Questions around Transparency in AI models with Tom Davenport
Often the question around bias is raised whenever the conversation turns to AI. Tom Davenport, author of “Working with AI: Real Stories of Human-Machine Collaboration” points out that bias is not limited to AI, but also finds root in many human decision-makers as well. Actually, according to Tom, the bigger threat is ignoring that working with AI is going to increasingly be a part of our human work experience.
Get even more insights from data and analytics leaders like Tom on https://www.thoughtspot.com/data-chief.
#thedatachief #aipredictions #2023trends #aitrends
Cindi Howson (14:28):
Yeah. So I think it, it, it's the human in the loop, um, or the human plus the AI that makes it the better combination. That's what speeds the diagnoses and makes, removes some of the bias from the radiologists. But maybe if it's full ai, does that, or black box ai, does that worry you about the degree that we rely on this in terms of bias at scale?
Tom Davenport (14:58):
Well, the bias doesn't worry me as much because humans are hugely biased. I think even more so than AI is generally, of course, you know, if AI models are trained on what on human humans have made for decisions in the past, that will probably fund some bias. But, um, or incorporate some bias. But what does worry me, I think if the models are not transparent, that, um, people will just refuse to use them. I mean, going back to radiology, there's a academic study which I really like, of three groups of radiologists in a big hospital in New York. And, um, two out of the three groups, they, they all, all three have used AI to kind of confirm or deny their, their initial judgments. And two of the three groups just kind of blow it off if the AI disagrees with them because totally opaque, you know, the models they can't make any sense of. And so they just say, forget it, they must be wrong. I don't know why. But, um, so we, we need to make progress, I think, on the transparency of these models. So we could have some sort of, I don't know, dialogue if you will, between the, what the machine thinks and what the human thinks. Um, and that, that kind of collaboration is not easy to create. Uh, you know, it requires both sides, um, compromising a bit.
Cindi Howson (16:25):
Yeah. Or I would say, well, transparency breeds trust. I think it's easy to create, if you build that into the design from the beginning, it's hard to go back if you have black box AI and not reveal Yeah, what were the inputs, what was the sample training data, things like this. So I think that's what makes it hard. Yeah.
Tom Davenport (16:48):
But it's really hard. I mean, on the most powerful form of AI that we have now is deep learning, um, neural networks and those are almost exclusively black box. Um, very, very difficult to make sense of them. And to take, um, G P T three for, for creating content hundred and 75 billion parameters. Oh, sure. Yeah. I, as a human being, I'm gonna make sense of what those parameters, what those loadings are and all those parameters and how did that come out with that particular bit of text or that image or whatever, not gonna happen.
Cindi Howson (17:24):
Yeah. Okay. Fair enough. I was wondering if you were gonna bring up, um, G G P T three, because it is just blowing up and the fastest ever to get to a million downloads. And I remember, um, a talk that we were doing, I think for Venture Beat in the last two years, and I asked you, were you concerned now that it started out open source, but now Microsoft being the main backer and owning it, and what did we think, was this a good thing or was this a bad thing? So what do you think or give us a prediction around who's gonna leverage it best or how in 2023?
Tom Davenport (18:05):
Yeah. Well, I mean, I, I think it's sort of a mixed bag on the, on the one hand I liked the fact that it was open source, but one of the big investors was Elon Musk. And frankly, I'd rather have Microsoft in charge than Elon Musk in charge of anything. I, I just told my Tesla because I really can't support that guy anymore. And, um, two, I think there's a lot of competition now to, um, GPT three and when G P T four comes along and so on. And so I'm not as worried about one company dominating this space. Um, okay. I do think that they sort of have the lead right now and chat G P T has been kind of wowing everybody in the past few weeks. But, um, certainly there are competitors and some of them are open source, you know, hugging face and mid journey and so on. Yeah.