SCAI Question 11

A Framework For Effective AI Adoption For Social Good

How can AI adopters effectively evaluate and apply AI models for social good?

Context & Assumptions

AI developers often focus on improving technology, while governments regulate AI to address societal risks like misinformation or crime. Yet, there is a gap in effectively integrating AI into social good applications by governments, Non-Governmental Organisations (NGOs), and social enterprises, and a lack of thorough evaluation to measure their real impact.

In the private sector, AI adoption is measured by revenue and costs, making it easier for adopters to assess impact. However, in the social sector, evaluating outcomes in areas like education, healthcare, or climate change is more complex and lacks sufficient financial and technical resources for analysis. An additional complication is when adopters pick up a successful use case from one sector or context and apply it to their own, without evaluating whether the model or outcomes are still relevant for their situation. This complexity means a higher reputational risk for AI companies and greater potential harm from poorly designed programs, placing more responsibility on the AI industry for effective adoption in social sectors.

The risks of early AI adoption include wasted investments and negative outcomes for participants, alongside reputational damage and potential overregulation for the AI industry. With growing interest in AI for social good, it is vital for the industry and adopters to develop a framework focusing on learning, piloting, evaluating, and capacity building.

For example, AI can boost teacher productivity and student learning in education, improve patient outcomes in healthcare, and provide better farming recommendations for climate change. However, rapid implementation without proper impact assessment can lead to negative consequences like reduced learning outcomes, incorrect health advice, or crop losses.

Question

How can AI adopters effectively evaluate and apply AI models for social good?

How can we offer a sociotechnical framework to AI adopters that enables them to:

• Accurately assess various aspects of AI models for social good use cases, including the dependencies (such as access to computational resources), utility (like model readiness and alignment with the proposed use case), and appropriateness (for instance, determining if general-use models are suitable for the intended purpose)?
• Pilot and rigorously evaluate AI use cases to comprehend their true impact compared to existing programs, ensuring that AI is being effectively adopted for social good?

Important considerations:

• It is crucial to emphasise socio-economic, cultural, and other differences that might lead to unintended consequences or worsen inequalities and biases when applying general-purpose algorithms in social sectors like education or public healthcare.
• AI adoption should not be rushed. It is essential to first pilot and rigorously evaluate AI-based interventions in real-world settings.
• The AI industry needs to allocate financial resources to support the “Framework for Effective AI Adoption for Social Good” proposed here. This includes funding for accessible in-person and online training for social sector organisations on integrating AI into their programs effectively; for conducting thorough evaluations of use cases; and for sharing case studies that highlight both successes and failures in these applications.

Indicators of Progress

We suggest the following potential approach:

• A Framework for Effective AI Adoption for Social Good requires that adopters:
  • Collaborate with the AI developer to gain a clear understanding of the potentials and limitations of the AI models, including the data they were trained on, the values underlying that model, and relevance to the local context.
  • Engage with researchers who possess global insights into both effective social program design and AI integration. Involve them from the design stage to develop a pilot for the AI-enhanced program.
  • Initiate a concurrent, independent, and rigorous evaluation, such as a randomised control trial (RCT), of the AI pilot. This is to accurately measure its impact compared to existing methods, offering insights into what works, what does not, and why.
  • Proceed to scale up the program only after integrating learnings from both the pilot and the concurrent evaluation. If the pilot is found ineffective, consider scaling down.
  • Disseminate the insights gained from this pilot and its evaluation to others who could benefit from integrating AI into their programs. Share through blogs or other open-source platforms.
  • Address the needs of governments, NGOs, development organisations, and social enterprises in the context of social good applications (not commercial organisations and applications). These entities often lack the technical or financial resources for the above analysis, which is more challenging as outcomes are not solely measured in terms of revenue and costs.
• Known challenges or obstacles to answering this question include:
  • Program implementers often lack access to the global knowledge that can offer valuable insights for effectively incorporating AI into social programs. Researchers, with better access to this information, can be crucial partners in this process.
  • There is sometimes an underestimation of the importance of tailoring the program to the local context. It is also vital to refine any elements that did not perform as expected.
  • Even when social sector adopters recognize the importance of these steps, they frequently lack the technical capacity or financial resources to design and execute such a pilot and its thorough evaluation effectively.
  • Tech and AI companies mostly do not allocate resources to assist adopters of their technology for social good. It is crucial for these companies to support social sector adopters in navigating this framework, ensuring that only relevant technology is adopted, and that it is done so correctly and appropriately.
• Broad criteria for recognising progress in answering the question:
  • The AI industry and adopters broadly use and adapt this “Framework for Effective AI Adoption for Social Good”.
  • The AI industry commits both financial and technical resources for:
    • Implementing a sufficient number of use cases across various sectors such as education, public healthcare, climate change, social protection, labour markets, and agriculture, and in diverse socio-economic and geographic contexts.
    • Publishing open-source studies that summarise the experiences of early adopters, including evidence from rigorous evaluations of the above use cases. These studies should focus on understanding what works, what does not, and why.
  • Building capacity of key actors in the social sector to effectively adopt AI in their programs. This can be via a combination of in-person training for government AI/IT departments and multilateral development banks (like the World Bank and Asian Development Bank), as well as virtual training for relevant NGOs worldwide.
  • Over time, as tech and AI companies, adopters, and researchers gain a better understanding of when and why AI is an effective tool in social sector programs, we expect to see fewer products exited and fewer pilots considered unsuccessful.