Paul Hopkins with Ethical AI by AppSite · Reflections · Part 3 of 3 · 3 min read
Series: Part 1 · Part 2 · Part 3 · The series
AI will believe whatever you give it to believe. So will you. Part three: why AI tries so hard to please you, and how a truthful scorecard keeps it honest.
In Part 1, I wrote about the grape juice stain we stopped seeing. In Part 2, about locus of control, and keeping control on the inside. This last part is about beliefs: what we choose to believe, and what we let our tools believe for us.
Lou Tice taught us that what we see is shaped by what we believe. The same is true for AI.
A word of caution about AI
These days, as the models grow, they tend to be more interested in pleasing you with a final answer, even when that answer is wrong, by suggesting it is exactly what you asked for. That confidence brings people in, but it doesn't get them very far. It can also mislead you if you don't yet know how to use AI or how to tell a good answer from a bad one, especially when you are asking about something you don't know.
At AppSite, we've taken those people-pleasing moments and turned them into actual work: results that stay tied to what you originally asked for. If you're building an app that gets someone from point A to point B, you want the route you asked for, not a confident answer that sounds right and sends them somewhere else.
A truthful 10
Here's how. The turning point came when I decided to put every request into a milestone in GitHub, with the issues attached to it. The milestone becomes the session. The logs, the conversations and the work all get packaged into it, under version control and with timestamps, so nothing gets lost and the original request stays at the centre.
My own stories feed the system, too. Every time a memory or an experience comes back to me, I bring it into the beliefs of the Ethical AI architecture. That is how it learns what a good result looks like, and how to score its own work. Scoring matters in AI, but only if what it scores against is in alignment with what you actually want.
We are also building a scorecard into each milestone. A milestone isn't met until it earns a truthful 10 out of 10. Why would we accept anything less?
I saw why this matters while writing this very series. I asked an AI to score it. The information it gave me was correct, but the moment it had a number to beat, it started looking for ways to raise the score. A clear, honest target is what turns a pleasing answer into real work.
If you are starting out with AI, my best practice is to begin with something you already know well. Then you can see how it interprets your request and how it works with you. Get very comfortable with that before you trust AI to tell you the truth about any topic where you wouldn't be able to say, "Hey, that's not right."
Pick carefully
It all comes back to core beliefs, and to keeping control on the inside. If you give AI ten things to believe in, it will believe them. And so will you.
So pick carefully.
← Part 2: Locus of Control · The whole series
A note on where we are: AppSite is currently in beta. We have active users, and we are inviting more people onto the platform as we develop. You can view the status of any of our modules, and our compliance work is in progress.
