037 – Generating AI Images: Between Bias and Absurdity

[00:00:00] Lily: Hello, and welcome to the IDEMS Responsible AI podcast. I’m Lily Clement, a data scientist, and I’m here today with David Stern. Hello, David.

[00:00:14] David: Hi, Lily. Looking forward to this follow up, because our recent podcast was then just before a very interesting article came out.

[00:00:25] Lily: Yes the recent episode called generating AI images: Do AI implementation suppress diversity? And in this episode we talked about quite a few things. For example, if I ask for an image of a scientist from an AI machine such as chatGPT or Gemini, it generally will give me a white male scientist.

[00:00:43] David: Not from Gemini, that’s the whole point.

[00:00:45] Lily: Well, yes, I guess I was trying to be diverse by saying more than just chat GPT.

[00:00:49] David: There are others who are not like Gemini.

[00:00:52] Lily: But okay, so let’s, let’s go into the article. As you say, between recording and publishing, this new article came out on Google’s AI tool called Gemini, where it’s faced a bit of backlash for generating, I guess we could call it, politically incorrect? I think that that’s the word that they’re using.

[00:01:08] David: They’re using politically incorrect? I mean, the point is that they’re including diversity in places where it shouldn’t be. This is exactly what is to be expected. They’re overcompensating.

[00:01:20] Lily: Yeah.

[00:01:20] David: It’s really hard to get it right.

[00:01:23] Lily: Well, just as an example of what we mean by they’re overcompensating. So someone’s asked it to generate an image of the US founding fathers, and they’ve put in it a black person, which wouldn’t have been the case at the time.

Or when generating German soldiers from World War II, having all sorts of races in there. And I think that the quote from the article was along the lines of, you know, in trying to solve the problem of bias, they’ve then created this whole new problem. Which is that it’s just absurdity.

[00:01:53] David: Yes. And I think this is the key thing, AI is in its infancy still in this sense and trying to fix one problem, you are going to create another problem because actually getting it right is hard. So it’s a very blunt tool which is being used at this point in time. It’s difficult to call it that because it’s a very sophisticated blunt tool. But it is something where actually doing this right is something where that’s layers and layers away from where we could be.

[00:02:34] Lily: I wonder how much input we need from other people on this, because I guess… Maybe I’m completely wrong here, but my understanding or the image in my head of how AI tools work is that there’s a lot of scientists, data scientists, statisticians, working on it, but you don’t have as many people in other areas working on it who might understand these.

[00:03:00] David: Well, there’s now a whole group in philosophy working on the ethics of AI, and they’re some of the groups who Google will have been paying attention to. Actually, a while back, they were at the forefront of this. They may still be but they were at the forefront of this at one point and then they actually got into some difficulties with it. That’s part of the history and they’re clearly making efforts.

So I don’t see this negatively, I don’t see this backlash as being necessarily negative against Google and what they’re doing, on the contrary, I see this as demonstrating that they are ahead. They are at the forefront because they’re making the mistakes that the first people to do something would make.

It’s not that others are doing this better, it’s that others aren’t trying to solve that part of the problem yet and therefore they’re not yet as far along.

[00:03:55] Lily: So as a result of this, Google’s paused their tool and they’ve kind of said like, you know, we’re working on this, we’re trying to fix this. But if you’re seeing this from that positive point of view, then it is great that they’re aware of this and that they’re working on it. What are your thoughts on that they’ve now then paused the tool? But yet we still have tools, other tools going out, which have in them these biases, not these absurdities.

[00:04:19] David: It comes back to, I think, one of our previous podcasts, where regulation is good for the big companies, and potentially good for the consumers as well. Because Google has too much to lose by having the bad press. So they’ve got to pause the tool, because they can’t afford that bad press. But they’re ahead.

So, really, if we wanted things to be developed well, I would prefer that the others paused their tools, and they didn’t pause their tool because their tool is ahead. It’s not perfect. It’s got its own problems, of course, any tool is gonna have serious problems. Anyone who doesn’t recognize that hasn’t actually understood the current state of AI.

So all of these tools are going to have serious problems. They’re going to have biases in, they’re going to have absurdities, they’re going to have things coming out, because that’s the nature of the technology in its current state. So the question to me isn’t: should the others have paused their tools and Google not have?

It’s the sort of, well, could we actually get the development of such tools to be more responsible? And of course the other bit of news which has happened very recently is the EU now has released regulations, the first sort of elements of regulations on this. I’m still not ready to comment on that except that this is a big deal. So maybe the dynamics in the EU will now lead to changes in exciting, interesting ways in the balance of how this works because of that regulation piece.

[00:06:16] Lily: Well, it feels, I guess, regulation is great, and it’s good to say this, and it’s good to say and now we’ve got the regulation, but I still wonder what does that still mean? Like, how can we still teach AI, or get AI to kind of understand… Maybe I don’t understand the regulation well enough to be able to know.

[00:06:38] David: Regulation is not the answer because the whole point is the technology can’t do this. So if the regulation wants to make sure that there’s no biases and no absurdities, then nobody would satisfy the regulation. But that’s not how the regulators, regulation of regulators, it’s not how they’re looking at this.

They are looking at sort of responsibility in interesting ways, which I think is important. I’m not saying that the regulation is necessarily good. It might stifle innovation.

[00:07:06] Lily: Yeah.

[00:07:07] David: I don’t know, I don’t have those answers.

[00:07:09] Lily: And I’m sure we’ll talk about that regulation in another podcast when you are ready to comment or when…

[00:07:15] David: When a little bit more is known about how it’s received rather than just it being out. So, you know, it might be coming soon, it might not. It depends a little bit on the reactions of people to this. But it’s part of this development, which I think the likes of Google are caught up in.

I suppose there’s one other big piece in the news around this as well, which is of course Elon Musk and OpenAI. That’s in the news at the moment a lot as well.

[00:07:45] Lily: What’s being said in the news about this, sorry?

[00:07:46] David: So Elon Musk is suing OpenAI because OpenAI was supposed to be open, supposed to have open source algorithms. And he’s suing them to say, well I was part of the team that sort of did this and that’s what was proposed and that’s what OpenAI was going to do. OpenAI’s algorithms are now being commercialized to the benefit primarily of Microsoft. And what he’s also done is released their AI algorithms as open. It’s a very interesting one.

And so, you know, it’s fast paced out there. We recorded that literally, what, a few weeks ago?

[00:08:30] Lily: Yeah.

[00:08:30] David: These three things are all really happening and bubbling up in these three weeks. And, okay. The one which is really directly relevant to the previous recording is, of course, Google. But, related to their response is the regulation. And related to the regulation is also the whole debate around the actually proper open AI, rather than OpenAI’s AI which is not open.

Those are all different discussions. And I’m not saying that any of these are necessarily getting it right, getting it wrong. You know, why is Elon Musk suing OpenAI? Is it because they’re really not making things open? Or is it because he fears that Microsoft is getting the benefit?

I don’t know. I just know nothing on this. It’s happening in different ways. I’d love to believe that Elon Musk’s pushing and really motivated by these technologies being open. And the fact that he’s released their work as open is big. And I think that’s important. What it’ll actually do, I don’t know. You know, are we going to have, suddenly, because that’s open, a whole set of researchers all over the world able to drive things forward and learn in a way that we couldn’t before? Maybe, but we’ll only know that in six months time, I would guess, because researchers are pretty slow compared to this.

Now, of course, AI researchers have to be fast because otherwise they’re out of date. So as research goes, this is fast, but in terms of industry, in terms of the way the AI world is moving, this is slow. Interesting times.

[00:10:25] Lily: Incredibly. And then while everything’s moving so fast, we have Gemini, Google’s AI tool is paused.

[00:10:31] David: And it’s this element of if they want to wait until they fixed all the problems with it, it’ll never be unpaused. But that’s not what they’re trying to do. So the question is, what are they, what can they realistically do? Pausing it was, in my opinion, not a technical or scientific decision. It was a political decision for Google, is my opinion.

It’s not that there’s something in the algorithms which somebody got wrong, which is going to be fixed in two weeks time. It is that they might be able to build a next thing so that what is wrong is not as noticeable in two weeks time. That’s the reality. The technologies can advance and my guess is you will have people working through the night at the moment to try and get something out which does not have those particular issues and so therefore will be acceptable and not a liability for Google.

[00:11:40] Lily: But there’s always going to be something, as you say, you know, we had biases, now we’ve got absurdity. I wonder what the next thing is going to be. But there’s always this ongoing need for this human oversight and AI can’t test everything, or the developers can’t test everything.

[00:11:55] David: No, that’s the key point. The developers can’t test everything. And so the key thing is, humans in the loop. And humans are in the loop, but are they in the loop in the right way? What is it which is there and how is that coming in? When you do something, they’re not in the loop at that point, but they’ve been in the loop at other points. There’s all sorts of work going on. The human effort behind these big models in different ways is actually immense.

[00:12:22] Lily: Going way back to our first podcast, or it might have been the second one on the post office.

[00:12:26] David: Yeah.

[00:12:27] Lily: Where back in the day, you have the machines reading the postal address and the postman would then give you that letter. Humans were in the loop there to correct, to then say, this isn’t my letter. So now we have the humans in the loop once it’s published.

[00:12:42] David: Well, if you think about Google’s problem, it’s actually a much better problem to have than biases. Because it’s much easier to get humans in the loop to recognize absurdity, or to recognize, you know, incorrect facts, than it is for them to recognize biases.

[00:13:01] Lily: That’s a nice point.

[00:13:03] David: So, instead of sort of pulling it out, what about actually having, you know, have you noticed something strange? If so, please respond. These are the sort of things where actually they could put humans in the loop at scale, just like in the post office, when something goes wrong.

And that could be done in very positive ways. And so I would argue, as I say, if it wasn’t Google, but it was some small startup. You can bet that’s what they’d have done. And people would have applauded them for it. Now, am I sorry that people would have applauded a startup and they didn’t applaud Google?

Well, Big Tech’s kind of dominant at the moment. I quite like an underdog story. That’s what we would be doing. But it’s not working out very well at the moment because all these AI startups are exploding in different ways, getting money thrown at them. And most of them are gonna… it’s gonna be a disaster. It’s a point at which actually I would prefer to trust Google. They actually have the expertise and the skills behind this to do this at scale, to try and get it right, to maybe get it out to get the feedback at scale, to improve it and make it much better. So in this particular case, I don’t think I want the underdog who’s going to have the biases and it’ll take too long to fix because they can’t work at the same scale.

I shouldn’t say that so much because we’re one of these underdogs as well, but we’re not in that race.

[00:14:39] Lily: We’re not, no, fortunately not. I don’t know how late you could be to that race anyway either. Can you just catch up? Anyway, that will be a whole other discussion, I’m sure. If you could just join the race at the point everyone else is, or a few steps behind, anyway.

[00:14:54] David: Well, this is exactly where, of course, it’s so exciting that people like Elon Musk have released their AI algorithms as open, because in theory that means you could just enter the race and there are groups who have done this and raised huge amounts of money very quickly just because they’ve got people with the right skills and expertise and so much value is placed on the skills and expertise. Maybe too much value, maybe not enough, I have no idea.

[00:15:23] Lily: Which skills and expertise are you referring to? Do you mean those physical skills for the AI kind of coding or do you mean more of those like soft skills.

[00:15:29] David: A French startup very recently, I need to remember the name, had basically three experts building these large models, large generative AI models. You know, top people from the industry, from the big firms who with real top skills put their startup out there, the three of them, and based on their skills alone, they had the biggest venture fund or whatever it was, in history.

[00:16:06] Lily: Wow.

[00:16:07] David: It was really, really surprising to me. But, at the same time, not. Because they were able to say we know what we’re doing, we’ve got a new company that can do this, with the right money we can get this to happen tomorrow. And that’s what they were basically saying. And so people bought in and said, okay, another player in the game with the right resources behind it could shake up the market and could actually do something. It’s a world I don’t understand. And if I’m honest, it’s not a word I particularly like.

It’s, it doesn’t seem… No, it’s not that it doesn’t. It is not effective. It is not efficient. And so, yes, some people will make a lot of money out of it. But, will it actually build good tools? I don’t believe so. In that case, I would trust Google, more than I’d trust that startup. Not because I don’t believe they have the skills, not because I don’t believe they have the ability to do things and to do things well, not because actually given the amounts of money they have, they can recruit more of their people who they believe can do it right. So maybe they have the freedom to do things differently. But because it’s startup land, they don’t have the structures in place in the same way to ensure they’re doing no harm.

And I would argue that Google’s act of taking Gemini down would be perceived as an act of trying to do no harm, whether that is or not, whether that’s the motivation or not, whether it’s just marketing and so on. But they are worried about reputation in a way that a startup doesn’t need to be.

So I feel that there are elements there where actually that’s potentially a good thing. I don’t know. It’s all so complex. Same time, I really like the idea of just a group of three guys who know what they’re doing, who actually understand the technical side, just having the billions of dollars needed to do something. Wow, that’s fun. That’s going to shake things up a bit, certainly. I don’t envy them because although they’re probably going to become very rich out of it, it’s almost certainly going to be full of moments of failure, you know, the most likely outcome is that at some point their dreams of what could have been remain dreams because they didn’t get everything right.

I hope I’m wrong. I hope they’re able to shake things up and build the next big thing. And they’re not the only one. There’ve been so many of these recently. So I’ve just taken that one as an example. But yeah, the startup culture combined with AI right now is, I think… It’s not the first time it’s happened. And the other times it gradually underperformed and was then left.

[00:19:33] Lily: When were the other times, sorry? Were they like in AI history?

[00:19:39] David: Yes, in AI history, if you go back in AI history, I think there have been three or four booms before now.

[00:19:45] Lily: Yeah.

[00:19:47] David: Trying to remember. I mean, one of them, of course, was around when AI finally won at chess, that led to a big AI boom, in a sense, yes, it’s finally there, it’s going to be able to do all sorts of things, which it couldn’t do at that point in time, and now it can. And so, I’m not saying these booms are totally wrong, there is substance behind this. AI is extremely powerful as we keep saying. What it can be done if used responsibly is huge.

With the current incentives, I can’t see how responsible AI stands a chance. It’s not about it being responsible. No, I shouldn’t be pessimistic about that. It needs people to care about doing things responsibly and not just rushing off.

We’ve been doing bits of work with organisations to try and move towards responsible AI, to try and encourage them to understand what that could mean, how they could do it, and done bits of work training lecturers on this. And I would argue that Google, in its latest attempt, was trying to do responsible AI. And therefore, it is not surprising that it’s the responsible AI one that got turned off.

[00:21:04] Lily: Well, I was thinking to say what, what then is responsible AI? Because there’s bias, and there’s this absurdity, I guess. Do you want bias or do you want absurdity? And as you said earlier on, it’s easier to catch those absurdities than those biases, which is something I didn’t think about before this discussion, but it’s absolutely something I agree with. And now I’m like, it’s also fun to catch absurdities as well.

Whereas biases, maybe that’s a little bit more frustrating. I don’t know, but much, much easier to define an absurdity to them than a bias. But then what is responsible AI? Is it more responsible to be biased or absurd? Well, if it’s more responsible to be absurd than biased, then it’s irresponsible to pause your model and just allow the machines out there to be the ones that are biased.

[00:22:00] David: No, I think, let’s be clear here. Changing history is irresponsible.

[00:22:05] Lily: Okay. Yes. Okay, that’s true.

[00:22:07] David: So I would argue that having algorithms which are imposing current standards to change history is not responsible. And therefore there are elements there where actually it is responsible to stop it. This is not just absurd, you know, you can’t trust it. Now, having said that, you can’t trust other AI either. So it’s not that, you know, people who put too much trust in AI at this point in time, they’re getting it wrong, they’re not being responsible.

I love the way you use AI in so many of your different pieces of work. But I would argue you use it responsibly because you’re always in the loop, very seriously, looking for biases, looking for absurdities, looking for, you know, it is not used to build a final product. It is used to be able to represent, improve, enhance your work.

[00:23:16] Lily: Efficiency as well.

[00:23:18] David: Definitely, yes, as well.

There was another news story around this which I think is fun.

[00:23:26] Lily: Okay.

[00:23:26] David: A, I think it was, an interesting entrepreneur who has used AI rather extensively in their work got kind of caught up because after this amazing advertising of Willy Wonka Land or whatever it was, did you hear about Willy Wonka Land?

[00:23:45] Lily: Yeah. Yeah. So Willy Wonka Land, was it in Glasgow?

[00:23:48] David: I think it was in Glasgow, yes.

[00:23:50] Lily: Yeah, they kind of sold it as a great idea, I didn’t hear so much about it but it sounded like that they sold it as this like huge amazing day because there’s this new Willy Wonka film coming out, and the reality was that it was a little bit…

[00:24:03] David: A little bit disappointing to say the least. Children crying.

But the point was the problem wasn’t really the event. The event was what it was. The problem was the advertising was AI generated and it looked amazing. And so people’s expectations for the event were totally out of whack with the reality.

[00:24:27] Lily: Ah, okay.

[00:24:28] David: This is the real point and this is where you look at what they had, and it was just, you know, that was gonna be fantastic. And then it was in a warehouse with a couple of balloons and what have you. There wasn’t even much of a backdrop, it’s the image I saw, but maybe there was a backdrop somewhere else.

[00:24:48] Lily: A small backdrop?

[00:24:48] David: I have no idea. I just read the news articles. The advertising was done with AI. Apparently, the actors were given scripts which were AI generated and made no sense! So the whole thing was just this AI generated thing, which was a total failure in all sorts of interesting ways.

[00:25:11] Lily: Yeah, I love it.

[00:25:12] David: Yeah, exactly. So, you know, the humans in the loop did try to innovate, there were some actors there who were sort of employed to actually deliver these scripts, which were absolutely insane. Anyway, I do feel a little sorry for everyone caught up, I apologise if I’m laughing at the situation, which was comical from the outside, but it really was not from within.

But it’s again, it’s the role of AI in this and it’s what AI can do. It’s power to make something which, before AI, you’d have needed real skill and you’d have needed to develop… The amount of money that would have gone in to actually developing those sorts of advertising materials and the rest of it. You wouldn’t put that sort of money in unless you also had the money to put in to making sure that the event was good.

[00:26:11] Lily: Sure. Yeah.

[00:26:13] David: There’s a sort of balance there. Whereas now, you know, anyone can make amazing advertising. Or something which looks really impressive because AI is pretty good at that. And therefore, what’s behind it is not as clear. And it often comes back to the humans in the loop, where they do that. Unless I’m mistaken, actually the person behind it was also an incredibly productive author who had within a summer managed to sort of write six or seven full books.

[00:26:48] Lily: Wow.

[00:26:49] David: Yeah. Of course there were signs that they may have been AI generated.

[00:26:53] Lily: I was going to ask, what is it books written since, since 2020?

[00:26:56] David: I don’t know, I’ve not followed up on this, I don’t know, this is just a news article to me. It’s one which relates to what we’ve been discussing in interesting ways, and it’s the other end, if you want, from Google’s absurdity.

[00:27:10] Lily: Yes, yeah.

[00:27:11] David: Because Google is one of the biggest companies in the world, with a team of real experts trying to de-bias their AI, that has created absurdity. And at the other end, you have an entrepreneur who’s using AI to absurdity, almost certainly making quite a lot of money and then losing quite a lot of money in quick succession in different ways.

[00:27:36] Lily: But still both absurd. Well, still this absurdity, it’s still not responsible because how it’s been used has resulted in a lot of disappointment.

[00:27:47] David: Yes, this is the thing. How can we possibly get, I’ve got to stop being less negative about it. The current systems, we are so far from anything responsible, anything which could be conceived as responsible AI. It’s amazing to me. And what’s really amazing is that the power on this, it’s really not clear to me. It’s not in Google’s interest as big tech. It’s not in their interest for AI to be irresponsible.

You know, it’s going to come back and bite them. That’s why they closed it down. If something serious happens, this is going to really hurt them. So it’s not in their interest. So big tech’s interest, it’s not in their interest for responsible AI to not be present in the system. It’s not in government’s interest, it’s not an individual’s interest.

So it’s not really, maybe it’s in some startups interests for us to have responsible AI. But how is it that something where so many of the key actors, it is clearly an issue, it is clearly going to create issues, how is it that this is not something which is… Or this is something which is so poorly understood because when you hear people talking about it, they’re just distracted and they’re discussing the wrong things.

I think we discussed before, you know, this big event at Bletchley Park got distracted by killer robots.

[00:29:27] Lily: Yeah. The summit.

[00:29:29] David: Yes.

[00:29:29] Lily: I don’t remember what episode that was, but that was an event back in November, early November.

[00:29:34] David: Absolutely. Where it was trying to look at AI safety and this sort of things. It’s just, well, if you don’t tackle the real issues of now, yes, have a few people looking at killer robots because that is a safety concern. And you know, there is now a drone, which is just another thing in the news, which has happened just recently. There is now a vehicle, a drone with a big mounted gun coming out of Russia. So there are killer robots, I don’t think it would yet have levels of AI to the extent where it would be considered a killer robot, but it is certainly in the Ukraine war, a real danger of an attack drone in a way which I think is towards that dystopian future that people are worried about.

So I’m not saying these things shouldn’t be of concern. I’m saying that there are elements which are much more dangerous, and actually solvable at this point in time that are not being addressed. These are the issues around, as Google sort of highlighted, the fact that we do not have, even the top experts, do not know how to build responsible systems, even when they’re really trying.

That’s a knowledge, that’s a research point. Why are we not really pumping money into that research. Why don’t we have all our top researchers, a lot of them, getting research grants to investigate that. Because it’s hard. And that is something which should, could and should be made, done through open models, actually looking at this in serious ways.

I think there’s real possibility. Why are we not taking seriously the ways in which biases are coming out in pretty much any system. So, Google is, which is where they’ve tried to counteract that. But that balance of this, their research on that, how much of that is internal versus how much of that is public?

How much could others be building from what they’ve done? Or is that their sort of in company knowledge that they’re building? They’ve obviously made progress, but how open are they with that progress? Because of course, as soon as they’re open with it, all their competitors will use it, and they’ve done all the work.

So how are we getting the systems to actually work so that we are encouraging the right progress? I don’t think we are.

[00:32:20] Lily: No. Very interesting. Do you have any final thoughts to round off this podcast?

[00:32:27] David: I want to finish on a bright note.

[00:32:29] Lily: I know you do.

[00:32:31] David: Because despite all this, I am so delighted about the news article that Google has tried to address biases and got it wrong. Because it was obvious that the first person to try and address biases seriously was going to get it wrong. It’s part of the journey. So in terms of learning, this was inevitable. This had to happen and it’s happened. So what’s next? Look forward to talking again in a few weeks when things have moved on again.

[00:33:08] Lily: Absolutely. Thank you very much, David.