What’s New at AI for Community
- Our free online meetup is August 20th at 10am Pacific, and the topic is building and using Skills in Claude and ChatGPT. Skills sound technical but are actually pretty straightforward, and they can be incredibly useful (for example, making sure everyone in the org uses the same branding standards for external communication). We’ll work through several examples of building skills (sign up here).
- Two online classes coming in September. AI Essentials is September 10th and 11th, no technical background required. AI In Depth follows on September 24th and 25th, which covers customizing your AI, deep research, agents, and AI policies. Both are two sessions across two days (register here).
- Still waiting to hear back on our Claude Corps application. Keep thinking good thoughts for us.
- As I’d mentioned last month, Emily had to leave, but we found an amazing person to replace her! I’d like to introduce everyone to Olivia Loy. She has great experience in marketing & PR, which means I get to spend more time on course updates & classes…and that makes me happy.
- If you know anyone who works at a community foundation or large nonprofit, please introduce them to us. Partnering with organizations is how we set up training efficiently, and the easiest first step is having them schedule a Zoom call with me.
Hot Off the Press
As always, so much in AI-land has happened over the past month, and a fair amount of it actually matters for the work you do.
Anthropic (Claude)
Opus 5 launched July 24th, and for most nonprofit work it’s the sweet spot in Anthropic’s model lineup. The Fable 5 model handles complex tasks better, but Opus 5 is just as good for many tasks, and at half the price. If you’re on a paid Claude plan, you probably already have access.
Claude in Chrome got updated so that the side panel is now a real Cowork session. Cowork is Claude’s mode for giving it an entire task instead of chatting back and forth — you describe the desired outcome, then Claude goes and does the work. You also have saved history and access to your skills and connectors. It’s most useful for the browser-based work where you can’t just use a published connector, for example pulling numbers from the web site of a system that doesn’t support exporting data.
Claude had a privacy stumble. Late in July, people discovered that Claude conversations shared via link were showing up in Google search results, and some of them contained sensitive material. Anthropic’s position is that sharing a conversation makes that content publicly accessible, which is technically true and also not what most people think Share means. Always be cautious when making a chat shareable, as the information referenced by that link could wind up in Google search results if you (or any recipient of that link) aren’t careful.
Claude adds digital fingerprints. Anthropic started embedding invisible, machine-readable digital watermarks into text generated by newer Claude models to comply with transparency regulations like the EU AI Act. It’s not magic characters, it’s subtle choices about which word to generate when building a response, which lets anyone with Anthropic’s detection key calculate a probability that Claude was involved. The technique of using word probabilities is similar to what Pangram does (see below). There’s no personal information in this watermark, so it can’t be used to trace text back to you. With AI-generated “slop” showing up everywhere, we’ve started seeing more proactive detection being deployed by platforms that rely on user-generated content.
OpenAI (ChatGPT)
ChatGPT Work launched July 23rd, and it’s OpenAI’s answer to the same problem Cowork addresses. It pulls context from your email, calendar, and other connected tools, then produces documents, presentations, and analyses. It asks before taking action and cites its sources.
The free tier got meaningfully better on August 7th. OpenAI made GPT-5.6 Luna the default for Free and Go users and removed limits on text-based chats. Separate limits still apply to files, images, and voice. If you’ve got staff or volunteers who can’t justify a paid subscription, the free version is now much less of a compromise than it was.
OpenAI also launched a small business program bundling tools, webinars, and education (details here). It’s aimed at small businesses rather than nonprofits, but the material is largely applicable and free to access.
Google (Gemini)
Google relaunched Google for Nonprofits, which is worth a few minutes of your time if you’re a current or potential Google Workspace user. Eligible organizations can book up to six 1:1 video sessions with Google experts to train staff on AI tools (details here). There are three new AI courses, a repository of 400 nonprofit AI use cases, and new research drawn from 6,000 nonprofits worldwide on what’s working and what’s blocking adoption. This is the most concrete free support any of the big three has offered our sector.
Gemini Notebook (formerly NotebookLM) now gives you more control over the artifacts (slides, infographics, podcasts, etc) that you ask it to generate from your source materials. For example, if you’ve requested a slide deck, you can choose between detailed & presenter formats, modify the length, and add instructions to help customize the results.
But Google is lagging, with no leading frontier model. I still recommend Gemini for deep research, Nano Banana for image manipulation, and Gemini Notebook as a very useful tool for compiling research across many sources and quickly generating slides and podcasts from the same.
AI Detection
Pangram is an AI detection tool, and it seems really interesting. Unlike most detectors, it doesn’t just give you a yes or no. It rates text on a scale from fully human through lightly or moderately AI-assisted to fully AI-generated, and it’s built to have a very low false positive rate — which is when it says something was written by AI, but in fact it was created by a human. That’s the part that matters, since wrongly accusing someone of using AI is a lot worse than missing a case where they did.
Substack has enabled it for the platform, which is one of the reasons why it’s showing up in the news. I ran this newsletter through it, and it correctly flagged most of it as being human-generated. It did say that a paragraph in the Final Thoughts section at the end had some AI-generated text, which surprised me.
One thing I noticed while testing. If I gave it an individual paragraph with more than about 100 words of text, it was generally better at avoiding false positives, but that increased the false negative rate compared to giving it the entire document.
If you’ve got staff, volunteers, or contractors creating content and you have explicitly asked them to NOT use AI, this can help assess content that doesn’t feel real…but please, please remember that it has issues with false positives (see my example above), so it’s a starting point for a conversation, not proof of anything.
Tip of the Month
Here’s something I say in every class I teach — you are the boss.
Here’s what that looks like in my own work. I ask AI to critique a new article for our resources page. It comes back with nine suggestions, many of which are good. And instead of feeling helped, I’m overwhelmed by it. So I put it aside, and it gets added to the stack of things I still need to do. What actually happens next is I go check my email.
The problem isn’t that the feedback was wrong. It’s that AI can generate results faster, and at a much larger scale, than I can evaluate and act on them.
That’s the moment to remember I’m in control. So instead of going and checking email, I need to tell AI what’s actually going on:
I only have 10 minutes. Give me a list of the most important changes to this article that you think I can actually make in that amount of time.
Or “I’m pretty fried today, so just give me the one change that matters most.” Or “Sort these by how long they’d take, and start with anything I can knock out in under two minutes.”
This also connects to a study I mentioned back in February, which found that at one company the use of AI didn’t reduce work so much as intensify it, through task expansion and more multitasking. This matches my experience, where if I’m not careful, I can easily spend most of my mental energy by lunchtime. Constraining what you ask for is one of the few practical defenses against that.
The only controls you have are the ones you assert yourself, via prompting. Which brings us back to where we started: never forget that you’re the boss.
Final Thoughts
I’m seeing more organizations recognize that they need a public statement about how they use AI.
This is not the same thing as an internal AI policy, and I think the distinction matters. An internal policy tells staff things like recommended tools, and where they’re not allowed to use AI. A public AI statement is for your board, your supporters, your donors, and your community. It answers a different question: not “what are we allowed to do with AI” but “what are we doing with AI, and why.”
Almost every nonprofit I talk to has thought about the first one. Not many have thought about the second. And as backlash around AI grows, the lack of public AI statement is going to get uncomfortable.
Based on our recent “Hard Conversations About AI” meetup, there are real issues here. If a nonprofit says publicly that it doesn’t use AI while internally leaning on it regularly, that’s a real disconnect. And the biggest risk isn’t that someone disagrees with your AI use, it’s that they discover it rather than hear it from you. Nonprofits run on trust, more than almost any other kind of organization. Not being upfront and transparent about a tool you use every day is a real threat to that trust.
The “why” matters as much as the “how.” Your supporters will typically understand the case for using AI to spend less time on administrative work and more time on your mission. What they won’t understand is finding out you weren’t being open with them.
We just published ours, at aiforcommunity.org/ai-statement. It’s short, and it leads with the issues before it gets to the benefits, which I feel is important.
Please take it, adapt it, and make it yours.