AI Safe Sandbox Webinar, Part 1 (September 2025)

We’re going to run through the essentials of experimenting with AI assistants like ChatGPT, Claude or Copilot safely and responsibly. We’ll cover the four big dimensions of AI risk, with practical steps you can start using today. We’re also introducing the scope of what you can do to mitigate risk, for daily use and for bigger-picture AI in organizations. So here’s our little welcome.

Yes, this is about us. AlignIQ is a mission amplifier business, the way I think of it, meaning that we help teams align AI with their purpose. While AI can be misused, in the hands of people who want to make the world better it can give time back. It can help people use today’s best tools and prepare for tomorrow, all without losing trust or their values.

Since humanity’s knowledge of AI is evolving so quickly right now, individual experimentation is essential to get the most out of these opportunities. There aren’t manuals, in a way. You have to play with it and figure out what works. But without guardrails, using these tools can also leak data, mislead us or erode trust. So let’s look at how to avoid those pitfalls.

What we’re going to show is how we’re thinking about the main areas of risk. There are four main categories that matter whether you’re experimenting, using AI daily, or thinking long term. Data safety is the largest category, but there are others raised in frameworks like NIST or the EU AI Act. Those frameworks are mainly written for developers, but we’ve translated them into guidance that works for everyday AI users like us.

When you’re using some of this new accessible software, in addition to data safety there are risks around the reliability of AI outputs, their truth and fairness. That in turn leads to risks in maintaining your voice, your purpose and expertise, and responsible stewardship of our world. Three of these are critical when experimenting with an AI system, so we spotlight how to mitigate them today: identifying safe data, detecting hallucinations, and maintaining stakeholder trust.

As you experiment with an AI system, it quickly becomes obvious that it lacks context. I think of it as an intern on their first day. The way to make it more useful is to give it data, and the more data for context, the better it gets. A lot of times when somebody first starts out, they’re thinking, well, what’s the big deal? When you start giving it more context is when it starts shining. But not all data is safe to give to this AI intern.

One way to think about the data is a traffic light system. Red data would be like client health information or unreleased intellectual property. That’s something you would never want to put into a public AI system. Amber data would be like drafts or anonymized notes. These can be used carefully if the private information is masked. Green data is things that are already public, already on the internet or general knowledge, and that is generally safe.

As a starting point, providing data that’s already online or not a secret is enough to get a lot of very useful applications, ranging from researching topics to summarizing articles to translating languages, including one of my favorites: obscure computer errors into English. It’s actually pretty good at that sometimes.

The reason safe data practice is so important, and why we have it first, is that there are two main sources of risk with these systems and data. First, what the AI companies are storing isn’t necessarily stored at the highest security compliance.

“Or we don’t have any way of knowing, really.”

Yeah. It’s a black box. There are Silicon Valley companies that move fast and break things. They’re not companies you want to trust with very sensitive information. That’s the one reason. The other is that your data is used to train new models by default, which means what you paste in can show up in a future user’s response. That’s why we say always turn off training in public chats. According to their policies, that should eliminate that risk.

Flagging hallucinations. Beyond the data risk area, in the truth area, is when AI output expresses high confidence but its assumptions are flat false. This can be confusing and frustrating, as the image shows. The technical term is hallucination, and if you’ve ever tried autocomplete on your phone, you can see how this happens.

AI assistants today are somewhat like an eager new intern. They are sometimes brilliant, but with no real-world experience. You wouldn’t send their draft to a client without checking it, and the same goes here. If you’re not sure about an assumption in the output, journalism’s three-source rule is something you can follow that would save you from a retraction, essentially.

As people begin experimenting with AI assistants at work, they get outputs that would have taken them significant hard work, which is part of why people are excited about the technology. But until these steps to flag hallucinations are taken, the work reflects the experience level of a book-smart intern.

This risk around whether AI output represents the whole truth leads to risk in maintaining trust when using those outputs. There could be a temptation to ship out impressive-looking AI work. However, your audience doesn’t know to what extent it represents your opinion or work, and trust is fragile. If AI shaped something, disclose it. Your donors, clients and staff need transparency. You always want to sign off as a human and never pass raw AI output along.

Online, people are calling it a bank, because you might have different ways to disclose different levels of assistance, or for different audiences.

So what would these policies look like in practice, and what kinds of AI policies do these organizations need to start developing? Here are some examples. For stakeholder trust, we start with a policy that ensures transparency. At the end of any of our posts online for marketing, in the posted media file or the blog post, we include a line that says we used AI for polish, not purpose.

For heading off hallucinations, we might have specifications on our AI policy for conducting AI research. For example: any factual claim generated by AI must be verified against at least two independent, reputable sources before inclusion in final outputs.

We might also create a resource sheet that can be pinned up at a project manager’s desk, something like the green, amber, red, just to keep something handy so you’re reminding yourself as a project manager what you can or can’t include if you’re doing research using an AI chatbot.