mitigate some of the things you're saying.

[B]

You're talking my language, Gabrielle. This idea of this trusteeship model in Blockchain, they call it the guardianship model versus the sovereign model. I think the one thing though that comes up to me which tends to be obviously because of the gdpr, is the idea of consent. And even if you have a trust, how do people, especially in areas where they might not be as digitally mature or understand their rights, so to say? So as you start talking about places where you have a population, as I said, that does not have a level of digital maturity to understand their rights in the digital platform. How do we do it in that way? Because one of the models around Blockchain that was happening was that an NGO would be able to register beneficiaries, but because beneficiaries didn't quite understand how to erase their data or how to manipulate their data through on the chain, which they have the right to do, then the NGO takes on the guardianship of that data. So kind of a similar thing to what you're saying, and it would manage the data on their behalf. But the issue that they kept running into was consent. And I mean, Richard, I think that that then translates down into the workforce as well. So now you've got a workforce that might also not understand what they're getting into and not being able to give informed consent, so to say, on doing the actions that they're doing. And so is it. I guess I'm mixing a lot of things here, but. But in general, the. The point I'm trying to say is consent, right? So a person that's part of the not monitoring, but the use or building of the tools, the monitoring of the tools, and then the person that's using the tool going to get to a point where everybody is going to be able to give consent. To your point, Gabrielle, on these collective trusts, for example, how do we fix this? Because it seems to me to be the large linchpin that nobody's able to.

[D]

Really break the challenge, especially with respect to Africa. Consent is the last thing individuals, especially the workforce, will be mindful of. And this is the reason. And it's a sorry state of affairs because the owners of these platforms have made the impression or have this perception that the workforce in Africa is very desperate for work. You call them to do something today, if they don't have something that will make ends meet, they will greatly commit themselves to that which they are being called for, if at all it is putting sense or it is to provide bread and butter. And so consent becomes the last thing, or even it is not the real deal at the end of the day for them simply because they are desperate for work. And the providers of these platforms think that these individuals are vulnerable. I have made them to be vulnerable. And that's the point. And so at the end of the day, even when you get to a point where you have this situation, it gives your employer the right to dictate how things should work on your end, at what time you should go home, at what time you should stop working on these platforms. You can work overtime without pay. You can do everything. So long. And remember, this is data. And we are dealing with outsourced work. And so that means we are dealing with three parties. The workforce, the outsourcing organization, and the organizations that create these platforms. And so whatever the workforce is asked by the outsourcing organization to do, they will render themselves to doing it without considering the nitty gritties. What are you consenting yourself to? What are you binding yourself to? Is it worth the fight? Is it worth the cake? And so that's a sorry state of affairs that we are dealing with, unfortunately.

[E]

Absolutely.

[B]

No, I really appreciate it. I mean, Naseem, over to you.

[C]

Thanks, Chris. This is generally a very interesting conversation. Speaking about our individual rights, population rights, kind Of I would say decentralized governance plus the policy top down governance models. And I really liked this thinking around, okay, third party trust holders I guess we could call them, where we could rely on for holding our digital rights and advocating on people's behalfs and organizations behalfs. So I guess one of my final thoughts is what is the role of humanitarian organizations like the UN or just generally when we speak about human rights, right, and what does it mean to have digital rights? Can we actually see it enacting where we can all benefit from it and how can we decentralize that? Should we actually think of a new model for digital rights that includes AI development use? And also. So it's kind of essentially misuse I would say when we see AI is actively being manipulated, maybe through red teaming we can stop that. Maybe it's certain hackathons, I don't know what are the solutions at the end of the day that we could foresee? Maybe Richard, maybe we can start from you from an African perspective and then Gabriela maybe from a policy perspective.

[D]

I think there is with respect to the United nations, there is need because United nations has the membership of this entity in which our country is also a member of this entity and has bind Kenya has bound herself to so many treaties in as far as workforce and labor is concerned and even human rights are concerned. I think there is a greater need to commit these countries to respecting fundamental human rights, which is the foundation of what we are talking about. It's the most important thing in as far as AI is concerned. There is a greater need for United nations and even big brothers like the United States and uk, Canada, Germany and so on and so forth. There is need to ask or to force countries. I don't know whether force is the right word to use, but there is need to ask these nations to commit themselves or to respect fundamental human rights, which is the most important thing. Nassim thanks Richard.

[E]

Gabriella yeah, I think first of all, developing an internal AI policy starts with clarity on principles and scope. So you humanitarian organizations can kind of start by asking, you know, how do we want to use AI to support our mission and what boundaries do we need to ensure we do no harm? And kind of my advice there is to align internal policies with the high, the highest standards out there. So starting from ethical principles and risk assessment, whether you're looking to the eu, AI act or nist, humanitarian organizations can look to these kind of mandates to systematically think about things like bias, robustness and transparency in their use of AI systems. And you need to tailor that to governance to scale, whether your own humanitarian org, like a big UN agency or a small local ngo. If you're smaller, you might have one person kind of double hatting as the AI ethics point of contact where a bigger agency can have a full committee. But the key is to just have somebody accountable for AI governance in the organization. That way as laws evolve, there is at least a person kind of keeping, keeping track of them. And I think no matter what size your organization is, the risk assessment threshold remains tethered to accountability. So accountability in humanitarian AI use means that humans remain answerable for any decision, even an AI driven decision. Right. And there's a reason that high risk AI in the EU AI act stipulates that AI should not fully operate autonomously in critical applications. So I think once a human in your organization feels they can't answer for an AI driven decision, I don't think you should use it even just from like a strict liability standpoint. Because let's say an algorithm might flag this family gets aid first, but a human manager should be responsible for that decision and should be ready to either accept, explain why they went with it, or be prepared to override it. And in practice that might look like a humanitarian group explicitly assigning responsibility to a person for each phase of the AI system life cycle, so to say. So who's accountable for looking at the quality of the training data? Who is our EU liaison? Who is accountable for validating the model recommendations against past decisions that we've made? So overall, I think if there is any doubt that the AI tool is kind of causing such harm, you don't need to use it. You can delay until those risks are addressed because a precautionary approach is key. And maybe a good example is beta testing an AI tool. You could run it on past data, you could run it in a controlled hypothetical pilot to actually verify it's addressing decisions and it doesn't produce dangerous errors. There's no reason to deploy it right away, especially if you're a smaller organization. And I think to my last point, I'll wrap it up soon. But we need to invest, humanitarian groups needs to invest in AI literacy so that the people overseeing these tools, they understand them enough just to intervene appropriately. And to your point, Chris, explain to the users where the data is going. So I'm a fan, especially in these higher stakes situations of simple AI models. So high complex, highly complex AI models can obscure the variables driving critical decisions. Right? It's difficult to detect biases or errors in Brian Christian's book the Alignment Problem. He shares this really fascinating story from healthcare where researchers were building an AI system to decide whether pneumonia patients need hospital admission or who can go home, who can recover at home. And the AI suggested that patients with asthma should be sent home, but that makes no sense medically. So it turns out the AI was learning from historical data because asthmatic patients go to the hospital right away, right? They get really careful treatment because it's a respiratory disease, so they end up recovering very well. So the AI understood this, misunderstood this, I should say, and concluded asthma patients are at lower risk because they recover really well. But it's confusing those variables, right? So if we have that AI literacy and we have a simple models that practitioners can clearly evaluate and correct the criteria influencing them, it's a faster way to become literate in these systems and it's a better way to explain to folks where their data is going because we can actually see it in real time.

[B]

Believable. Wow. There was a lot there and that is amazing. No, really, I mean it's just kind of like mind blown moment there when you really think about it, about the application of that. And I think that when you talk about the investment, I really do want to thank Innovation Norway for funding this podcast because the reality is that there is a lack of knowledge. There is a knowledge deficit currently within the humanitarian sector and within the development sector because it's not our bread and butter, it's not what we do, definitely not what we can do now because of all the cuts and everything that's been happening. But the reality is that these investments, wfp, Nassim, just to call out WFP and the investments that they're making with you and your team and the folks that are doing it. So there are certain organizations that are making the investment, I think, which is a really positive thing. But we have a long way to go. And so I want to thank you, Richard and Gabrielle for joining us today. Nassim, I want to leave the last and final question to you to give it to the panel and it's been such a meaningful conversation today. I'm just really mind blown over to you, Nassim.

[C]

Yes, I really enjoyed the conversation today, especially that I resonate with a lot of it. I loved how Gabrielle just closed the conversation talking about the principles and defining the principles even within our own organizations, within our own businesses and developing the capacity to roll them out. Right. And I think that's very critical and I can say that we do that in wfp and I'm sure that a Lot of organizations are in the same boat. So I guess my last question would be what is your message and what is your hope for the future of AI, Especially considering our podcast that it is for humanitarians. And a lot of our audience are humanitarians, some of them still skeptical about AI and is thinking about the dystopian future and some of them overly excited. So over to you and thank you so much.

[E]

Yeah, I think. Well, I want to start with fairness, transparency, accountability, human well being, not just about AI outputs. And if AI ethics continues to kind of ignore real world externalities, it risks becoming a discipline that just legitimizes the expansion of AI into all areas of life. And for humanitarian orgs, definitely not saying there should be no AI. There's clearly scenarios where AI systems, if responsible, designed and thoughtfully implemented, can significantly enhance emphasis on enhanced human capabilities and humanitarian missions. But I think I just want to be clear that AI, or should I say more, more complex AI is not an inevitability if it does not enhance these capabilities and real kind of ethical deliberation. It demands asking not just how to ethically implement AI, but whether certain applications of AI should exist, exist at all. It's inherently social, it's political, and it's a moral judgment rather than a technical one. So even if you have robust fairness and transparency standards, and that's not even a guarantee of the AI system being fair or equitable. Right. Sometimes the very act of automating inherently moral and altruistic choices like humanitarian work, it may signal that efficiency is more valuable then kind of a more compassionate, deliberate understanding which can erode trust and importantly humanitarian service uptake by people on the ground, but ultimately social cohesion.

[D]

Thanks.

[C]

It was a pleasure.

[A]

Thank you for joining us on humanitarian frontiers in AI. We hope today's conversation gave you new insights into how AI is transforming humanitarian efforts and the steps we need to take to ensure sure it's done ethically and effectively. If you enjoyed this episode, be sure to subscribe and stay tuned for more discussions with leaders and innovators at the intersection of technology and humanitarian work. Together we're exploring how AI can bring real change to communities in need. Keep pushing the frontiers of possibility.