AI for Humanity: A Conversation with Allison Cohen

Automatic transcript of the recording (English); speaker labels are automatic. Timestamps open YouTube at that moment.

0:00Sacha: Welcome back. I'm pleased to have Allison Cohen with me today. Allison is a senior applied AI project manager at MILA's AI for Humanity team, where she bridges the gap between AI research and social impact. She's delivered notable projects, including an AI policy search engine for OECD and award winning research on misogyny detection. As host of the world we are building podcast, she explores nuanced perspective on AI development, Allison has been recognized as one of Inspired Minds top 50 influential women in AI, and is a frequent speaker at global forums, including UNESCO focusing on social beneficial AI practices. I'm so pleased to have you with me today, Allison. Welcome.

0:48Guest: Thank you so much for having me, Sasha. It's really great to be here.

0:51Sacha: I want to start out with like a very easy question. Can you explain what AI for Humanity means?

1:00Guest: I can definitely AI for Humanity, at least at MILA is sort of defined as how can we make sure that the AI ecosystem, whether that is from the vantage point of policy, from the vantage point of AI development within academia, from the vantage point of practitioners within industry, how can we make sure the AI landscape is humane, that acknowledge and respect human rights, that are considerate of AI ethics, and are being done in ways that are also scientifically sound. And within that team, of course, there are many levers that we use to promote responsible AI across that sort of value chain, if you will. But within my team specifically, I'm working on developing high beneficial AI tools that we see in some way as servicing humanity, which is obviously a very lofty goal and in no ways do we see ourselves as saving humanity as a whole, but making some small contribution that demonstrates AI's potential to help society in important ways.

2:26Sacha: Personally, to work on AI for Humanity at Mila.

2:30Guest: So beyond just being someone that's very interested in how this technology can make a real difference in our society, which I'd like to think I am, I also come with a background that is in no way technical. I did an undergraduate degree in international development and have a master's in global affairs and so my interest in artificial intelligence initially came from that background presentation of there's so many problems in the world, they're so multidisciplinary, multifaceted. How can technology be used as a force for good in that context? And so from my perspective, I've always wanted to work at that intersection of yes, interesting scientific breakthroughs, cutting edge technology, cutting edge technology, cutting edge technology, cutting edge technology, going to change the world. But how can we do it in ways that meaningfully intervene in a global ecosystem that is so complex?

3:35And how do we navigate it with the right people at the table so that the tools that are put in place in place, have the type of impact that we really want to see because so much of my training coming from that social science background was around interventions that were incredibly harmful. And so I saw its potential, but I also saw AI's potential for harm. And being able to bridge that gap as I do with Mila in the role of AI for humanity has been something that is really fulfilling and aligned with that background that I have, again, in the social sciences.

4:14Sacha: Mila is quite like a big player in AI in Montreal. What inspired Mila to place such a strong emphasis on AI for humanity in its mission?

4:29Guest: Great question. I think, to some extent, there's something somewhat Canadian about wanting the technology that we're building to be incredibly human rights respecting. And Mila is in some ways emblematic of Canada's strategy when it comes to artificial intelligence being one of the three major academic poles where the Canadian government is infusing financial support to see Canada's AI ecosystem grow. So to some extent, I want to say it's sort of a set of Canadian values that have been imbued into the AI ecosystem of Mila. And then to some extent, I also want to say it comes from our leadership team. There is a real origin story around Yoshua Bengio and a number of other professors at the University of Montreal who wanted to create an AI research lab that was highly socially oriented and socially motivated.

5:31They recognize the power of this technology and wanted to make sure that what we built was built for good. And I think there's also, you know, it really connects back to the academic orientation of being interested in doing cutting edge research, but to some extent, not just for the sake of it, but for the sake of supporting a broader goal of creating meaningful research that's both used in industry and used in academia. But ultimately, that underpins something that is broader in terms of its social utility, I guess.

6:08Sacha: Mm hmm. And we've seen the excitement around the I grow in the past probably two years, like crazy. How how do you balance the excitement of AI with the responsibility of ensuring that it benefits everyone?

6:25Guest: Yes. So it is very hard to balance these two things, because in part, there's a lot of interest in moving fast and breaking things, as we know, is the motto that has sort of come out of Silicon Valley. And given how how much of an impact is likely to have and given how scalable that impact really is, there's there's almost this desire to be the first mover to have secured that kind of impact as quickly as possible. And what we know from working with people who are, you know, marginalized or deeply affected by the social issues that we're looking to engage with, there is a real distrust around artificial intelligence, in part because many of these tools are either deployed against marginalized communities for the purpose of surveillance or otherwise. Or because they're deployed too quickly and end up causing harm in unanticipated ways.

7:28And what I've realized in doing this work is that if we are going to be successful in deploying AI tools that have the type of impact that we're looking to see, we need to kind of move slowly because we need to build trusting relationships with the community groups and individuals who will be adopting this technology or using it for their own purposes, their own purposes, their own communities. So I think it's, it's really important that we set a more reasonable pace for creating and shipping these solutions, so that we account for the real human element of, of their deployment, not to mention building some of this technology responsibly requires a lot of multidisciplinary engagement. So working with social scientists, working with community members, of course, working with the machine learning team, and doing that successfully over the entire lifecycle of an AI project takes time.

8:30And I think we need to start respecting the amount of time that it takes so that we can be more intentional, more responsible and ethical once we are shipping these tools. So

8:44Sacha: what you know when you say, um, scientific and technological progress should benefit everyone. Can you, can you elaborate how do you define and measure that benefit?

8:58Guest: Mm hmm. Mm hmm. So because we're situated within an academic research institute, and to some extent, our work is in, you know, academic publications and data sets and improved models, we can measure some of the success and impact in those terms. Um, you know, how many citations have we received, to what extent do our models improve relative to the baseline. Um, and I think that that's, again, it, it acknowledges the fact that our work is situated in an academic research institute, but it may not account for the type of impact we're really looking to have, which goes beyond academia. And I think when we are measuring impact, we need to make sure that that impact is measured according to the people that are using this technology. The people that have built this technology alongside of us.

9:58Um, what do they think is a high impact, uh, metric that aligns well with what it is we've built? Um, because I think there's this real drive, particularly in Silicon Valley to measure the success of a lot of tools in terms of their uptake, in terms of click through rate. And how you measure success will end up, of course, having a big impact on how you build these tools. If you're measuring click rate, you may optimize for, um, you know, click bait titles or manipulative tactics. And we don't want to do that. We want to make sure that we are incentivized in the design of our tools to create solutions that will actually be useful. And so we need to make sure to measure ourselves, um, in alignment with that. So really make sure that our metrics of success are grounded in community outcomes.

11:03So

11:03Sacha: you mentioned that AI risk should be considered and addressed during every phase of development. What does this look like in practice?

11:13Guest: So I think in practice, it means making sure that to some extent you have a reflection embedded throughout the AI lifecycle about whether the decisions that we're making are highly ethical or may have significant risks that we should be looking into before moving ahead. And in part, the only way to do that is to make sure you have an organizational structure or set of incentives that allows you to pause and allows you to, for example, work with someone who is specialized in AI ethics, um, or does represent community members and has those interests at the table. certain decisions about, for example, how to find the right data set or how to label the data set in appropriate ways. I really sort of can't stress enough that having time for reflection, but also making sure that people are there that would be able to identify risky behavior, risky decisions, risky outputs, and sort of slow the process down accordingly feels like it is an important step to take.

12:37And there was an article that came out recently or a publication with Raman Chowdhury that outlined the major organizational elements that companies that produce responsible AI products often have. So it's really about recognizing the role that organizations have as a whole in ensuring that on a project by project basis, people are taking the necessary steps to identify and action on irresponsible practices in the AI development process. It's a big effort, but I think it's worth it.

13:22Sacha: And at least that's from my experience, most times when you increase the understanding of like a technology, it usually comes with a lot of benefits. Could you tell me about some of the initiatives to increase AI literacy?

13:39Guest: So there are numerous initiatives at Mila designed to increase AI literacy, which is awesome. We have the trail program, which is an AI literacy program or an AI ethics literacy program, I would say for researchers and for practitioners. For researchers, the goal of the trail program is to teach them about how to infuse ethical concepts into their research. I think there's now the requirement in many publication venues to describe an ethical impact statement. So essentially describe some of the ethical considerations or implications of your research. And this is an area that many computer scientists don't necessarily get a rigorous education in. So I would say AI literacy, AI ethics literacy is something that we're working a lot on. We have an AI literacy program for professionals as well.

14:45We have what's called the AI for Good Lab, which is a research program designed for, I want to say, bachelors and masters female plus students who are looking to get sort of hands-on experience with applied AI projects. We have the Indigenous Pathfinders program, which is designed for Indigenous students to get more exposure into the field of AI and help sort of kickstart their career in that field. We also have a summer school on AI and human rights. So there are a number of initiatives that are happening within MILA, all designed for the purpose of training, which and literacy, as you're sort of describing, which is so important because the field is still so new. And we need to make sure, especially the communities and the professionals who are going to need access to these types of fields have the training that they need to be engaged in this process and to be making sound policy or to be making ethical research.

15:53Or, for example, for Indigenous community members to be as involved as possible. So, yeah, I would say it's important to check out a number of those programs that are being run at MILA.

16:05Sacha: Topic, what unique challenges do historically marginalized groups face in accessing AI educations or even careers?

16:16Guest: So there's a number of barriers. I think I was having a conversation recently with one of my colleagues, Lindsay Chartrand, who's leading the Indigenous and AI program at MILA. And she talked about how it's never as impactful to talk about or create training programs, for example, in terms of the entire group of marginalized people and their access to, for example, the field of AI. Because the challenges that each community faces does end up being, to some extent, different. But you can sort of break it down across the entire chain of what it takes to get access to the field of artificial intelligence. So of course, to some extent, it's the type of education you have access to in primary, middle and high school, whether you are socialized to see yourself in a STEM career, whether your schools have the resources to include, whether it's computer science training or science training more generally.

17:30I also heard how, you know, in certain indigenous schools, there was no science teacher, for example. And so if you're not getting that kind of exposure early on or if you're not socialized to believe you can have a career in this field, there's going to be a number of people that are sort of left behind at that stage. And then, of course, it's a question of access to university programs and the amount of money that that requires. And, you know, the rigor of some of these programs, the competitiveness of some of these programs and the question of whether you need to balance that program with an external job to pay for that education. And of course, that that makes it harder. And then again, you have questions of who has access to networks of people that are hiring for careers in these spaces and who is promoted when they are given access to certain companies and who's retained over time.

18:34So there are numerous steps to getting people into the right spaces to be able to have a career in AI and intersectional marginalized identities and demographics makes that process incredibly hard because we know the field of AI is overwhelmingly male. male and very oftentimes the community is quite privileged from an economic standpoint as well. And and it's again for the reasons for the reasons I've mentioned, you need to be have gone through pretty rigorous schooling, get in the right doors and sort of be socialized to believe that that there is a career for you in a field that is, again, predominantly male and sort of Western educated.

19:24Sacha: on the I think it's a bit of a different topic, but how how do you how do you measure the success of like an AI for humanity initiative?

19:36Guest: Yeah, so it sort of goes back to what I was talking about before in terms of making sure that our metrics of success reflect genuine needs and priorities of the community group that we're deploying this solution within. I'm going to try to think of an example. We're creating a solution now that is in the field of human trafficking. And if we measured success just according to the number of uses of the tool, for example, we might be incentivized to deploy this tool among law enforcement officials who have very easy access to AI tools to training programs, have a lot of incentive to use this kind of technology and therefore the number of uses that we might have from that community might be quite high. But because we have decided that the values and priorities of this project are more to serve survivors of human trafficking and support them in rebuilding their lives than it is to support law enforcement agents looking to intervene in a human trafficking case, we've really defined success in terms of community trust and community uptake of the tool.

20:53We're still in the process of rolling this tool out. So we haven't actually measured success just yet. But as you can see with that example, the way that you measure success really orients how you design some of this work. And we want to make sure that our tools are all oriented towards socially beneficial outcomes.

21:15Sacha: I'm a big fan of AI for Humanity. And I heard of one great example recently, I think it was about two days ago. And I don't know if that fits exactly the definition of AI for Humanity, but I think it was a very positive example of how AI can be used for good. And it was it was basically that you could leverage those cameras that the NYPD or anyways, like police in general has. But it was in the context of of in New York and you could leverage those cameras to basically recognize like emotional, emotional features on a person that you that a policeman or would would arrest and then give tips to the police person on how to deescalate the situation.

22:19So if you would see that the person you're arresting is being is stressed, for example, you would leverage that and then have a suggestion in your ear for the police person to say, OK, now it's time to take a step back. This person is stressed. You could try to like speak slower, like in the escalate the situation. I think that was a wonderful example of how I can be used to to help out in a good way rather than using, for example, these cameras for detection as the as they do a little bit in China from what I've heard. But I don't know what you think about that. I want you. I want to give you the opportunity to. I don't know.

23:05Guest: Yeah. I mean, I think it's a really interesting example because. It always comes down to how this technology is actually built and then how it is deployed and integrated into law enforcement, for example. There is, of course, plenty of controversy around emotional interpretation using artificial intelligence, because what kind of data set are you using and does it predominantly contain instances of white men and women, for example, or men and women from Canada, but then can't generalize to other cultures where emotion might be expressed very differently. And, you know, so it's really a question about how making sure that the data set is built effectively. I think there is also just some controversy over the scientific soundness of interpreting emotion through facial expression.

24:08And then also, you know, does the tool have a predisposition towards reminding law enforcement to deescalate the situation, in which case maybe it's net good, you know, regardless? Is there the risk of a false positive where the model says to the law enforcement? They're really relaxed. You can be more aggressive or whatever it is. So this is sort of when you start peeling back the layers of whether the tool is going to be human rights respecting in practice, whether it is really AI for humanity or whether it is disguised as being socially oriented, but in reality creates harms that are much more detrimental than the benefits.

24:50Sacha: I want to ask you. I want to ask you. I want to ask you a future looking question. Where do you see like the biggest opportunities for AI to benefit humanity in like in the next five, 10 years? Do you have any predictions? I

25:05Guest: think in the next five to 10 years, AI has incredible potential to support with particularly accessibility issues or challenges that people may face. I think whether that is access to goods and services that were previously quite expensive. And I know that there is still a lot of controversy over the use of AI as therapists or as lawyers or as anything else. And maybe, you know, the technology won't replace these professionals fully, but maybe it actually can drive down the cost significantly so that more people can gain access to important services. And then also in terms of accessibility of the world, we know that there are plenty of people who have, whether it's neurodivergence, whether it's physical challenges, and if there are technical solutions that can help people navigate the world in a more equitable way, I think that that's really exciting.

26:12And then, of course, all of the potential in scientific research and breakthroughs around treatments and drugs and just other types of rehabilitation, things like that. I think that's just so exciting. So all of those things have me very hopeful that AI will continue to be hopefully a net benefit to society. I'll

26:37Sacha: give you the floor to share anything with the audience, anything you'd like.

26:42Guest: Yeah, so I would say that artificial intelligence is a topic that's getting a ton of attention, and rightfully so. But I think it's always important to make sure that we are staying very critical of the narrative that's being put forward. Because, again, we know how exclusionary the field is, and we know how exclusionary it is to sort of be the spokesperson for the field. So all this to say that we need to make sure to be sort of educating ourselves from a wider perspective on how the technology is built and how to make sure it's done ethically. I'm thinking about all the examples of the exploitation of data workers and the fact that that's sort of been swept under the rug for a really long time. But I really hope that people start to get more engaged in the field and the barriers to entry start lowering because it's going to continue to have such a dramatic effect on our world.

27:47And I hope that more and more people are involved in having a say over what this technology is capable of.

27:57Sacha: Such an honor to have you with me today. Thank you so much.

28:01Guest: Such a pleasure to be here. Thank you, Sasha.