What’s New at AI for Community
- Our next AI Essentials class is July 8th and 9th — two sessions, 9:30am to 11:30am Pacific each day, no technical background required (register here).
- The following week we’re delivering AI In Depth in Berkeley, CA at the David Brower Center on July 14th (register here).
- Our free online meetup is July 22nd at 10am Pacific — and the topic is “Hard Conversations About AI,” which is how to talk about AI at your nonprofit without triggering the “are we replacing people?” reaction. Come with something specific you’re wrestling with (sign up here).
- And we’ve got a lengthy list of in-person classes and presentations in July, including the Bear Yuba Land Trust offsite, two AI Essentials classes for the YMCA, and both AI Essentials and AI In Depth with the Central Valley Community Foundation. If your organization would benefit from a customized in-person presentation or class, reach out and let’s talk.
Quick note: our amazing program coordinator Emily is wrapping up over the next few weeks, and we’re looking for part-time admin help — someone passionate about nonprofits and curious about AI. If you know someone, point them at aiforcommunity.org/admin-position.
Hot Off the Press
Anthropic (Claude)
Two product updates since last month. Fable 5 — the advanced model forced offline by a government export control directive in mid-June — had its restrictions lifted on July 1st and is back. Anthropic also released Claude Sonnet 5 at the end of June, now the default for both Free and Pro accounts. If you’re on Claude and haven’t changed any settings, you’re already using it. It performs close to the more expensive Opus model at a lower price point — for typical nonprofit work like drafting, data analysis, or grant research, you should notice better results.
Anthropic also launched Claude Tag, which connects Claude to your Slack workspace so it can monitor channels, respond to assigned tasks, and flag important updates. I’ve been trying it out, and it’s potentially very powerful, but kind of painful to get configured properly.
And this one is aimed squarely at nonprofits: Anthropic launched Claude Corps, a fellowship program that places early-career AI talent inside mission-driven organizations for a full year — fully funded by Anthropic. Fellows spend twelve months building AI tools and systems for their host organization, at no cost to the nonprofit. If your nonprofit could use a dedicated person to help build out your AI capabilities, it’s worth a look at anthropic.com/claude-corps.
Also, we’re applying to be a host organization ourselves. Think good thoughts for us.
OpenAI (ChatGPT)
ChatGPT’s market share dropped below 50% for the first time. It’s still the most widely used AI assistant by a wide margin, but users are clearly willing to try alternatives — which is healthy, since competition is a big part of what’s driving all three services to improve so quickly.
On the practical side, the scheduled tasks feature got significantly more capable. You can now ask ChatGPT to send a weekly summary on a topic you’re tracking, remind you about recurring projects, or flag updates without you having to remember to ask. A new “Scheduled” page in the sidebar manages everything.
Google (Gemini)
Google released new image and video generation models this month, with a model name only an AI company could love — “Nano Banana 2 Lite“. This version is free for US users in the Gemini app. If you’ve been meaning to try AI-generated graphics for event flyers, social media, or program materials, this is an easy place to start.
The Bigger Picture
A New York Times piece (June 29) flagged an emerging blind spot worth knowing about. Researchers at Boston University and Boston Consulting Group found that managers review AI-produced documents less carefully when they think the work came from an “AI employee”. Specifically, they missed errors they would have caught otherwise, apparently because they didn’t feel personally accountable. About a quarter of the companies surveyed already include AI agents on their org charts. The takeaway: the review step matters more now.
Separately, a large Oxford/UK government/Stanford/LSE study found that AI systems are now reliably better than expert humans at text-based persuasion — even when experts chose their own topics, prepared thoroughly, and were offered significant cash incentives. In one test, AI was nearly three times more effective than professional fundraising canvassers at generating charitable donations. For nonprofits, AI can be a useful tool for donor appeals and advocacy writing — and it’s worth knowing that persuasive AI-generated content arriving in your inbox is getting more effective too.
Tip of the Month
Most people think of tools like Codex (ChatGPT’s agent mode) and Claude Code (in Anthropic’s Desktop app) as developer tools. And that’s how they started. But a new OpenAI research report tells an interesting story about what’s happened over the past year.
The fastest-growing Codex users at OpenAI aren’t engineers — it’s everyone else. Lawyers, finance staff, and recruiting teams now use Codex for more than 85% of their AI work, up from almost nothing in August 2025. Non-developer organizational users grew 189-fold over that period.
What changed is the kind of work these tools handle. They’re designed for longer, multi-step tasks you hand off and let run — the OpenAI report found more than 80% of users had made at least one request representing 30+ minutes of human work, and a quarter had submitted requests representing 8 hours or more. That’s a very different model than typing questions into a chat window.
For nonprofits, the practical application is tasks that normally require babysitting multiple rounds of AI chat: summarizing a set of program reports, pulling together a grant research document, analyzing survey responses across multiple files. You describe what you want, hand it off, come back and review.
But…Claude Code and Codex have a slightly more technical interface. I still think it’s worth downloading the app and trying on a low-stakes task. For example: launch the Claude Desktop app, switch to Code mode, paste in a couple of recent program reports, and ask it to summarize key outcomes in one page organized by program area.
Final Thoughts
I’ve been watching prompting advice age out faster and faster, and I think that’s actually a useful signal.
A year ago, “give AI a persona” was standard advice — something like “you are an expert California labor attorney” before asking your question. The idea was that framing the AI as a domain expert would improve accuracy. There was early evidence it helped. But Wharton’s Generative AI Lab recently tested persona prompting across current frontier models and found it doesn’t reliably improve factual accuracy anymore. Earlier research from the same lab found that threatening AI (“I’ll tip you if you do well”) and chain-of-thought prompting (“think step by step”) have also lost most of their effectiveness with newer models.
I still think the persona framing has value — just not for the reason people thought. When I imagine talking to “a California labor expert,” I naturally change how I approach the conversation. I give it more challenging tasks, and I eagerly share what I already know. The improvement isn’t from the label I gave the AI — it’s from the quality of conversation that framing prompted me to have.
That’s where I think the real shift is happening. Working with AI agents is starting to look more like management than prompting. You’re briefing a capable system, giving it a goal and context, and coming back to evaluate results you didn’t personally produce. The skill that matters is shifting from “which phrase unlocks a better answer” to “clearly defining the goal, giving useful feedback, and knowing when to push back.”
So when the nature of productive work shifts, which skill investment actually matters most? The current prompting techniques are already going stale. The ability to clearly communicate what you need, give meaningful context, evaluate results critically, and redirect when something isn’t working — those are management skills. And unlike prompting tricks, they don’t expire. I think organizations are going to find that they need to invest in teaching everyone (no matter where they are in the hierarchy) how to become better managers.