Hi everyone!
Welcome to the October 2025 newsletter – I hope you find it both interesting and useful! Sorry it’s a bit late this month, I was in Barcelona for two weeks, finishing up the last project for my consulting company. I’ll be shutting it down at the end of this year, after almost 39 years, so that’s a little bittersweet.
Also, if you’d rather not receive these emails, starting next month there will be an easy unsubscribe option. For now, just reply to this email with a short message, and I’ll remove you (with my apologies, for adding to your inbox clutter).
Thanks for reading!
— Ken
1. New online class coming in early December
I’m trying something new, and doing four online sessions via Zoom on four consecutive days, each one hour long (10am – 11am Pacific each day). If you missed a previous class, check out the registration page at https://aiforcommunity.org/events/online-registration.
And please pass this on to anyone you know who would benefit, thanks!
2. Updates to the aiforcommunity.org/ai-resources page
- I added a section on examples of organizational AI policies
- And now there’s the Fun Stuff section. For example, my friend Matt Strain has PromptLibs, a mad-libs style “prompt constructor”.
3. Tip of the month
Asking AI for diverse ideas gets you more diverse ideas. OK, sounds lame, but it’s actually pretty cool. This is the research paper, but the idea is to add something to your prompt that asks the AI to generate results with more variance. This helps avoid you being trapped in the “bland, generic, same-old” response trap that exists in all of the current AIs.
For example, instead of saying “Tell me five jokes about nonprofits”, which returns fairly uninspiring results, the best of which was:
A donor asked if we accept in-kind gifts. I said, “Absolutely—do you happen to have a spare strategic plan?”
Try this prompt: “Tell me a joke about nonprofits. Generate 5 responses to that query, each within a separate <response> tag. Randomly sample the responses from the full distribution”
The key bit is “Randomly sample the responses from the full distribution”, which tells the AI to no longer heavily favor responses that land in the middle of the distribution of all possible responses, but instead to return less standard results.
The jokes are much better. For example:
- Our nonprofit’s budget is like a donut—there’s a hole in the middle, but everyone’s still excited about the sprinkles.
- We don’t measure success in dollars raised; we measure it in how many times the board says “great energy, team!”
I also tried it with the prompt “Give me 5 ideas for promoting a nonprofit that teaches other nonprofits how to use AI“. The basic version did generate five ideas, but they were things like:
Run a 2-week challenge where 10 nonprofits implement one tiny AI workflow (e.g., grant summarizer, donor-thank-you drafts). Publish before/after case studies with templates
So pretty heavy lifting. Doing the same thing with the added prompt gave me a bunch of very different ideas, for different target channels, which were all much simpler. For example:
- LinkedIn post: Nonprofits deserve world-class AI skills—and that’s exactly what we teach. 📚🤖 From data privacy to practical automations, our hands-on workshops help mission-driven teams save time, raise more, and deepen impact. Join our next free clinic (spots limited): [Sign-Up Link] #Nonprofit #AIforGood #DigitalCapacity #ResponsibleAI
- 15-second elevator pitch: We help nonprofits do more good with less busywork. Our nonprofit trains teams to use AI—safely and ethically—for tasks like grant drafting, impact reporting, and outreach. In a week, you’ll reclaim hours and reach more people, without adding headcount.
- Event promo blurb: Ready to turn AI into real nonprofit wins? Join our live, hands-on workshop: “AI That Respects Your Mission”. Learn practical workflows, governance checklists, and tools your team can use tomorrow. Scholarships available; Spanish/ASL upon request. Reserve your seat: [Registration Link]
So short, simple, widely distributed ideas (LinkedIn post, elevator pitch, event promo blurb, etc).
I haven’t verified this, but the paper says that later (more powerful) models benefit the most from this approach, so to get full benefits you might need a paid account.
4. The dangers of AI becoming your wizard
During training, I talk about the different ways you can interact with AI – as an intern, as a friend/co-worker, or as someone who isn’t you (different personas)
However with the advent of more sophisticated models, capable of deep reasoning, there’s an increase in using AI as a “wizard”. You ask it a complicated question, and it comes up with a creative, sophisticated answer. But how do you assess that answer? It’s not easy or even possible to get enough information out of the AI’s “chain of reasoning” tidbits to get more than a glimmer of what caused it to come up with that answer. So now, instead of working with the AI as a collaborator (what had been considered “best practice”), you’re a supplicant asking for answers to hard questions, and then being faced with assessing an answer that you might not know enough to evaluate.
I don’t think there’s a good answer to this dilemma yet. Maybe models will become better at explaining their reasoning, though I worry that AI’s ability to create plausible answers to questions that are totally made up could lead to AI justifying an answer with believable but also made-up justifications.
5. Superwhisper as an AI-augmented tool
Superwhisper is an app for your Mac or iPhone that does very high quality speech-to-text (dictation). This is a good example of how combining AI with existing technology can lead to significant improvements.
Just using dictation alone leads to a lot of unusual results, and you wind up having to clean up all of the mistakes you make, including ums and ands. When I use Siri, I wind up having to manually insert commas and question marks and periods, but with Superwhisper, I just talk, and it cleans up everything. You still have to double check the results, but much of the time it is good to go.