Google report assesses how AI can best serve humanity

Google.org is today releasing a report that details how social impact startups, nonprofits, government policymakers, and academics can use machine learning to address some of humanity’s biggest problems.

The assessment is the result of analysis of more than 2,600 applications — from 119 countries — to the $25 million Google AI Impact Challenge, a global open call for projects that aim to use AI to serve humanity.

Google used AI to arrive at its conclusions about how human-centered organizations can better use AI, specifically tapping natural language processing and clustering analysis techniques to review applications and internal assessment of applications. The AI analysis was supplemented by interviews with impact challenge applicants and recipients.

Today’s assessment of challenges and opportunities appears to be the first such analysis from a tech giant on how civil society can improve its implementations of AI.

In May, 20 organizations from various parts of the world received Global AI Impact Challenge grants, following the initiative’s debut last fall.

Recipients include:

  • Colegio Mayor de Nuestra Señora del Rosario in Colombia is using computer vision and satellite imagery to detect illegal mining operations known to contaminate local drinking water.
  • New York University and Fire Department of New York (FDNY) teamed up to reduce response time for 1.7 million annual emergency service calls.
  • American University of Beirut in Lebanon will use machine learning in an attempt to help farmers save water used for crop irrigation and food production.

But in assessing the 20 funded organizations to see how much cloud credit would be useful for each, it became evident that their approaches were not always best for their particular project, including missed opportunities to do compute on-device versus in the cloud, Google.org principal Brigitte Gosselink told VentureBeat in a phone interview.

Ongoing initiatives that may tie into the report released today include efforts to help social impact organizations gather and label data, as well as collaboration with Google Cloud engineers to ensure efficient use of compute power.

The report strongly recommends the creation of partnerships between organizations to reduce redundancies and pool resources. For example, more than 30 applications received for the Global AI Impact Challenge focused on the use of AI to identify and manage agricultural pests.

“We believe partnership between organizations with deep sector expertise and organizations with technical expertise are the most actionable near-term opportunity to develop and operationalize the use of AI for social good,” the report reads. “Common forums where organizations interested in using AI for social good could share missions and needs with technical experts may help facilitate connections.”

Partnerships can also help address the data labeling and data collection challenges some applicants encountered.

The top recent example highlighted in the report was a Facebook Department of Defense CrowdAI effort to create xBD, a labeled data set of 700,000 satellite images after eight natural disaster events. This data can be used to train AI systems to identify areas most in need of assistance.

Applicants in crisis response, economic empowerment, and equality and inclusion sectors were more likely to face a lack of meaningful data sets, while those working in health, environmental, education, and the public and social sectors were more likely than others to already have access to necessary data.

Google’s review of applications also uncovered a lack of experience deploying AI systems. This is consistent with the report’s finding that 55% of nonprofit and 40% of for-profit social enterprise applicants had no previous experience with AI. The report also frequently found a fuzzy understanding of the limitations of AI and of what’s required to bring a project from concept to completion.

“Many applicants that were new to AI needed a better understanding of the types of data roles required, and others found it difficult to compete with private sector organizations when hiring technical talent,” the report reads. “Throughout our review process, we saw that even the most mature organizations underestimated the time and resources needed to prepare and maintain the data for algorithm use.”

Data science talent shortages that afflict many businesses are especially acute in the social impact world, but private businesses that typically have more resources or in-house IT teams also encounter challenges. A fall 2018 PricewaterhouseCoopers survey of executives in more than 60 countries found that only 4% say they have successfully implemented AI for their organization.

The report makes no direct mention of compute power limitations, but a lack of compute power could reduce the kind of model experimentation that can lead to breakthroughs.

“One of the things we’ve been talking about as we watch the process with our winners is actually doing a follow-up deep dive and working with some of our cloud engineers to maybe build out a little bit of a framework we can share with the [social impact] sector that speaks to … the kind of questions you’d want to ask to do this kind of audit of your own cloud requirements, so that you’re being as efficient as possible,” Gosselink said.

While the report evaluates roadblocks and potential opportunities, it makes no attempt to offer data-driven assertions about the kinds of efforts that offer high return on investment for funders or have the highest impact on human lives.

“I think the jury’s still out on that question to some extent. Part of the reason that we chose such a diversity of winners for our AI challenge, and frankly part of the reason that we did the challenge in the first place, is that I don’t think we as a society know yet where AI is most effectively applied for the highest leverage on social impact outcomes,” Gosselink said. “To some extent, I think we’re still trying to identify those highest use case applications across the board.”

Some other illuminating facts about Global AI Impact Challenge applicants:

  • More than half of the applications came from organizations with fewer than 25 employees.
  • Proposals received as part of the Global AI Impact Challenge touched upon all 17 of the United Nations’ sustainable development goals.
  • Health projects represented more than 25% of the pool of more than 2,600 applicants, while projects to help the environment, education, and economic development were also common.

For lawmakers and government officials writing or enforcing policy, the report encourages more public education efforts and training grants to speed progress and overcome limitations that come as a result of technical talent shortages, and subsidies to support investment in physical infrastructure underpinning AI in regions where it is lacking.

Finally, some applicants created plans to test data sets and model outputs for bias in their applications, but, ironically, the report also asserts that some humanitarian or social impact organizations need to improve their assessment of AI solutions to better understand potential negative use cases.

“Understanding and managing responsible AI use is both difficult and non-negotiable,”the report reads.

The fact that the challenge was only circulated in English and reliance on nongovernmental organizations to share the Google Global AI Impact Challenge application limit the scope of results.

The Global AI Impact Challenge concludes in February. Google has not yet determined whether a second challenge of $25 million will be held.

Google isn’t alone in championing of AI for good initiatives. Microsoft also has AI for Earth, accessibility, and humanitarian action initiatives underway, and the United Nations is exploring a range of AI for good initiatives, including how AI can be used to achieve Sustainable Development Goals (SDG).