AI beyond the hype: What nonprofits really need 

Remember when we were all told to drop everything and learn about blockchains? Or the metaverse? Generative AI (genAI) is having that moment, but it’s just the latest piece of technology. It won’t eliminate intractable social problems, greatly increase your nonprofit’s fundraising, or solve your personnel problems. All the challenges of applying tech in nonprofits still apply.  

Every nonprofit leader knows those challenges. Funding is tight and donors don’t love paying for technology. In part due to that underinvestment, the hardware and software tend to be out of date. Many nonprofit staff and volunteers lack tech skills and, at the same time, resist change. It’s hard to make the leap to genAI use for programs if you have no data to speak of, or if your team lacks data skills. 

GenAI use for grant writing: The stuff that mostly works today 

Even with these challenges, many fundraisers have been using genAI tools for grant writing and similar tasks. I think of these tools as “spellcheckers on steroids” that go beyond correcting spelling or grammar into rewriting a 500-word answer to fit in a 250-word grant application box. But remember. just as spellcheckers often suggest you correct things that are right, genAI “makes stuff up,” without regard to reality. Making up a non-existent donor spouse or an incorrect factoid about your programs can have consequences. GenAI for grant writing works well only if knowledgeable staff reviews the output.  

GenAI use for programs: Don’t inflict bad AI on the people you serve 

The more exciting—but also riskier—uses of genAI are in programs. This is a lot more challenging than asking for help with a grant proposal. 

In one infamous example, the National Eating Disorders Association fired all of their human helpline counselors and replaced them with Tessa, the genAI chatbot. After screenshots were shared of Tessa telling people to start counting calories, the opposite of modern practice for eating disorders, it was shut down after a week. Tessa’s use of massive amounts of data from the internet to offer dieting advice (rather than experts) could cause harm to people with eating disorders.  

This underlines the key limitation of general-purpose genAI solutions like ChatGPT, Gemini, or Claude: They parrot the data they’re trained on, without understanding of or empathy for the individual user. They’re trained to sound more convincing than a person, even when they’re 100% wrong.  

The key elements of successful genAI uses for specific program needs are guardrails and context. Guardrails are there to stop harmful answers from getting through to community members. Context trains the genAI tool to respond based on trustworthy content from your organization or your field. Unless there’s a product out there that’s designed for your use case, it’ll cost too much time and money for the typical nonprofit to get this right. You wouldn’t dream of putting an untrained volunteer into the middle of a life-or-death situation; you shouldn’t put a general-purpose genAI tool in your programs based on the false belief it knows what it’s doing. It doesn’t.  

The ‘right’ genAI tool for your nonprofit 

So, how can you assess your AI use cases and needs? First, do the basics: Create reasonable internal policies around genAI use, like not sharing confidential data with commercial AI companies. The nonprofit tech incubator Fast Forward has a basic free AI policy builder.  

Second, improve your organization’s ability to collect and use data to enhance your programs. Data you collect now about your programs can be used to train AI applications to work better in the future. Look for tasks that resemble repetitive drudgery, such as required data entry, as likely candidates for genAI augmentation.  

Above all, 99% of nonprofits should wait for a product to be developed that they really need and will truly benefit their work. By the time a genAI solution is packaged in a product, it should have been tested with and built for organizations like yours, with guardrails and context. You’ll be able to talk to other organizations about the product: If two or three sensible nonprofit peers are raving about a genAI deployment, it’s probably worth looking into.  

AI-enabled tools are also being developed by tech nonprofits for use by less technical organizations. A nonprofit tech partner who shares your mission focus and values is more likely to prioritize safety for your communities and to understand the context in which you work.  

We’ll have within years products that make it easy to feed your vetted content to help answer questions with more context and guardrails. However, you’ll always need to invest the time to supervise them, as they’ll inevitably make mistakes. You always owe it to your community to ensure new technology is doing far more good than harm.  

What to expect from genAI for nonprofits going forward 

We should see steady feature improvements in genAI tools, not only for development tasks such as grant writing but also for common nonprofit needs like social services case management.  

Larger nonprofits with the resources to develop their own genAI solutions are already showing possibilities. Quite a number are using an approach known as retrieval-augmented generation (RAG) to add context and guardrails: An organization might feed a few thousand questions asked by clients and answers created by expert staff, and ask the RAG solution to model its answers after the expert answers.  

These are still early days. Nonprofits should wait for the industry to figure out what organizational applications make the most sense, and then adapt those to the social sector context. Similarly, watch larger nonprofits for successful experiments, and wait for products based on those to arrive for you to test on your needs.   

Photo credit: Jacob Wackerhausen via Getty Images