How Experts Are Approaching Bias In AI

Forbes

This conversation between artificial intelligence industry experts about bias in AI and how to best govern AI to unlock its potential. This conversation took place at Imagination In Action’s ‘Forging the Future of Business with AI’ Summit in April 2024.   0:00 Introduction 2:24 How Does AI Play Into Bias? 6:40 Can AI Sense Personal Rhetoric Through Computers/Programming? 10:29 Building Bias Detection Tools: What Will That Entail? 16:03 How Does Cultural Diversity Play Into Data 18:24 How Framing On Topics Can Help Build More Receptive Data  Subscribe to FORBES: https://www.youtube.com/user/Forbes?sub_confirmation=1  Fuel your success with Forbes. Gain unlimited access to premium journalism, including breaking news, groundbreaking in-depth reported stories, daily digests and more. Plus, members get a front-row seat at members-only events with leading thinkers and doers, access to premium video that can help you get ahead, an ad-light experience, early access to select products including NFT drops and more:  https://account.forbes.com/membership/?utm_source=youtube&utm_medium=display&utm_campaign=growth_non-sub_paid_subscribe_ytdescript  Stay Connected Forbes newsletters: https://newsletters.editorial.forbes.com Forbes on Facebook: http://fb.com/forbes Forbes Video on Twitter: http://www.twitter.com/forbes Forbes Video on Instagram: http://instagram.com/forbes More From Forbes:  http://forbes.com  Forbes covers the intersection of entrepreneurship, wealth, technology, business and lifestyle with a focus on people and success.

Transcript

00:00 Good to see you all.

00:05 My name is Michael Deela.

00:07 I am the director of a project called Oslo for AI, which is a project to do diplomatic

00:16 work between different AI communities who are not talking well to each other, but who

00:23 need to for the sake of governance and much else.

00:27 I'm excited to get to the conversation with the panel.

00:32 Before I do, I want to ask you some questions.

00:36 So if you could answer them with a show of hands, a clear, nice, big raise of your hand.

00:44 The first question is, do you think that intelligence is well understood?

00:50 Raise your hand if you think yes.

00:53 Okay.

00:55 Do you think we have a well-established consensus on the potential harms of AI systems?

01:03 Raise your hand if yes.

01:05 Okay.

01:06 Three of you.

01:07 Good.

01:08 Do you believe we have existing models of governance capable of regulating emerging

01:14 technology?

01:15 Raise your hand if yes.

01:17 Okay.

01:18 So just based on that very negative show of hands on those questions, you know, I think

01:26 that's a very good benchmark for where we are in the governance conversation.

01:32 If we were going to be doing really careful, thoughtful governance work, we would have

01:38 better answers to all of those questions.

01:40 So I'll just, you know, my first offer to you is those are three questions we should

01:45 all be thinking more about and talking more about.

01:51 Each of the panelists today, and I hope what are going to be non-obvious ways, have something

01:57 different to tell us about how to think about governance.

02:02 So I'm going to go to each of them in turn.

02:06 Dennis will be first, Malur next, and Kim will follow up.

02:11 So first of all, Dennis, can you tell us a little bit about your background and then

02:18 you can tell them about the mistake you made that landed you on this stage.

02:24 Sure.

02:25 My name is Dennis Gleason.

02:26 I am a recovering political analyst in the CIA and analytic methodologist.

02:33 I can't stress the word former enough in that everything that I'm about to say is really

02:39 kind of my opinion.

02:41 I made the horrible mistake of committing thoughts to the internet, which was, there

02:48 was a really interesting article in MIT Technology Review about how AI grokked its way to understanding

02:55 basic arithmetic.

02:56 And one of the conversations Michael and I have had for about a year now is the volume

03:03 and diversity of writing about AI is so huge, it's hard to actually keep abreast of what

03:09 everyone is saying.

03:11 And so my question was, well, let's run the thought experiment.

03:16 If we took everything that's been written about our hopes and our fears and our concerns

03:22 about AI and put it into an AI and said, "Hey, AI, we need you to generate 100 generations

03:33 of AI governance policy."

03:37 What comes out after 100 iterations?

03:41 And so that's how I ended up here.

03:43 All right.

03:46 Malur, you are running technology development at an AI company called Latimer.

03:53 Maybe you can tell the folks a little bit about what Latimer is, the problems you're

03:58 working on, and where you are in the play.

04:02 Sure.

04:03 So Latimer is a large language model that is trained on black history and black culture.

04:09 And the intent is to make sure that you have inclusive voices, inclusive representation

04:15 of the data, and therefore less biased.

04:18 My own background, I started my career in machine learning 34 years ago, building reinforcement

04:24 learning algorithms in C++ and C. And I can't believe that 34 years later, here we are talking

04:31 about governance and how AI should be managed and how bias should be managed.

04:38 Yeah.

04:39 Thank you.

04:40 All right.

04:41 And Kim, I know a lot of your career has been spent doing UX and design research.

04:46 You are on the cusp of finishing a PhD in the intersection of rhetoric and AI.

04:52 And I know you have some thoughts about language and the architecture of language and the effect

05:00 that it's having or could have on governance.

05:02 Yeah.

05:03 So I'm going to step back a little bit and just give a little bit wider, broader stroke

05:06 of my stance.

05:07 So I identify culturally as French Creole.

05:11 I am from one of the Karakum nations, St. Lucia.

05:15 And so I speak a language that they also speak in Haiti.

05:18 And so my interest really in this space is multifaceted.

05:23 Before doing that, of course, not before, but rather after, I also led a very large

05:27 design teams.

05:28 I led the first team at a company in Canada, now defunct, called Blackberry.

05:33 And I raised the team from four Blackberry folks, from four interns to about 56 folks.

05:40 And so that's up in Waterloo.

05:42 And I also led a team at Autodesk, taking the first instance of the cloud version of

05:47 a tool called Maya.

05:49 It's a 3D animation software.

05:51 And so I did that as well.

05:52 So I'm not that kind of designer who does interfaces.

05:55 Let me just leave it at that.

05:56 So my interest really was in 2017, when I was leading the team at Autodesk, one of the

06:02 things I noted was that my design team was shrinking.

06:05 It was getting smaller and smaller.

06:07 And part of the reason was because we're proceduralizing a lot of the interaction with the users.

06:12 So instead of having a user go from one screen to the other, they might go from one screen

06:17 to the other, and AI would do the rest.

06:19 And so that, I would say, concerned me as a people leader.

06:23 So I decided to take a break and head back to finish a degree, because I had family.

06:28 I wanted to go back and really think about design at not that physical level of what

06:32 happens in the UI, but rather what happens at a very intellectual level of the ontologies

06:37 that we think about when we think about designing.

06:40 And so my area of, well, hopefully, expertise is in a space called computational rhetoric.

06:46 And so it's at the confluence of mathematics, computer science, as well as the humanities.

06:51 And so I'm here today with the group for many reasons.

06:54 I think there's a lot of overlap in both of the things that we do.

06:58 But one of the things that Michael and I have been talking about a lot is I'm very much

07:02 inspired by Lev Minovich.

07:03 And 20 years ago, I read his book around the meta-- or rather chapter, "Metaphors of Computing."

07:10 And it struck me then that the language by which we currently discuss AI is so fundamentally

07:16 flawed.

07:17 And so that's what I hope to unpack as we discuss here on the panel.

07:22 So I will pause for now.

07:24 Hopefully, you'll have some other questions for me.

07:26 Actually, I want to talk about something.

07:27 So one of the things that interests me where Milo and I overlap is that we have a pending

07:32 project looking at the impact of AI on CARICOM countries.

07:37 And the reason is because CARICOM countries represent a lot of the-- it provides a groundwork

07:44 for surfacing a lot of the pains that we see around AI bias, transparency.

07:48 We are a creolized part of the world.

07:51 And I have advanced the idea that I think it lends itself to very good laboratory testing

07:56 ground for testing small before we scale.

08:00 So we could weed out things like biases.

08:02 But I will talk about that.

08:03 I have a lighting talk this afternoon.

08:05 So please join that.

08:06 And I'll talk about that more.

08:07 All right.

08:08 Milor, maybe you can tell us a little bit more about how LATAMR is modeling bias and

08:15 engaging with bias.

08:16 Because I think it's non-obvious, just out of the box.

08:21 Yes.

08:22 For the last year or so-- well, even longer than that, I've been focused on bias in general.

08:30 But specifically around large language models, we've been working on that for a year or so.

08:34 And the challenge with bias is that it's manifested in many different ways.

08:40 First of all, the data that exists today that most large language models are trained on

08:45 are, as we know, are trained on a mainstream data that's been scraped from the internet

08:50 and largely represents, I'd say, 70% white male population data.

08:59 It's not representative of the larger global audience.

09:01 And therefore, when a LLM has to respond, it's responding based on that data and based

09:11 on whatever nuances it has, interpretations it has made from pattern recognition, et cetera.

09:16 The second part, of course, is there's human bias.

09:19 The way that your data is labeled, that causes-- and how it's stored and how it's searched,

09:24 that causes an additional level of bias.

09:29 And the third is the algorithmic bias itself.

09:33 Because algorithms are trained-- machine learning algorithms are trained to be successful in

09:40 delivering whatever task it is meant to do.

09:44 That is its only purpose.

09:46 And it's going to do that regardless of what occurs around it.

09:50 So that's why you have companies that are building large language models putting guardrails,

09:55 which are like Band-Aids, to say, hey, you know what, here's what-- don't say these things,

10:00 don't say those things, avoid these topics.

10:02 And you'll see when you're talking to Chad GPT or any of those, you'll get those type

10:05 of responses.

10:06 What we've done in LLM is, A, we've trained the data.

10:10 We've incremented the data using a rag bottle.

10:12 So we've got additional data from 100 years of newspapers, dissertations, books, textbooks,

10:20 et cetera.

10:21 To make sure that the data set is more wider and obviously more diverse.

10:29 And the second thing we're doing is we're building bias detection tools in order to

10:35 be able to-- because bias is so subtle, you can detect-- a lot of the bias detection tools

10:41 are really focused on toxicity and those sorts of things today.

10:44 But very little is based on subtle bias.

10:46 So much of bias detection is a human thing.

10:51 And how you perceive bias depends on your own background, your culture, and whether

10:56 you're offended by something or not.

10:58 So that's a real challenge for us.

11:00 And that's what we're building.

11:02 Any company that wants to use an LLM has a bias score that they can say as a reference

11:09 to say this is either positive or negative.

11:11 And then they can decide how they want to mitigate it.

11:14 Yeah.

11:15 Yeah.

11:16 I think there's one more layer of bias that we tend not to talk about, which is given

11:21 that generative AIs are so intuitive to use, we don't talk about bias in the prompts.

11:29 We just type what's out in our mind.

11:31 And so the question is going through that chain from the data to the algorithm to the

11:38 user, how do biases either amplify themselves or cancel one another out?

11:46 I think that's something that's interesting that I don't see talked about a lot is the

11:51 relationship between biases throughout the discovery process.

11:56 Well, and one of the things we were talking about last night is because of the way that

12:03 generative systems work, when we interact with them and bring our own biases to them,

12:09 the way we get feedback from the system isn't consistent either.

12:15 So we have an unstable model for not only the bias we're bringing into the system, but

12:22 the bias that's coming back out.

12:27 I wanted to talk or have you talk a little bit more, Dennis, about how this poses particular

12:32 kinds of problems in the intelligence world.

12:36 And as AI tools appear to be, of course, changing the landscape of anything, everything, how

12:49 intelligence work might benefit from this kind of technology and why it isn't benefiting

12:59 in some of the ways in which you would hope or want it to?

13:01 Sure.

13:02 I think I'm going to kind of zoom out a little bit and say I think this is true of all analytic

13:08 work.

13:09 That, you know, whether you're talking about an intelligence analyst or whether you're

13:13 talking about a professor doing research or somebody at a think tank, you know, we talk

13:20 about the need for like explainability and transparency and trustworthiness.

13:25 And that starts with how data is handled.

13:29 And so I think in order for generative AI to play a greater role in this high end analytic

13:37 work, regardless of where it's being done, it starts with making the data a lot cleaner,

13:44 kind of enriching it, being very clear about the source of the information so that, you

13:50 know, the user can say with confidence, it gave me this answer based on this information

13:56 and I can push or put against it if I have a problem with the source.

14:02 The problem is, in my experience, right, is if you sit down and you talk to a group of

14:07 analysts, they like universally say if the data was cleaner, we could do this amazing

14:13 stuff.

14:14 And it's like, great, let me go talk to the developers.

14:19 And you go across the hall and you say to the developers, you know, the analysts are

14:23 talking about clean data and it would really help them and the developers are like, clean

14:27 data would be so awesome.

14:29 It'd be like, great, we actually have both halves of this equation agreeing on the problem.

14:36 How about we all fix it?

14:38 And suddenly everybody's counting ceiling tiles.

14:41 Because like data cleanup is thankless and it's unsexy.

14:46 And yet it's really critical as like the base of everything that goes further downstream.

14:53 And so I think until that problem is like addressed or people start working on it, it's

14:59 going to be hard for AI to contribute to the precision work, precision analytic work.

15:07 There is, if I may, I asked Perplexity, I think the CEO is here somewhere today, a question.

15:16 I asked him to come up with a bunch of quotes and I just want to quote it here.

15:20 And it came up with a quote that said the challenge with AI is not to create a Terminator,

15:25 but to create a Terminator that takes orders from John Connor.

15:28 So for those of you who are Terminator fans.

15:31 And it was attributed to Elon Musk, which is a whole issue of bias on its own.

15:36 And when you dig deeper, you realize that Elon Musk never said that.

15:40 But the AI assumed that people like Elon Musk must be saying these things.

15:45 And for me, I have a problem with that quote fundamentally because if leaders are thinking

15:49 that way, that is a problem because we need to be thinking about not creating a Terminator

15:54 at all and really thinking about so that John Connor and his descendants don't have to face

16:01 judgment day.

16:02 Yeah.

16:03 Yeah.

16:04 So, you know, building on something you said earlier, Malur, about how Latimer is including

16:11 new kinds of data and layering that into foundation model data.

16:18 You know, Kim, you have been talking, I know, and thinking about doing work in CARICOM countries

16:25 in particular because, one, it's a well-defined sandbox, as it were.

16:32 But also there's a lot of linguistic diversity, a lot of cultural diversity, which has the

16:37 potential to enrich models.

16:39 Do you want to talk a little bit more about how that might work?

16:41 Yeah, I'll talk about that briefly.

16:43 So it's a project that is actually still at its very infant stages.

16:48 We're just trying to scope, like, what is it we need to do?

16:51 So I come from a country where it's a Creolized country, but it's a French-English Creolization.

16:57 There's also Papiamento, which is maybe Dutch and English, and all of these.

17:02 And so one of the things that we find, and if you follow the works of folks like Timnit

17:05 Gabrou and Joy Balunwini, from here, you will know that the data that we've captured

17:10 is really data coming out of the global north, right?

17:13 And we need to extend that.

17:16 And the Caribbean itself represents, a lot of people don't know that the Caribbean has

17:20 over 300 islands with people living on them.

17:23 Many people will say to me, "Are you from Jamaica?

17:25 Are you from Haiti?"

17:26 There are lots of islands and countries in between with very nuanced, deep histories,

17:31 linguistic histories that we can learn from.

17:33 So I think that's a place to start.

17:35 But I want to take the conversation elsewhere just slightly to talk about how many people

17:41 have read about bias in AI, transparency?

17:45 How many people, just by quick show, want to say, "Right."

17:48 Now every time I start reading these articles, I just stop.

17:52 I'm tired of reading the same article that AI is bad, AI is not transparent, and all

17:57 of that.

17:58 And so I do write a blog where I pour my thoughts in there.

18:01 It's really a primer for defending next semester.

18:05 But one of the things I think is that the crisis that we see in data is not just a data

18:09 crisis.

18:10 It's a humanity crisis, right?

18:13 We keep on talking about things, and I've actually written a blog to that topic.

18:16 But I'd like us to really step back and think about new ontologies of discussing this really

18:22 transformational technology.

18:24 For example, one of the splits that we're seeing in the AI space right now is one where

18:28 we have folks who are saying centralization, and we have folks saying decentralization,

18:33 autonomy, control, specialization, generalization.

18:37 We need to be able to frame these conversations first before we begin to talk about the cleanness

18:42 of data to some extent, because that is foundational to how we begin to talk about that.

18:49 The other thing I also talk about a lot is that we need to move away from that concept

18:53 of data leaks, right?

18:54 When you think about leaks, they can become very messy very soon.

18:58 So what is the next thing we need to talk about?

19:00 Is it data oceans?

19:01 And how do we actually frame that conversation before we start thinking about this need for

19:07 innovation and speed in AI?

19:09 So just to get back to your question, Michael, yes, I do believe that the work that we're

19:16 doing in the CARICOM nations, although I do live in Canada, I will be giving a talk there

19:20 in June, so I can't be here in June.

19:23 And it is within-- it's with the Caribbean Studies Association.

19:27 So the theme this year is AI and Caribbean economies.

19:32 So hopefully that is a conversation we can begin to have here with our peers in North

19:38 America and the global north to say, hey, we're open.

19:41 We want to have you test the AI, because we don't want to be the recipient of AI that

19:45 was not created for us, right?

19:47 So these are the conversations that I think we've been having collectively.

19:52 And I think, yeah, I'm going to pause for a while, because I can go on and on and on.

19:55 I just want to add to that.

19:57 Thank you.

19:58 Thank you.

19:59 I just want to add to that.

20:01 You know, Kim and I were talking about-- we were approached by a group in St. Lucia to

20:08 help with learning, because one of the languages she's speaking, or Patois, is fast declining.

20:18 And how do you make sure that those languages are-- you know, we have historic history documents

20:24 in Sudan and all other countries' languages, dialects, that are disappearing.

20:29 So how do we use AI to prevent that?

20:31 But what I want to talk about is really-- you know, when I talk about AI and bias, a

20:36 lot of people raised hands here.

20:37 And there's a lot of fatigue around AI, and there's a lot of fatigue around bias.

20:41 People usually roll their eyes and say, you know what?

20:43 Yeah, I've heard it all.

20:45 It's not relevant to me, right?

20:48 And maybe it is.

20:49 But a counterpoint is that the bias has huge implications in terms of business, right?

20:55 In terms of revenue impact, because of brand image.

21:00 It has impact on the downstream, in terms of how you govern, whether it's criminal justice,

21:07 health care, health insurance, any of those factors.

21:10 All of those have big implications.

21:11 Yeah.

21:12 So that's a great-- I'm going to start to land the plane.

21:17 You give me a great inroad to-- when I was working in the Defense Department, the leader

21:22 that I worked for had a great saying, which is, bad news doesn't get better with age.

21:29 To Kem's point, the problems that are sitting underneath the technical problems-- and you

21:35 all know that the technical problems are many, and they're interesting.

21:39 But there is yet so much more complexity in the human level of the problems, in language,

21:45 in culture, in the inescapability of bias.

21:51 And I wanted to sort of put in mind for all of you a question that we'd like to leave

21:58 you with, which is-- there's a very lucid question about governance that was a headline

22:05 for an OpenAI blog post last year.

22:09 And the headline was, "How Should AI Systems Behave, and Who Gets to Decide?"

22:15 And the question that I'd like to leave you with is, governance is about the design of

22:22 intelligent systems.

22:24 So who designs those systems, and who gets to decide how they work?

22:32 Thank you.

22:33 Great afternoon.

22:34 Thank you, Michael.

22:35 Thanks, guys.

22:36 Thank you, Miller.

22:37 Thank you.

22:38 Thank you.

22:39 Thank you.

22:40 Thank you.

22:41 Thank you.

22:42 Thank you.

22:43 Thank you.

22:44 Thank you.

22:45 Thank you.

22:46 Thank you.

22:47 Thank you.

22:48 Thank you.

22:49 Thank you.

22:50 Thank you.

22:51 Thank you.

Category

Transcript

Recommended