Ep 201: When Trust Becomes the Interface | Heidi Chow

And that's kind of one of the key things that I keep in mind when thinking about trust with AI, transparency.

It's easy to interact with any model and assume that that trust has been sort of pre-established and that what, because it's so confident in what it's sending back to you.

But at the end of the day, it's a promise that it can't fully keep.

You know, it's kind of a trust me even when I'm wrong.

And to your point, I mean, users rarely see a confidence score.

They can't see a training set.

All they see is what's in front of them.

Okay.

that gap between what the AI is capable of and what the user believes it's capable of.

I mean, that's not just an engineering problem, right? That's got to be a design problem too.

I would say so, but I would be, I would recommend really thinking about if your user needs to know all of that information, right? Like we know when we're training in an ML or an AI that it does give back a confidence score.

And so while I don't have a dictionary definition, right? Let's Google it real quick, just in case.

Confidence score, because I don't want to mislead people.

I have my impression of what a confidence score is, but we are here to not spread lies, right? This is

actually great.

This is, I think, the first time we've done real research live on an episode.

This is great.

Okay.

Google has spoken and it says a confidence score is a number showing how sure an AI or computer model is about its own answer prediction.

It usually looks like a decimal between zero to one or percentage from zero to 100%.

And so this typically is more of an internal thing.

I've worked with it when I was designing software.

the AI aligner fit scan while I was on bite, which we can talk about a little bit more.

Let's talk about

that.

And that's kind of one of the key things that I keep in mind when thinking about trust with AI, transparency.

It's easy to interact with any model and assume that that trust has been sort of pre-established and that what, because it's so confident in what it's sending back to you.

But at the end of the day, it's a promise that it can't fully keep.

You know, it's kind of a trust me even when I'm wrong.

And to your point, I mean, users rarely see a confidence score.

They can't see a training set.

All they see is what's in front of them.

Okay.

that gap between what the AI is capable of and what the user believes it's capable of.

I mean, that's not just an engineering problem, right? That's got to be a design problem too.

I would say so, but I would be, I would recommend really thinking about if your user needs to know all of that information, right? Like we know when we're training in an ML or an AI that it does give back a confidence score.

And so while I don't have a dictionary definition, right? Let's Google it real quick, just in case.

Confidence score, because I don't want to mislead people.

I have my impression of what a confidence score is, but we are here to not spread lies, right? This is

actually great.

This is, I think, the first time we've done real research live on an episode.

This is great.

Okay.

Google has spoken and it says a confidence score is a number showing how sure an AI or computer model is about its own answer prediction.

It usually looks like a decimal between zero to one or percentage from zero to 100%.

And so this typically is more of an internal thing.

I've worked with it when I was designing software.

the AI aligner fit scan while I was on bite, which we can talk about a little bit more.

Let's talk about

that.