Evaluate whether AI is right for your nonprofit organization, and discover how you might take advantage of powerful machine learning technology.
You’ve seen it in your newsfeeds, in your inbox, and in action movies — artificial intelligence, or AI, is having a moment. The chatter is everywhere, and it can be challenging for organizations, especially those that tend to have fewer technical resources on staff, to know how these new tools might help (or hinder) their work. While machine learning is increasingly used by companies across the private sector, few nonprofits have integrated this kind of technology into their operations to date.
What is AI and why does it seem like it’s everywhere these days?
The variety of terms can make it difficult to understand the AI landscape.. AI is the process of using computers to help perform tasks automatically that could previously only be done by humans. To do this, we use a process called machine learning that allows us to “teach” computers how to tackle these tasks. Most machine learning techniques revolve around simple arithmetic. But by using modern computer processors to perform these calculations at high volumes, data scientists can create models capable of solving complex problems.
What kinds of problems can machine learning solve?
Many AI models take a series of input data points and use them to predict aspects of new data points. For example, USA for UNHCR is using images of refugee camps to create models that can predict how different buildings are used. This makes it easier to figure out whether enough services are being provided to support those living in the camps and how resources might be allocated most effectively. Similarly, Benefits Data Trust analyzed data from callers looking to enroll in public food assistance benefits in Philadelphia and created a machine learning model to predict which benefits might be most helpful. This dramatically reduced the time to get benefits to those in need. Stories like these sound almost like magic, but they represent a lot of hard work!
How can you tell if your organization might be able to use it?
Consider the following questions.
1. Do you have an appropriate question in mind?
Many organizations considering an AI solution skip this step and assume that they’ll be able to take whatever data they have and “find something useful in it.”
Trying to create a machine learning model without a clear question in mind is like trying to take a trip without a destination in mind.
You might wind up somewhere, but there’s a high likelihood you’ll spend a lot of time and money driving around aimlessly. Typically, questions that are well-suited to machine learning include those that seek to predict a category (“Given a dataset of water points, which ones are likely to break in the next year?”) or a number (“How much money is this donor likely to give next year?”).
2. Do you have access to useable data?
A solid dataset is the one of the most important requirements for a useful machine learning model. Unfortunately, it can also be one of the most difficult inputs to gather. Machine learning models that make predictions to answer their questions usually need labeled training data. For example, a model built to predict if an email is spam will need to be trained on a dataset of emails with labels indicating whether each training example is spam or not. Labeled training datasets can be tricky to obtain and often require creativity and human labor to create them manually before any machine learning can happen.
3. Do you have access to the people to build and maintain the models?
Once you have an appropriate question and a rich training dataset in hand, you’ll need people with experience in data science to create your models. Lots of work goes into figuring out the best combination of features, algorithms, and success metrics needed to make your model effective. This can be a challenging and time-consuming process and requires constant attention to maintain your model over time.
4. Do you have a plan for managing issues around ethics, bias, and interpretability?
Machine learning models reflect the underlying biases and inequities implied in the data used to train them. If not designed mindfully, AI can reinforce these problems in many ways. Even something as seemingly simple as a binary column in a dataset can obscure nuance that data scientists can’t ignore. As a result, it’s critical that organizations considering using AI carefully consider the ethical implications of the models they create, especially how they’re created in partnership with the communities they seek to serve and how they address the diversity of modern society. Interpretability techniques used throughout the design of an AI can also help identify and reduce the bias present in your data and make it clear why your algorithms are behaving the way they are.
5. Is it worth the trouble?
Given the challenges implied by the previous four questions, organizations must ask whether or not the potential benefits that machine learning might bring are outweighed by the costs and difficulties of creating and maintaining an ethical model.
Does answering “no” to any of those questions mean my organization can’t use machine learning?
Maybe. If you don’t have a good question in mind, your organization is probably better off focusing first on clarifying its theory of change. Similarly, if your organization doesn’t currently gather and store data in a structured way, investing in core data collection and reporting tools should be a higher priority than any machine learning models. If this is new to your organization, tools like Airtable can make getting started easier.
What if you’ve got data and good questions, but don’t have the data science expertise on staff to design and implement a machine learning model, much less maintain it over time? Organizations like DataKind or Hack4Impact support networks of volunteers who can help get nonprofits up and running with new AI or machine learning tools.
Another option for organizations interested in using AI to deliver on their mission is DataRobot’s new AI for Good program in partnership with GlobalGiving. DataRobot’s team of experts will assist in identifying and refining potential machine learning projects from your data, and guide your organization through the creation, implementation and maintenance of predictive models using their standalone platform. DataRobot’s tools are unique in their ability to automate the most time-consuming components of the machine learning process, such as feature engineering—figuring out the ways to combine data to form the most useful inputs to a machine model—and algorithm selection—figuring out the best type of model to use to achieve your machine learning goal. This automation and DataRobot’s expert resources can potentially tip the answer to “Is it worth it?” from a “no” to a “yes.” If your organization is interested in being part of the second cohort of AI for Good: Powered by DataRobot participants, look out for the opportunity to apply, coming soon!
Is your nonprofit interested in using machine learning to grow your impact? Learn about DataRobot’s AI for Good program.
Featured Photo: Send 100 Girls to STEM Camp in Nigeria by Women's Technology Empowerment Centre (W.TEC)