Ed. note: Throughout the past two years, the Goldhirsh Foundation and its LA2050 initiative have provided complimentary AI (artificial intelligence) consulting, training, and workshops to the Los Angeles impact community and beyond. The work is led by Jen García, Goldhirsh Foundation’s AI EIR (executive-in-residence). Below – and in other posts available in the AI section of our news feed – García shares practical insights, real-world use cases, and emerging research to support nonprofits’ responsible and ethical adoption of AI tools.

As you can imagine, there was much to cover when 75 social impact leaders gathered this month at the headquarters of the Fox Foundation – an LA2050 Grants Challenge funding partner – for two-and-a-half hours of AI fluency training.

Here are three takeaways from the session, based on the questions and comments in the room:

1. Your approach to learning AI tools is important. I hear a range of self-talk bubbling up to the surface in conversations during our trainings, ranging from “I’m bad at tech” to “I am old, so I don’t learn as quickly.” And although these admissions may feel true, we need to reframe our approach entirely. Let’s trade any limiting self-talk with curiosity, playfulness, and hope. More effective word choices may sound like:

• “How can I apply what I know so far about AI to my current problems in my org?”
• “How exciting that a new technology has popped on the scene at this point in my career. So much to learn!”
• “Well, that attempt didn’t work. Let me try another way.”

This type of approach is one we urge our young learners to take – let’s engage in the same manner and watch how spongy our minds become.

2. Context is king. Whenever I hear grumbles about large language models’ (LLMs) outputs, I wonder – or directly ask – what the prompt was. Without a thorough yet concise prompt and added context input, the output will usually be lackluster. Always remember an age-old truth that applies to more than just this context: garbage in, garbage out. In the training, we covered a research-backed advanced prompting formula** that ensures the user exhausts their brain first (which is ideal for continued development of cognitive skills – more on that in a future post), providing as much context within the prompt as possible, and then adds any other files or data pertinent to the task. This type of methodical, systematic approach, though time-consuming, will save loads of time later.

3. Think alongside the AI tool. Applying the large language model’s generative abilities to our social impact use cases is very much an art. The non-deterministic nature of the LLMs makes it such that no two answers will be the same. In addition, a creeping dependency on the tools to “think” for us – whether it be in making this email “more professional” or less “emotional,” or communicating with donors – is not the recommended approach for myriad reasons. One is that these tools, as magical as they may seem, do not fully understand the layers of nuance that we do in our day-to-day tasks. Can they learn? Sure, but only up to a point. Instead, we need to think alongside the LLMs. Asking the LLM to provide suggestions along with its reasoning will augment our approach – either by taking its suggestions, keeping them in mind, learning from them, or scrapping them altogether. Whichever we decide, we are critically thinking and applying the vast amount of experience we hold.

** Note: In the workshop, we covered the expanded version of Caltech Center for Technology and Management Education’s RADOE prompt formula (Role, Action, Detail, Outcome, Example).