Holding
"AI for Good" Accountable
The “AI for Good” movement is going unchecked. The tools at its center (LLMs) are trained on data collected without consent, come with a significant environmental cost, and produce biased, incorrect, and even dangerous outputs.
Yet they're pushed onto vulnerable communities (who absorb the risks) in the name of social impact, despite having little evidence that these solutions are even wanted, let alone helpful.
The Kaleidoscope Project was started to uncover that evidence and push “AI for Good” towards community authority and governance, honest measurement of both benefits and harms, and the minimization of dependence on LLMs and Big Tech.
>$15M DE-RISKED AS DILIGENCE LEAD
>100M DE-RISKED AS DILIGENCE SUPPORT
>200 LEARNERS PROVIDED TECH LITERACY EDUCATION
OUR GUIDING PRINCIPLES
01
Evidence & Transparency
02
Big Tech Independence
03
Agency & Ownership
04
Surface Collective Input
05
Invest in Foundations
01
Evidence & Transparency
Honest measurement should move on from vague metrics like “Lives Impacted” towards quantifying the “Net Effect” (benefits minus harms) of an AI’s deployment. Numbers can’t fully measure impact so they must be paired with longitudinal, depth-based interviews. Transparency in this measurement and other aspects (e.g., policies, data retention) helps to ensure accountability, lift others and move towards field-wide standards.
02
Big Tech Independence
Partnerships between Big Tech (especially LLM developers) and CSOs require scrutiny. Big Tech gets increased usage from embedding their tools in core social services and an improved reputation from associating with impact, while CSOs risk theirs on widely distrusted tools. That distrust is warranted given opaque data practices and model interpretability, contracts with surveillance/"defense" tech and the unilateral power to change core functionality or pricing at any time.
03
Community Agency & Ownership
There’s an enormous gap between those who decide to build, and especially fund, “AI for Good” and those who interact with it. When AI replaces a person in a service that someone depends on, a choice is forced onto them. While the lack of consent is problematic alone, it’s compounded by undisclosed use, retained data, and the lack of standards around quality. Even our own attempts at comprehensive evidence frameworks can’t encapsulate the nuanced benefits of human interaction. Regardless of how good any builder, funder, or evaluator thinks an AI tool is, the people who will use the tool should have governance power to decide how and whether it’s deployed and clear ownership of their data.
04
Surface Collective Input
A founder’s “lived experience” is important, but it doesn't reflect the diversity within a community. What’s more critical is community input at scale. “Community design” is an established principle in social impact but is often treated as a checkbox or afterthought. True community input should happen before prototyping so they can shape the foundations. This community involvement should be codified into governance mechanisms and evaluation methods so that it can’t be overlooked when inconvenient.
05
Invest in Foundations
While AI is sold as a leapfrog technology, no AI application stands alone. Most require devices, internet access, and digital literacy. Without these foundations, AI can exacerbate divides within a single community. Before deploying AI, first audit and, if needed, fund the foundations. This includes education about what happens to people’s data and where these tools fail. Investing in these foundations also avoids the “saviorism” trap of deploying a culturally ignorant tool or creating unforeseen maintenance burdens. The right investment should give communities power and capacity to innovate on their own.
How We Work
We partner with philanthropies, governments, or anyone else navigating social impact in the age of AI.
Diligence
Comprehensive, technical risk and opportunity assessments of "AI for Good" philanthropic investments.
Policy & Education
Creation or education around responsible AI governance frameworks for governments or CSOs.
Capital Allocation
Development of fund strategy or distribution of capital for tech for good solutions to reach your impact goals.
AI Evaluations
Development of automated assessments that measure an AI solution's ability to reach your impact goals with a quantifiable score.
Our Impact
>$15M
de-risked as diligence lead
>$100M
de-risked as diligence support
>200 learners
provided tech literacy support