GivingRadar

AI for Nonprofits

How we use AI, and what it is good for

Giving Radar collects more than any person could read. AI is how we read it.

This page says what we use it for, what we keep it away from, and what we have learned that a nonprofit can use.

What AI does here

  • It labels every email. What the email is for, the one thing it wants you to do, which persuasion devices and feelings it uses, and who it says it comes from.
  • It reads the sector’s news. Topics, tone, how each story is framed, and what it asks the reader to do.
  • It reads video transcripts. Five questions for every video: who speaks, what about, why, how it is told, and what it asks you to do.
  • It listens to podcasts. What the people in each episode think, predict, recommend and learned, each with a quote from the transcript as proof.
  • It writes the first draft of Giving Radar Daily. A person reads it, corrects it and approves it before it is sent.

What AI does not do here

  • It does not count. Anything with a right answer, like word counts, links, dates and duplicates, is done by plain code, which is exact and never changes its mind.
  • It does not guess. A video or episode without a transcript is not labeled. A title is not enough.
  • It does not take orders from what it reads. If an email or a page contains text addressed to an AI, it is recorded, not obeyed.
  • It does not publish. It drafts and labels. People decide what it means and what goes out.

How we keep it honest

  • Fixed answers only. Every question has a closed list of allowed answers. Anything outside the list is dropped, so labels cannot be invented and results can be counted.
  • Tested against people. We labeled 200 emails by hand and counted how often each model agreed with us. Our bar is 85% on the main questions, and we test again whenever the instructions or the model change.
  • Quotes are checked by code. Every podcast statement must come with words copied from the transcript. Code searches for them. When they are not there, we know the AI rewrote or invented them.
  • More than one model. We choose a model for each job, compare them on the same material, and log every call.
  • We say where it has not been checked. The news and video labels have not had the hand-labeled test yet. Until they do, we treat them as a rough map.

If you run a nonprofit

We do not sell AI and we do not consult. But putting it to work on fundraising material has taught us three things we would tell any fundraising team starting out.

  • Ask closed questions. “Is this email mainly fundraising, advocacy or engagement?” gets an answer you can count. “What do you think of this email?” gets an essay.
  • Check it on a sample you have judged yourself. Fifty of your own emails, labeled by you and then by the AI, tell you more than any vendor’s demo.
  • Keep a person between the AI and your donors. Let it read, sort and draft. Let a person decide what is sent.

Read more

The Method page has the full machinery. Each source has its own post on the blog: