Mond(AI)y Coffee

Meeting Notes

Atlanta isn’t San Francisco. It isn’t “Silicon South,” or any other such term. Atlanta is Atlanta.

So when we get together to talk about responsible AI, the tone, content, and goals are a little different. Most people in the room have worked for a big corporation. Here, building with protections and safeguards in place from the jump matters more.

Jonathan Kyle of Ten Sparrows led the conversation with Principles of Responsible AI, picking up a question from Pass the Aux: what happens when an AI agent learns to work like one of your engineers, and then that engineer leaves?

The framing: responsible AI doesn’t mean “no risk.” It means the risk is managed, visible, and bounded. “The vendor says it is safe” isn’t enough. Clear use cases, known limits, human control points, and a way to roll things back are.

Six principles, each with a real example of what goes wrong without it:

• Human impact. Average accuracy can hide who gets hurt. The COMPAS risk scores had similar accuracy across races, but Black defendants who didn’t reoffend were nearly twice as likely to be labeled high risk. Who carries the downside when the model is wrong?
• Transparency. A classifier told huskies from wolves with high accuracy. It turned out it was looking at the snow in the background, not the animal.
• Privacy. Your professional duties don’t transfer to a chatbot. Lawyers were sanctioned in Mata v. Avianca for filing a brief with cases ChatGPT made up.
• Safety. Air Canada was held liable when its support chatbot gave a customer the wrong refund policy. What an AI says on your website can count as a promise.
• Oversight. Accountability can’t be handed to the model. Someone has to be assigned, logged, and able to step in, as misuse of Flock license-plate cameras has shown.
• Provenance. The one agents make urgent: what shaped this output, who owns what the system learned, and can a bad or legally barred memory actually be removed? (Short answer: machine unlearning isn’t magic. Design for rollback before you deploy.)

Three takeaways:

1. Responsibility starts as architecture. Define the purpose, the people affected, the data rights, and the failure modes up front.
2. AI needs evidence, not vibes. What can’t be seen can’t be governed.
3. Where learning persists, control must persist. If an agent remembers, you need a way to trace, own, and revoke what it remembers.

And the engineer-twin test, to try on your own AI workflows:

• If the engineer leaves, what learning can remain?
• If that learning is poisoned, who can prove it and remove it?
• If the agent makes a bad decision, who answers for it?
• If a law, license, or court order requires erasure, can the system comply?

Thanks, Jonathan, for raising your hand and saying you wanted to share your thoughts on the topic. It was a great conversation, and one we’re going to continue. A community of builders who aren’t just building cool stuff, but building it responsibly, is something we believe in and want to lean into.

More to come.