Shaping responsible AI with dignity

How do we ensure that in our work to harness AI’s potential, we don’t inadvertently undermine the very values that guide our work. 2025 might be remembered most of all as the year of AI. It is an essential pillar of IDinsight’s new strategy. Part of that strategy involves figuring out these exact challenges.

At IDinsight, we believe the answer isn’t to slow down or retreat from AI. Instead, it’s to lead thoughtfully. We are committed to ensuring that as we embrace AI solutions for social good, we do so with rigor, transparency, and an unwavering commitment to upholding human dignity in everything we build and support.

AI for good in IDinsight’s strategy

With our deep understanding of data and evidence in the sector along with our close proximity to implementers within and outside government, we at IDinsight feel we have a strong opportunity and responsibility to contribute meaningfully to the global effort of advancing AI for social good. 

At IDinsight, we aim to do this through three critical pathways:

1. AI evaluations: We are taking our expertise in evaluations and adapting/innovating it to help organizations evaluate the impact of their AI solutions. We are not just measuring overall impact but are keenly interested in critically understanding who benefits, who gets left behind, and what unintended consequences emerge. Our approach balances the need for rigorous evidence with the reality of rapidly evolving technology. You can read more about it in our recently published blog post. 
2. AI products and tools: We are working to build and scale AI powered solutions that can address common sector-wide solutions. These products will be developed in close partnership with the communities and organizations that will use them ensuring they are grounded in real-world needs and contexts. Do check out our latest release, Evidential – a free, open‑source experiment engine purpose‑built for nonprofits
3. AI advisory: We are working alongside partners to help them navigate the complex landscape of AI adoption. We are supporting our partners in many ways, from identifying where AI can genuinely add value, to surfacing risks and to building organizational capacity for responsible implementation. We recently launched a Fractional AI Engineer program to support early stage organizations with AI adoption and scaling. 

Beyond our external work, we are also committed to thoughtfully integrating AI within our own teams. We recognize that leading this conversation requires us to model the practices we advocate for. This involves experimenting carefully, learning from our experiences, and being transparent about both successes and failures.

What does responsible AI mean for us

“We need to bring a dignity and inclusion lens to our AI evaluation work. Asking how people experience these tools is just as important as measuring reliability.” 

 

– IDinsight CEO, Becca Gong Sharp, at UNGA80

While AI’s potential to transform development outcomes is real, we are also acutely aware that this potential comes with profound risks. The stakes are simply too high to move forward without careful consideration of what could go wrong. 

The challenges are both technical and deeply human. AI systems can embed and amplify existing biases and thus systematically disadvantage already marginalized groups. They can erode trust by replacing human relationships that are often the foundation of effective interventions with impersonal interactions that leave people feeling devalued. They can create new inequalities where those with resources, digital literacy, and motivation benefit while others fall further behind. They can constrain human creativity and judgment by replacing the expertise of frontline workers with rigid standardization that misses critical context.

In practice, these risks could show up in real-world settings in the form of many unintended consequences. Education AI tools intended to assist teachers, for example, might unintentionally reduce their flexibility to tailor lessons for their classrooms. Health chatbots developed to expand access could weaken the relationships of trust between health workers and communities. And agricultural apps that benefit digitally connected farmers might further disadvantage smallholders with limited access to technology.

But recognizing these risks doesn’t mean rejecting AI altogether. It means building differently. At IDinsight, we are committed to developing and supporting AI that is fundamentally grounded in respect for human dignity. This means going beyond standard ethical checklists to ask deeper questions: Does this technology enhance people’s sense of agency and autonomy? Does it recognize and value their existing knowledge and relationships? Does it serve to empower rather than replace? Does it protect the most vulnerable from harm?

Our long-standing work on dignity along with our expertise in doing ethical and locally rooted research provides the foundation for this approach. We are extensively working to adapt these learnings and practices into the AI realm and constantly reflecting how this would look like as new innovations and solutions emerge. 

Putting dignity into practice

Translating our commitment to dignity into practice in these emerging areas means anticipating ethical challenges and unintended consequences before they occur. Rather than reacting with protective measures after harm has happened, we approach these risks as inherent—not exceptional. We design and build with the assumption that unintended consequences can always occur. It helps us be proactive and ensure we consider all possible kinds of harm that the AI solutions can cause now and in the future. 

Our understanding of dignity in AI is far from final. A review of eighty-four AI ethics instruments classified eleven commonly referenced principles, one of which was dignity. But while many of these guidelines refer to dignity, most fail to define it clearly. An analysis of existing case law and treaty interpretations identified four major conceptions of dignity:

• Non-instrumentalization: AI should serve human purposes without reducing individuals to mere instruments for achieving those purposes. 
• Protection of vulnerable populations: AI ethics must prioritize the well-being and dignity of vulnerable populations and ensure that they do not exacerbate existing inequalities.
• Recognition and exercise of self-worth and autonomy: AI systems should respect and promote individual autonomy, empowering users to make informed decisions and maintain control over their lives.
• Wider notions of protecting humanity: AI must consider the long-term implications for the human condition. This involves careful reflection on how AI could impact human skills, social interactions and our very understanding of ourselves.

We find those valuable. Yet as we do this work, we have valued the thinking of scholars like Lorenn Ruster, Sue Anne Teo and Rose Mutiso, as well as collaborators at IDinsight, CGD and Agency Fund, who have encouraged us to think beyond a purely protective stance, to look at how technology will affirm dignity, and the hard material and technical constraints that may prevent this.

That leads us to a set of ideas that inform the Dignity Initiative’s thinking, as we engage with partners at IDinsight and beyond:

• Protect what makes interventions work: Before introducing AI, we need to identify the core elements that drive impact in existing programs. If AI risks undermining these elements, we need to either redesign the solution or honestly acknowledge the trade-offs involved. Efficiency gains mean little if they break the mechanisms that actually change behavior.
• Measure people’s experiences: AI evaluations often default to metrics that are readily quantifiable like engagement rates, task completion, cost savings,etc. We realize to understand true impact it might require us to also assess: Do people feel respected in their interactions with this system? Do they feel listened to? Has their autonomy been enhanced or diminished? These experiences are harder to measure but they are essential for understanding whether we are truly serving people’s dignity or inadvertently undermining it. When we do so, we can consider where efficiency and dignity are in tension, and where they are mutually supporting.
• Ensure human oversight and preserve human judgment: Where AI recommendations affect high stakes decisions that impact people we must ensure that sufficient human oversight exists and that the final call on these decisions are based on human judgement. This means providing humans with information beyond just the AI’s recommendation, creating space and time for deliberation, ensuring humans have the authority to override AI and holding humans accountable for outcomes. 
• Build in participatory processes: Stakeholders affected by these AI systems should have genuine power to influence AI system design. And this participation should be meaningful, not tokenistic. This should go beyond providing feedback and involve meaningful co-design, on the models pioneered for public policy by Hilary Cottam and others.
• Design for heterogeneity: AI solutions that work “on average” can still cause significant harm to specific groups. We must intentionally design for diversity across literacy levels, digital access, languages, disabilities, cultural contexts, etc. We should measure and analyze how effects differ across subgroups while paying particular attention to who might be left behind or disadvantaged.
• Transparency: People deserve to understand how AI systems that affect them work. This includes information on what data is being collected, how decisions are made, and what the limitations are. Transparency isn’t just about technical explainability. It’s about communicating honestly with affected communities in ways they can understand.

These are not fixed rules yet but an evolving set of reflections we are actively testing, refining, and checking ourselves against as we navigate AI’s complexities. 

Moving forward together

The principles guiding responsible AI in development are still being written. The standards are still being set. The choices we make now about how to design, evaluate, and deploy these technologies will shape development outcomes for millions of people in the years to come.

The question isn’t whether AI will transform the social sector. It already is. The question is whether we will do so in ways that truly serve the communities we are meant to support, that protect and promote their dignity, and that make their lives genuinely better.

We are committed to ensuring the answer is yes. And we are eager to work alongside partners, communities, and colleagues across the sector to make it happen.