What’s Enough, What’s Safe, and Who Decides? Reflections on “AI for Good” 

The imperatives that AI brings forward are connecting science, academia, technology, and policy more than ever. Recent discussions at the AI for Good Summit reflect the need for an ecosystem approach on how AI can be applied thoughtfully, efficiently, and inclusively.

This approach aligns with Caribou’s own thinking on who shapes AI’s trajectory, and how. Our report on AI Innovation Ecosystems highlights the importance of the academic community, big tech, and policymakers working together, not in silos.

As an organization committed to identifying the opportunities that digital ecosystems hold to support inclusive and sustainable societies and economies, the AI for Good Summit offered Caribou space to reflect on how AI integrates into the next wave of technological advancement and transformation.

These are my reflections on the recurring themes that stood out as key to shaping the wider AI-for-good landscape.

The ITU Secretary-General’s warning during her opening address resonated throughout the Summit and brought forward the imperative of “more with less.” As Clemence Kyara, CTO at Code for Africa, put it: sometimes there can be “just enough tech,” and instead we should focus on the affordability of solutions and what technology is actually needed to solve the problem. A need to ensure practicality over hype.

Kyara also raised the question of data overloading when developing solutions. Getting access to public data, processing it, converting it into machine-readable data, ensuring its quality, and making it accessible and free is costly. The question is then what data is needed and what is worth keeping. 

This point resonated with David Cox’s, VP for AI models at IBM researcher, presenting how smaller versions of large language models (LLMs) as a software can be adequate, less energy- and cost-hungry, and safer, as opposed to LLMs such as ChatGPT requiring ever-growing models. 

The “less is more” imperative also aligns with scientists’ positioning, a community that reminds us that nature holds the keys to autonomy and efficiency in complex environments. Singapore University presented their in-depth study of the movement of octopus arms in order to deliver an AI submarine exploration device as the most efficient solution.

In an environment of urgency, another imperative stood out: the need to iterate. Aleksandra Bilika of Help.NGO stressed the need for rapid prototyping with “no regrets”: test AI agents and chatbots, learn quickly, and keep what works. Fermín Selva, CTO at CRIES, echoed this with a reminder that safe spaces for testing are essential. That means sandbox environments and closed-loop agents where sensitive data is protected. 

Policy actors also highlighted the need for an iterative approach to regulating AI given the pace of technology’s advancement. Given the current geopolitical fragmentation on the “right” approach to adopt in terms of AI governance and safety, iterating instead of looking for the perfect solution in one go will be key. Yann LeCun, Chief AI Scientist for Facebook AI Research, also presented his belief that the technology will advance progressively, allowing for innovations to be designed for safety.

Additionally, policy actors outlined that not all sectors should be regulated under the same policy. The need for sector-specific regulation echoes the position of another AI scholar, Helen Toner,  interim executive director at Georgetown’s Center for Security and Emerging Technology. In one of her lectures on “Who’s Actually Governing AI?”, Toner underscored that it is impossible to do policy for “all of AI,” highlighting the need to focus on sector-level AI policies, as AI in healthcare will look different from AI in automobiles.

Many consortia were announced during the Summit, including the UN-supported AI Skills Coalition and AI Hub, and the AWS-supported CTO cohort.  However, beyond these specific expert partnerships, the third imperative that came across different sessions was the need for society at large to be more involved. As Geoffrey Hinton emphasized, the public “needs to know enough” to put pressure on politicians to regulate responsibly. It’s not just experts who should be making calls; communities should judge whether tech truly serves them, tying inclusion to lived realities. This point resonated with speakers in the humanitarian sector and policymakers.

These three imperatives will stay with me as Caribou continues conversations with its partners and clients about designing inclusive AI solutions, considering what is enough, what is safe, and who gets to decide.