AI Safe Sandbox Webinar, Part 2 (September 2025)

Our bread and butter is really about sussing out regular and advanced use cases for AI. Since data safety tends to be a choke point for AI in business, we need strategies for amber data: the stuff between “yes, you may absolutely put it into an AI chat” and “no, you absolutely may not.” For anything in between, we need some sort of masking if it’s for a business, so we can get the most out of the chatbot safely.

I’m sure you’ve all heard concerns about over-reliance on AI. In policies, we want something that reflects when people should step away from AI, not just when they should use it. In terms of trustworthy data, it’s necessary to identify when outputs are biased and could reflect poorly on an organization’s motivations, if they’re client-facing communications. Those are the daily-use things to look out for.

Then we have bigger-picture strategies and scaling AI for these types of organizations. Our advanced training is going to center on access and compliance. We’re researching the new EU AI regulations, and in the USA, the National Institute of Standards and Technology recommendations, to give better compliance advice. But we’re not lawyers. It’s not legal advice, and we’re not going to call it that.

Many nonprofits, and certainly the people working for them, are really concerned about the negative environmental impact of data centers. So we want to demystify those dangers and also create incentives to use them more energy-efficiently. Finally, we want to ensure that the AI policy accurately reflects the goals of mission-driven organizations. That’s really where the “align” in AlignIQ comes in. It’s the center of our mission.

Recent reports are showing that 95% of AI business projects fail. This is from a very recent MIT study that came out in the last week. We don’t want you to be part of the 95%. The difference is starting with policies and clear practices before chasing shiny tools. When I was teaching students how to write college essays at my last job, one of my favorite sayings was: well begun is half done. You just need really good preparation.

In our AI governance quick start, we help organizations develop policies like these to create the strongest possible outcomes with AI, and use our data science background to determine feasibility. People need to know that not everything can be done with AI.

If this isn’t applicable to your situation, we’d ask you to think of someone in the social impact sector whose organization would benefit from AI but doesn’t know where to start, or has started and feels like the 95%: knee-deep in a sunk cost, not knowing why AI isn’t working for them.

With webinars, especially free webinars, we’re trying to give an impression of who we are and what we think about the material, and give general advice. An organization with a communications or compliance department won’t even need to think about some of this. But we want to work with smaller organizations, often with only ten employees, who are all contributing to making the world a better place and aren’t thinking as much about things people in other companies would know.

[Answering a participant question about AI and search.] Fundamentally, Google is in trouble, because if you’re just looking for some information, Google is no longer the best way to do it. These chatbots are connected to the internet, so they run the search and then summarize, the same way Google is now doing it.

“Basically they had to come up with something to keep some traffic on google.com, and that’s why they’re doing it.”

It doesn’t mean ChatGPT is going to give you better results. It means more people are interested in the simplicity of getting a precise answer back, even in some cases if that precise answer is not accurate. There are ways to turn that off in Google.

“AI Overviews.”

I do think it’s better in some cases. If I’m looking for a bunch of different resources, like which news organizations have posted something about the EU AI Act, I’m probably still going to use Google, because ChatGPT is only really good at that in its research mode, or whatever they’re calling it now. That mode is super energy-hungry, it comes back with only three choices, and I’m never going to know whether those are the three best resources.

If you use one of these AI chatbots, whether Google’s Gemini, Copilot, ChatGPT or Claude, which is another really popular one, what they’re really good at when connected to the internet is running a search, extracting the part of the information that’s useful to you, and giving you that.

For example, with the dog we saw earlier: if we’re worried about something with the dog and do a Google search, a lot of times what comes up is a website that’s not that good, with a lot of paragraphs repeating themselves in different ways. Its search engine optimization might be really high, but it’s not that useful, and it takes a long time.

“You have to sift through a lot of stuff.”

Yeah. And to some extent that explains the popularity of these systems for that purpose. A lot of those websites aren’t much use per word.

Setting healthy boundaries is still an open question. This is a tool, but in the same way a cell phone is a tool people don’t always have healthy boundaries with…

“Why are you looking at me like that?”

No, not you. In the same way that gambling isn’t safe for some people, there’s a need to understand whether this could be bad for a particular person. So I’m not going to always say AI is the way to go. It depends on the person, and it also depends on the task. At work in particular, I do think it can be extremely helpful when used as an assistant, with pretty rigorous standards for when you would use it and when you wouldn’t.