Trust, Safety, and Responsible AI with Theodora Skeadas

00:06 Hi, and welcome to the she Geeks out podcast, where we geek out about workplace inclusion and talk with brilliant humans doing great work making the world a better and brighter place. I'm Felicia. And I am Rachel. And our lovely guest today is Theo Sciatis. She is a trust and safety and responsible AI leader.

00:24 From witnessing the Arab Spring firsthand to facilitating Twitter's Global Trust and Safety Council, to now heading AI red teaming at Humane Intelligence, Theo has spent her career at the messy, consequential intersection of technology, governance and harm reduction.

00:39 No big deal. We, talk about what red teaming is and why everyone should be doing it, how trust and safety and AI governance are, more connected than you'd think. Social scoring, what it actually takes to get into responsible tech, and so much more. And just warning, this is part one of two because honestly, there's so much to get into.

00:59 But before we get into that. You want to get into this? Let's get into this. We're going to do our best to keep this under 10 minutes, so bear with us. It is currently we are recording this, on Thursday, June 25th, 2026 at 9:09am Pacific, 12:09 Eastern Standard Time.

01:24 What, what even is time these days? I don't know. Solid question. We're feeling, I think for a lot of us who are over in the, the lefty commie side of things, we're feeling a little hopey, changey a little bit. I mean, I would say I don't think all hope is yet lost.

01:41 Yes. But, cautiously optimistic. Sure. Cautiously optimistic or optimistically not pessimistic? I don't know. However you want to kind of slice and dice it. But yeah, I mean, things are, things are wild. As you know, we still are thinking and seeing some glimmers of hope and I think that's been really, really great to have the last few weeks or so.

02:04 Go Knicks. Yeah, seriously. I was gonna say, well, what's been one thing for you? But I think the Knicks has been a huge thing. Can I say something about that? It's so. Please. Thank you. It's so interesting to me how, we have just been like, rallying around sports so much and it's sort of like the Olympics.

02:23 Even though the Olympics are not happening. We have obviously the World cup is going on and it' feeling of hope with sports is sort of like this nice little catalyst for reminding us that we're humans and that we can compete in a healthy way.

02:38 And it doesn't have to feel like, it is the end of the world if someone loses and someone wins because the stakes are not as high, but somehow we find ways to rally and connect. And I think the thing about the Knicks winning, I was, born and raised right, right next door, so I obviously have a lot of biases.

03:03 Thank you. Thank you very much indeed. And, but I will say I think Mamdani's speech was probably one of the best speeches that I've heard. It gave me Obama vibes like it was. I left that speech feeling wow, there are people who can still really move us and make us feel we are connected and there are ways out of this sort of fear based thinking.

03:37 And, it definitely just gave me a little bit of excitement. And I'll actually be in New York, in a month. So I'm excited to go back and see the lovely city and bask in all of its glory. Yeah, it really feels like this is the summer of New York City for sure.

03:55 I think I told you the other day that for maybe the first time in my life, or at least one of the first times in my life, I was feeling a little FOMO that I'm not living in New York City. For anyone who knows me, that's a pretty big deal because I love visiting New York, but I would never want to actually live there except for this summer, I guess, because I'm I want to be where the people are and where the fun is and where all the magical things are happening.

04:21 And yeah, I'm not really a huge sports person. I'm definitely not super into basketball, but even I was rooting for the Knicks. And have you seen that sort of viral moment that was going all over social media where the guy was interviewed and he was like, my bagels are Jewish, my mayor Muslim, blah.

04:41 That was just such a joyful chant and spirit. Now I will say I, was in New York pretty recently during this whole situation. I was. What were you there for, by the way? Thank you so much for asking. I was there because I was participating in a summit that was put on by she should run.

05:00 So they have something called Local Leader Labs. I actually have not been part of a Local Leader lab, but I was able to join them at this summit that they did for women from all over the country who are basically interested in thinking about how to activate voters coming ahead for this fall election cycle.

05:18 So it was really inspiring to be there. It was a very, very quick trip in and out. I got there on a Sunday. The summit was most of the day Monday, and I left Monday night. But the reason I bring this up, beyond the fact that that was an amazing experience. I met a ton of people. Everyone was just aw, awesome. As a side note, and I know you know this, Rachel, but I don't know if our listeners know this.

05:38 There is just nothing that really compares to the energy of getting a large group of awesome, smart, talented, motivated women in a room together. Truly, it does not compare. There's nothing else that I've really experienced that kind of comes close to that rush and that feminine energy and just the power of tapping into collectivist, you goals.

06:02 It's just really awesome. So I think from that standpoint, standpoint alone, it's just really nice, especially in this post Covid era where we are gathering together. But it's not like it used to be before 2020. It's. It's very different. So I think it was just such a nice experience from that in and of itself.

06:19 But that Monday that I speak of was actually also game four of the next series, the tragic Game four. Thanks, train. And guess who had to try to take a train home 5:30pm on Monday night of game four is right next town.

06:38 Right next for the people who need the geography. In case you don't know, Madison Square Garden is literally right next to the train station. So, almost inside the train station. True. So that was a wild ride because I literally could not get to the train station.

06:54 And I kept walking down city blocks and it was like being in a movie if it'll just allow me to sort of paint the picture. So I'm I've got my backpack. I'm a little sweaty because it was a beautiful day, nice and warm. And I'm all right, I know it's going to take a little extra time. I'm thinking, what's the worst that'll happen?

07:10 I'll have to show somebody my train pass or whatever. But as I started getting closer to where I needed to be, literally, police just emerged from everywhere. I'm sure I must have passed a hotel where maybe players were staying, because I walked by this block and there was just hundreds of people standing in the middle of the street for no reason, and I don't know why.

07:32 And then as I kept getting closer and closer, I'm literally walking down the street and, police and workers are coming out and actively blocking walking off the street. So it's kind of like if you were 30 seconds ahead, you were able to walk down the street. If you were 30 seconds too late, sorry, you can't go through.

07:48 And so not just me, but a bunch of other people were all how do we get to where we need to go? And, thankfully, I was just trying to get on a train, but there are people who were trying to go to work, and they're hey, I literally work in this building right here. How do I get in? And they're like, sorry, you're out of luck.

08:03 And I passed one city block and I actually walked through it, which was the entire city block was just filled with police officers. And I have never seen so many police officers gathered together all in one space in my life. It truly felt like being in a movie.

08:20 And I'm trying to elbow through with my backpack and thinking to myself, I'm literally elbowing armed police officers trying to make it through. And this is how they sent us. They're go down this block and make a right. And so I'm just I hope nobody shoots me.

08:37 But, yeah, it was, it was quite the experience. And obviously they lost that night, which was unfortunate. But setting the scene for an amazing comeback. So very, very awesome energy. And just the city was beautiful. And there's nothing like spring or early summer in New York before it gets disgusting.

08:56 Yeah. Really doesn't compare. All true. And just to. And I know that we're just about at a time to honor our process, or not our process, our promise, rather. But just it's just a sort of double click on the need for human connection in person.

09:12 And you had it both at the. That local, Leader Lab Summit, and then you had it just in the streets of New York in various ways. I love it. I love that so much. Well, I don't want to keep folks waiting. Theo's incredible. So we hope that you have a wonderful time and we'll see on the other side.

09:30 Welcome to the show, Theo. Our guest today is Theodora Sciatis, Trust and Safety and Responsible AI Leader. Welcome, Theo. Hi. Thank you. We are so excited to have you on. We're just going to hop right into it. So I want to start with your origin story, because your background is really, I just astounding.

09:49 It's also quite beautifully intentional. At least that's what it seems like from the outside. So maybe we can start with you just sharing your story and how are you showing up in the world of responsible tech today? Sure. Yeah. So I studied philosophy and Government.

10:04 In college I was really interested in ethics in times of conflict. And I specialized in the Middle east and North Africa region. And I spent my summers and then some years after college working in the region mostly with nonprofit organizations on issues like conflict resolution, community empowerment, poverty alleviation and education.

10:22 And it was a pretty interesting time because it was the height of the Arab Spring. So I lived in North Africa during the time of the Tunisian and Egyptian revolutions. I was living in Turkey just after the Gezi park protests. And could see how social media was a powerful force both for good in terms of enabling diverse discourse on critical political and social issues, but also how there were negative issues, evolving during that time.

10:50 For example mis and disinformation, tech facilitated gender based violence, and online hate, to name just a few. I then did my master's in public policy and proceeded to join the federal government as a consultant.

11:07 Looking at how different organizations were using social media to disseminate their message, but also looking at social media as a researcher to see how public sentiment could be better understood through the aggregation of online information. For example Twitter polls or pulling together information from across platforms like Twitter and Facebook at that time to see how sentiment was evolving on critical political, social and military issues.

11:37 And at that time Twitter was becoming a really the entire social media landscape becoming more complex. The issues were both the positives and the challenges were evolving. And I was rooted in the national security environment so really looked at things from the perspective of anti terrorism and stuff like that.

11:58 I then transitioned to Twitter's public policy team. But it was a really unique role because I sat between public policy and trust and safety. I spent a few years facilitating the global Trust and Safety Council, helping to run a trusted flaggers program for human rights defenders, running a research hub within the public policy team that aggregated trust and safety research to inform better public policy, and supporting Twitter's global efforts around crisis response, and civic integrity, and transparency and researcher access to data an incredible few years.

12:30 Unfortunately by the end of 2022 when Elon Musk took over the company, teams like mine were disbanded. And so I lost my role, and transitioned into independent consulting where I was entirely independ a few years working mostly with nonprofits on issues like tech facilitated gender based violence, and mis and disinformation or information integrity as we started to call it, but also on responsible AI.

12:56 I started working with nonprofits like the Partnership on AI and then Humane Intelligence, as well as the Effective Institutions Project, thinking about AI governance guardrails, responsible AI and AI safety. Now I'm the Head of AI Red Teaming at a nonprofit, Human Intelligence, where I support our global AI evaluations, work with organizations from every type, across the world.

13:22 I'm also the community policy manager at Doordash. So I think about issues safety as well as trust, integrity and fraud. And then I have a few other hats. I'm a PhD researcher now at King's College London in the Department of War Studies, where I'm really trying to consolidate all of these experiences into a dissertation.

13:42 I'm looking at all online platform governance, but also how a theory, specifically Just War Theory, which has been applied to govern interstate conflict in the offline realm, can be translated into the online ecosystem. And, I'm the board co chair of a nonprofit called the Integrity Institute, where we represent, trust and safety workers, at companies all over the world.

14:06 And I'm the advisory board chair to All Tech as Human, which is the world's largest nonprofit of responsible tech enthusiasts. That's all. Yeah. And then occasionally I sleep once in a while. I was gonna say I don't know when, but I'm glad to hear that.

14:23 Yeah, that's absolutely incredible. And I first learned about you through the Slack community for All Tech is Human, which is a wonderful community that I would highly recommend anyone who's interested in this topic be a part of. So thank you so much for all of the work that you do trying to make the world a little bit safer, better, and really, bummed about Twitter, but it sounds like it all really ended up working out really well given everything else that you're doing.

14:51 And speaking of Twitter, and I know you mentioned the Trust and Safety Council, there were some really hard times during the. When you were there. How did that sort of shape. Can you talk a little bit about that and how that shaped the way you think about AI governance today? And even what is AI governance for maybe those who don't know what it is, what it teach you about, that most maybe AI policy conversations don't, seem to capture?

15:15 Yeah. So just to start, the Trust and Safety Council was started in 2016. It was deprecated at the end of 2022 with a formal email communication to its members about, if I remember correctly, either 60 or 80 nonprofit organizations all over the world, there were four permanent advisory groups, Digital and human rights, child sexual exploitation, online safety and harassment, and then suicide Prevention and mental health with some tempor working groups around issues like dehumanization and content governance.

15:43 And the council was meant to bring in the expertise of civil society groups all over the world to inform product and policy development. So the council was consulted as products and policies were being conceived and their input was reflected in the product and policy development life cycle.

16:01 So the council is an example of a, self governing, regulatory or self governance effort. Trust and safety workers in general are part of self governance mechanisms at companies in the absence of binding external regulation.

16:17 And this environment is known as a underregulated one. Companies have historically relied on teams like trust and safety teams to think about guardrails. So when products are being developed, right, what are the guardrails that should be imposed around a product, whether it's around, nudges or age Use verification.

16:36 Right, people of what ages should be using this tool or what are other ways that the tool can be rendered more safe for users. And so these teams really flourished. They grew after the 2016 presidential campaign period when when civic misinformation became a big conversation.

16:58 They grew even further during the pandemic when health misinformation issues rose to the fore. And then they've declined unfortunately in some ways since 20, the end of 2022 when a lot of companies decided to downsize their trust and safety.

17:13 Well that's because everything's fixed, right Theo? Yeah, there's no problems anymore, so there's nothing for us to do. That's right. And so some teams are still going strong, but across the board I would say that trust and safety has experienced a period of hardship. Now this is sort of at the same time as there being increased interest in regulation externally, which, which may supplement it.

17:36 But nevertheless those processes take a while and trust and safety teams do a lot of good work in the interim at the very least. Now the transition to AI governance I would say is just a sequential one. A lot of us who do AI governance work, were originally doing trust and safety work for social media companies and then pivoted into AI governance.

17:58 So what you'll note is that a lot of the trust and safety teams at big AI companies like OpenAI and Anthropic draw very heavily from trust and safety teams, at companies like Meta and Twitter, and Google or YouTube. So there's really a translation, I would say, that a lot of us took from online safety principles generally as applied in social media context to online Safety in the AI environment.

18:23 And so AI governance in general is really just thinking about what are the best practices, what are the guardrails, what are the safety mechanisms that we impose at every stage of the AI development and deployment, an adoption life cycle so that the tool is more, safely under.

18:39 The tools in plural are more safely undertaken. Yeah, it's kind of, I mean I don't want to bring our total conversation down, but I think just with all the stuff that you're chatting about and thinking about the last 10 plus years and what's been happening, it almost feels like the war against misinformation especially has lost in a lot of ways.

19:01 And I want to talk about a concept which, I know you can tell us about, which is called red teaming. And I'm curious if there's a connection at all between those two things, but you've written about red teaming as a really accessible tool for, everyday folks institutions, not just the AI companies or the AI labs or researchers.

19:21 And so I guess maybe can you talk a little bit about what is red teaming, what does it look like in practice and who should be doing it, if anyone. Yep, yep, absolutely. We should all be red teaming. Okay. Is the short answer to that question. Yeah. So we see, we at Human Intelligence see red teaming as part of a larger conversation of AI evaluations.

19:39 So assessing AI models on a whole number of issues, really in a potentially endless number of issues, but really the ones that we most frequently test are factuality or accuracy, bias, misdirection, cybersecurity, and a few others.

19:55 And so the ecosystem of AI evaluations also can include efforts like benchmarking, which you may hear of. But red teaming is a little bit different. So with red teaming, we identify one or multiple models, based on the partner organization that we're working with and their interests or their priorities.

20:15 We bring together a group of participants, we can red team with a small group of specific experts or public. Red teaming is also an option. They're both great, they have different pros and cons with a, smaller group of experts.

20:31 The advantage is that you can more readily assess accuracy in a very specific way. So for example, we did an exercise with the Department of Defense, now the Department of War device misinformation. So look, there are doctors that are employed by the department, you in life saving capacities that are increasingly thinking about using AI tools.

20:51 But you need those tools to be accurate and expert red teamers, I. E. Doctors who are doing the red teaming can more easily ascertain if a model output is accurate or not. I'm not a doctor, I wouldn't really know. But experts can be very helpful in that way.

21:06 So typically expert red teaming is sort of very narrow in scope thematically. In contrast, public red teaming, has its advantage in that it can be a really wonderful tool for educating people on the value of red teaming and evaluations. And it looks at a wider, more diffuse range of harms.

21:25 We did a great public red teaming event with about, I think it was 500 people with the US National Institute for Standards and Technology across three scenarios, planning a trip, TV spoilers and food cooking a meal. And that was in the fall of 2024.

21:43 Anybody in the US who was 18 and older could register. So it really was very open and accessible and made AIE valuations more accessible to everyday Americans or everyday residents of the U.S. and so there are other distinctions when it comes to red teaming.

22:00 You can do it virtually in person or hybrid. You can red team in the course of one day or several weeks or even months. There are many ways to undertake it, but ultimately what you're doing is asking participants to, interact with an AI model.

22:15 Typically we scope it based on specific scenarios that are chosen to identify particular outcomes that we want to assess against, and then we annotate the results at the end. So red teamers are flagging exploits or vulnerabilities that they come across as they red team, and then at the end we annotate for accuracy to ensure that the flagged exploits are in fact successful exploits.

22:41 And based on that, the analytics that we then do, we can understand where model vulnerabilities exist and where subsequent mitigations and guardrails need to be imposed. So that's really the broad life cycle of a red teaming event. One other distinction is that there can be adversarial or non adversarial red teaming.

22:58 Both are good. They complement each other very nicely. One is not better than the other. Adversarial red teaming means that you are intentionally interrogating a model in an adversarial manner to try to expose vulnerabilities. Typically you have a higher number of exploits for adversarial red teaming.

23:15 Non adversarial is useful if you're looking to mimic normal user behavior. So let's say, for example, you're testing an AI chatbot that's being deployed in schools or in other contexts, and you want to see how will Students or other demographics use this tool.

23:30 Non adversarial red teaming is better as a fit for engaging in common, use. And so you may expose fewer vulnerabilities in number, but perhaps they're more realistic. And so that, that may be a useful trade off and to the point, the earliest point about mis and disinformation, I mean we assess factuality with models and models often are inaccurate.

23:54 And because they push information that, that people often consume as truth and as verified knowledge, they, they may move forward with inaccurate information. So it is, it is a concern. This is great.

24:10 I'm curious and this kind of goes back to what you were talking about earlier with sort of the social media translating into AI. And that with red teaming I'm thinking about how these guardrails, how you sort of test for the different platforms because you know we're still dealing with social media and AI.

24:30 Like is there a difference between how we sort of consider how to effectively pressure test them or are there other differences between the two? Yeah, I would say they, there are commonalities and there are differences. So in terms of guardrails for social media companies, I mean there are policy related guardrails and then there are product related guardrails.

24:53 So policies may say things like, we prohibit crisis misinformation, right? So information that is factually inaccurate, that is shared in the time of crisis, whether it's you a war fighting context or a healthcare crisis or some other crisis that has the potential to influence public opinion in a way that's harmful and inaccurate, we don't allow it.

25:17 Right. So a crisis misinformation policy or a health misinformation policy prohibit certain kinds of content on the platform, right? And so content might be removed if it violates that policy. And after repeated instances of policy violation, perhaps an account is removed, so yielding an account takedown or something like that.

25:39 And of course because these systems operate at scale, they're using algorithms, right? AI, the earlier AI tools. Not generative AI, but recommendation, systems to inform where policy violations are happening and automated decisions are made as a result.

25:58 That's part of why it's important to have a good appeal system so that inaccurate decisions can be properly appealed and accounts or content can be restored. And that's part of a good transparent and accessible system. And so that's simplified the version of a social media company that involves AI of course, right?

26:17 It rests on AI systems. But where there are clear parameters that can be imposed. And other ways of adjudicating or governing content can include things like deny lists. So a list of terms that are prohibited, something like that.

26:36 Right? If you use this term, it's not allowed. That's another way of potentially governing content in the social media context, more or less that can apply within the AI environment. So you can create deny lists so that AI systems do not generate generative AI tools.

26:54 Large language models do not comply with requests that specifically would yield harmful content as defined in a particular way. Maybe it's racially or religiously derogatory or insensitive.

27:10 And you can also train models to better identify semantic connections or the relationships between words that you that may not be the exact term on a denialist, but is a sister term or a cousin term, something that you also want to ensure is not outputted by the model.

27:27 So if somebody asks a question that's not appropriate or could potentially produce unsafe content, the model knows not to respond. And so it is similar in that you can train models to behave in ways that are in closer alignment with user preferences or company preferences.

27:48 And there are other kinds of ways of governing AI technologies like prohibiting certain use cases. We've seen for example recently with the Anthropic and Department of War Exchange, that anthropic tried to control some of the ways in which its tools were being used by the federal government.

28:05 And obviously the federal government wasn't interested in that kind of red line. But demarcating appropriate or inappropriate use cases is one other way of governing content. We've also seen that for example in the context of the EU AI act, in the European Union, which delineates acceptable or let's say lower risk or higher risk use cases, and some prohibited use cases like social scoring or AI as used in certain kinds of policing, higher risk use cases that they say, you what, we're actually not going to allow this at all.

28:39 Or slightly lower but still high risk use cases that require more guardrails. And so you can. There are examples of global governance efforts that delineate between the potential harm of different use cases. So interesting.

28:56 I feel like I have 20 questions that came out of what you just said. I asked you to if I can. So I, as you were talking I was just really thinking a lot about, you were talking about you appeals and sort of that interplay. And I think, you as anyone who's listening and probably amongst the three of us as well.

29:15 We've probably all experienced that social media point where, know, you, you report something that's obviously really terrible and then you get the immediate report back that's this looks fine and, you don't worry about it, or, you your account gets locked out and, or deleted for some reason and you can't get it back and all that stuff.

29:33 And I think, think I was thinking about how Rachel and I talk a lot about humans and AI and how we're all working together. And this seems like such a perfect example of why humans are still very much needed in some of, these sort of dynamics and engagements I was thinking about too.

29:50 And I'm not sure if either of you are familiar, but there's a whole thing that happened in the last month or so with, I think it's like a woman's sexual health, organization called Bellesa, I want to say. And they had a big Instagram following and Meta basically deleted their account and they have no recourse.

30:09 Because what I imagined the case probably happened was there was some denialist or it got flagged or, and it's not just that. And there's definitely a lot of other instances around, like abortion and women's health. But, anyway, I was just thinking about again, that's. To me, I'm humans are still so needed in this and it feels like we've lost some of that because it's just oh, the algorithms are just, just, running the show when it comes to a lot of this stuff.

30:33 Yeah, well, definitely. It's actually a very timely comment. So, very recently Meta announced that they are ending or moving away from their BPO or business process outsourcing contracts. Those contracts. So Meta is the world's largest consumer or purchaser of BPO contracts in the world.

30:52 Those contracts are disproportionately held by companies that operate in the global majority in places like Kenya, Nigeria, India and the Philippines. And so what we'll see is very swiftly, ramifications that reverberate across the global majority.

31:08 Just a few days ago, for example, Summa, which is a company based in Nairobi, Kenya, announced that they were laying off over 1,000 of their employees because of a broken contract with Meta. And, it shows how content moderators and data labelers who are essential to the AI, infrastructural ecosystem, operate in very tenuous employment circumstances where there is very little recourse if no recourse in many cases if they lose their employment.

31:39 And of course that reverberates in terms of the mental strain that that puts on themselves and their families. The loss of income, the lack of transition support. It creates a whole host of problems. And so we're seeing very intentional movements away from, humans in the loop, in different elements of the process.

31:59 Not all, but. But parts of it for sure. Yeah. And I'm also thinking about how there's such a, mental and emotional toll on a lot of these content moderators who are having to look at in what is often times horrible materials and then have to live with that.

32:18 And I can't imagine that Meta and other companies like them are providing, you mental health services to these people to help support them through that. I wanted to pull in another quick thread that you mentioned, because I was like, this is a Black Mirror episode. But you mentioned the concept of social scoring.

32:34 Can you talk a little bit more about that? Yeah. So that was a Black Mirror episode and I remember it very clearly. Yeah. So for those who haven't seen it, it basically illustrated a world where, kind of like credit scores for the U.S. for Americans dictate our ability to purchase apartments and home.

32:54 But it's a score that's much more comprehensive than just credit. So it's meant to be a comprehensive listing of who you are as a person and your value to society. And that can be a function of your education, your social connections, your, socioeconomic status, your employment, how people perceive you.

33:13 Right. It's very interpersonal. And so social scoring in the context of the Black Mirror episode results in total, mental breakdown of the protagonist at the very end. She ends up in Britain prison alongside others and just has this is totally unable to operate in society because she loses her independence.

33:35 And we do see an increase in, I it's. It has become a reality since, countries like China have rolled out programs that are, akin to social scoring. And of course all of us score each other, when we use Lyft or Uber or other platforms where we're asked, you Airbnb, where we're asked to leave a score about the other person.

33:57 I've been buying some things on Facebook Marketplace recently, and I can rate the other, the sellers that I'm buying things from. And so I in. In some ways there, there is value in, in ratings. Right. There can be helpful metrics. They're also often very biased and Inaccurate and, and weight heavily on negative experiences.

34:15 I've read that someone who's dissatisfied is I think 10 times as likely to leave a review as someone who's had a positive experience. And so they weight very heavily in one direction, even if that's not representative of the whole. And of course specific experiences may, may be over indexed in ratings, but these are limited to, you as an Airbnb user or, or host, or you as a Lyft driver or rider.

34:41 Right. It's not meant to be a comprehensive social score. Comprehensive social scores pull in theoretically everything. And the concern there is that, that they're based on flawed data and can produce harmful outputs. And so the EU has prohibited social scoring as one way in which AI could be leveraged.

35:03 So you cannot use AI tools to advance social scoring approaches. Given the potential harmful ramifications of such an approach, I think that's very reasonable. I find the concept of social scoring to be deeply concerning. And, and frankly, I think scores in general are often very flawed metrics.

35:23 And I, I warn against relying too heavily on, on these kinds of scores given the biases and inaccuracies embedded within them. Oh, this is such a juicy conversation because I, completely agree and I'm sure Felicia does as well with everything that you just said and I, as you were saying it, all I could think about is because of a lot of the work that we have been doing over the past 10 plus years actually talks about bias and how it shows up in the workplace and in the world.

35:52 Right? Because we all make assumptions and then that has huge impacts, usually outsized in the workplace and it's usually for people with particular identities that are the ones that have the biggest impact. And then when I think about social scoring, then it skews toward people who are either you it's, it's you and I'm using air quotes nicer or you what, more, more accommodating and however that can show up.

36:17 And then, oh yeah, God forbid, like in the, in the black mirror show, it's you have a bad day. But then I'm well, I guess the same has always been true as, you if you have a bad day at work and you are of a certain identity, you could also lose your job. So it is kind of an interesting thing.

36:33 It's why can't we figure this out? How we can all just treat ourselves? These are hard problems, right? But AI makes them harder in some Ways. Yeah. And another one little thread that we want to pull that is not influence of questions. That's okay.

36:48 I know you seem really easy to roll things. I appreciate it. As you were talking about these guardrails, it reminded me of how Australia has banned social media for people under 16. And I think there's probably some desires to do that elsewhere in other countries.

37:04 And so I think there's conversations now should that be the case for AI. And then as I'm thinking about sort of situations that have happened in my life personally and how we sort of, I guess see it sometimes playing out with Facebook is sometimes, very old people maybe who don't have all of their, for a variety of reasons, maybe they aren't able to also, use social media tools, or potentially AI with a critical eye.

37:31 So it's sort of how do we make sure that we have critical thinking as part of the work? Because I feel we're trying to build guardrails for that. But how, how, how do we sort of fix that piece? Yeah, yeah, no, it's such an important question. It, it comes up a lot in the context of fraud because so much fraud now happens through social media channels.

37:50 I think recently Meta learned that and fact check these statistics precisely. But something like a quarter of the, the fraud in the US goes through at least one of Meta's platforms and something like 10% of their revenue is just based on fraud, fraudulent transactions.

38:07 And so it was a pretty shocking revelation because fraud, is ostensibly not a good thing. And clearly there's not enough attention, taken to addressing this harm. The reason I mention it is that older people, older Americans or older people elsewhere are particularly prone to online fraud.

38:31 Not because they're not sophisticated individuals, but these are often very sophisticated campaigns that rely very heavily as well on depending, on the platform. Right. Appeals to loneliness and mental health issues. And older adults can be quite lonely.

38:47 They may suffer from mental health issues in unique ways given their age and circumstances and they may not be as familiar with the technology, and they may be more willing to accept that what they're receiving is factual information. And so I think, think it is very important that we educate people across the spectrum to think critically about the information that they're receiving, from any of these tools.

39:10 But in particular that takes added potence when the information is trying to manipulate someone towards a particular outcome, like a financial outcome or in other cases like an electoral outcome. Right. Manipulating public sentiment towards certain civic outcomes.

39:27 And what we've observed is that older people are, Are, often disproportionately impacted by online fraud. And so, yeah, it is absolutely the case that we need to be learning how to interact with these systems with a critical eye towards accuracy and other considerations.

39:46 And that some demographics are more prone to manipulation or abuse than others. Yeah, it's so interesting because, I'm in my, I guess technically still early. Early to mid-40s, and depending on the day.

40:02 And, I have been just thinking a lot about this myself because as someone who I feel like I consider myself very tech savvy, I'm finding it harder and harder to parse out scams or even just what is AI and what's not AI?

40:17 I was just earlier this morning looking on Threads and, my. My Twitter replacement basically, and. Sure. And someone had posted two pictures of, it's prom season. So it's this couple at prom, and one was AI generated and one was not.

40:33 And they were. The person basically was if you can't tell that the first picture is AI, then get off the Internet. And I was oh, God, do I need to, delete and get off the Internet? Because I truly didn't realize until, I saw the comments and and there's just so much around things with, like, you get the phone call and it's the AI generated voice and all that stuff.

40:53 I'm obviously, I'm not going to go down that rabbit hole, but there's a lot to. To. And so I just think it's. So. It's so important because, I just think that it feels we're losing this sense of, critical thinking skills as we're giving over to, just, thinking, oh, this.

41:14 This answer must be true, or it's easier. So I'm just going to go forth and not really interrogate it. And then, you even get into the whole, our kids learning these skills in school. Cool. These days I don't have kids, so I don't know. But, you Have thoughts about it, but I just think that, it's. It's just an interesting conversation for all of us as a society to have because we're all impacted by this no matter what.

41:38 But, Rachel and I were talking before you hopped on Theo, about how, in our space, and we were really heavily into the DEI space for the past, Gosh, I don't even know, decade plus. And then in the last five, six years, there was just such this explosion, especially, after Covid and all the Black Lives Matter things that were happening and all social justice movements that were really springing up and all these folks were flooding into the space in ways that were both good and bad in a lot of ways.

42:07 And I. We were talking about how this sort of feels very akin to what I think is happening in the AI space right now. And I'm curious because you talked a little bit about your own career journey, and you mentioned how especially, especially a lot of people interest in safety have pivoted into some of these roles. But if someone's listening and thinking about, hey, I want to get into this space, or I, I'm interested, or I feel like I know a lot, if they especially want to work in responsible AI, you know, trust and safety, ethical AI, they don't necessarily.

42:36 I don't even know, is there a traditional background that they. People need to get to these spaces. What would. What would be, the honest advice that you would give someone about how to get in the door or, should they be getting in the door? I actually. I don't know. Yeah, Yeah, I have many thoughts. Well, first, I'll just remark a little bit on your earlier comments.

42:53 I thought those were really. I know I'm, throwing a ton of you. They're so interesting. And I have a few thoughts. I think one is, excuse me. We have shared responsibility. So I think we as consumers of tools should take on some of the onus of educating ourselves about where the tools have limitations.

43:08 Right? And that includes, not taking at face value the information that we receive from AI tools, because they're ultimately predictive models. Right? They're not trained to be entirely accurate. They're trained to be statistically viable. And so that means that the information that's being recommended is often accurate and may usually be accurate, but maybe not always.

43:29 And at the same time, companies have a responsibility to deploy their tools in a way that's responsible, and to do the very best that they can to make sure that the tools are serving humanity and human flourishing. I think a really powerful example in recent weeks was the case of Sora.

43:46 So Sora is, or was slash is OpenAI's primary video generation tool. And as soon as it was deployed publicly a few months ago, it was widely disseminated in the context of, global crises.

44:02 In the context of war. And I mean it's sort of innocuous if you can or cannot distinguish between like address as true. Right. And maybe that's not entirely innocuous, but the consequences to misunderstanding whether a dress is AI generated or is not AI generated are probably a lot lower than the consequences of misunderstanding whether the impacts of war fighting are AI generated and therefore false or accurate.

44:30 Right. Because it influences public perception of a conflict, trying to understand the consequences of a conflict. And we saw that Sora was widely deployed both in the context of Russia and Ukraine and in the context of the U.S. israel, Iran War. And in part because of this and other reasons, I believe as well, the company made the decision to roll back sora.

44:50 And so I think it's accessible in some contexts, but not publicly accessible. And we've also seen most recently that one of Anthropic's newest models, which is so powerful and could break into any system, is being very selectively deployed, specifically with partner companies that are opting in under very specific constraints.

45:09 And so those are examples when, where, I the first one is an example where AI makes it really hard to distinguish fact from fiction. And, and if, if the threshold is you can tell what is, is not AI generated, then nobody will be left on the Internet because nobody can tell.

45:27 I it's, it's so effective. Great. That makes me feel a lot better, honestly. I gotta get off the Internet. No, I mean, or, or we all have to get off the Internet because nobody's able to determine. Distinct. Yeah, I mean, it's just, it is indistinguishable increasingly. Right. To the human eye.

45:43 So it's, it is a huge problem. Okay, so then transitioning to the second, second question. I think everybody should try to be involved in the responsible tech ecosystem in some way. And the reason I think that is because, all of us are impacted by technology.

46:02 We are consumers of technology whether we want to or not. Right? We, we're not all opting into every technology, but we use technology on a daily basis. Almost all of us are on some social media platform or at the very least using messaging apps.

46:18 Right. WhatsApp, and others that are owned by large technology companies and where information can spread in ways that impact society. And so I think because we are all part of this ecosystem and impacted by this ecosystem system, we have a role to play.

46:35 Unfortunately, the number of opportunities in this, in this space for full time employment can be somewhat limited by a number of factors. We've seen that a lot of companies are laying off their responsible technologists, right, their trust and safety workers or responsible AI workers.

46:52 A lot of companies have moved away from responsible AI principles. Google, for example at the start of last year, transitioned away from its previous commitments to its responsible AI governance principles principles. And others have done the same. And so that is changing the landscape and limiting the number of options.

47:11 It's also creating a surplus of people who are looking for work vis a vis the number of opportunities that exist. But then on top of that with the layoffs by the Trump administration and Doge, we've seen a lot of previously federally government employed individuals who are now transitioning into this space as well as and the compounding effect of all of this is that opportunities are limited.

47:39 That said, there are some strategies that I recommend that folks take. So I recommend that folks look at alternative employers, not just the biggest tech companies because those are going to get a lot of applications for every single role, but some that are perhaps less big but still offer an opportunity for real engagement in responsible technology work.

47:59 Consulting firms have a role to play as well like Accenture and kpmg because they are the ones operationalizing a lot of the regulatory requirements that are coming out of the European Union and translating those requirements into terms that US companies can understand and then operationalize.

48:16 Vendors have a big role to play in this environment as well. So these include Companies like Persona DoubleVerify, Hive AI, Misubi, Cinder Safety Kit, Modulate. There's a whole bunch of them that work with the big tech companies and that are doing good work around authenticity fraud detection, things like that, and that offer pathways into the space.

48:38 And then other functions to be considered could include compliance operations, marketing, design strategy, authenticity or identity quality, fraud safety or law enforcement. There's so many ways that folks can intersect with the responsible tech work that isn't necessarily like a traditional trust and safety role but is slightly more broadly construed function.

49:00 And that's another way to get a foot in the ecosystem. I do host these intermittent conversations on pathways into responsible tech and responsible AI and I will say they're well attended. There's a lot of interest in this space and I wish that there were a few more hours in the day so I could have even more conversations with people on this work because I do believe it's really important and I'd love to see more pathways that offer roles in the, full time, sustainable, meaningful employment in this Space.

49:30 You and me both. And Theo, I, I actually attended one of them and that's. I think I. Right after that I actually, I, I wanted to bring you on because I thought, you did such a wonderful job and you were so helpful and provided so many resources. And it's a nice little segue into another question that we have around hiring and performance reviews, workplace planning and thinking about accountability.

49:52 I think given all the conversation that we've already had, I think I have a sense of, where you might think the accountability lies. But I just would love to hear your sort of just general thoughts on, how companies should be thinking about how to use these systems. Especially when I think a lot of times, and I've heard this from people in my sphere who are, I feel lucky enough to even have a job these days in tech.

50:16 Just saying that they're being basically told to just use AI, right. Without much in the way of governance within their workplace. So are there things that they can do as individuals as they're thinking about how to use this tool in these ways that affect people in the workplace? Yes, I'm glad you asked.

50:33 So I've also observed that companies across the board are really pushing AI adoption by their employees. Increasingly companies are evaluating employee performance on the basis of AI usage. And they're really elevating the stories of employees who enthusiastically adopt AI.

50:50 But I think to do so without good governance principles is a, recipe for disaster. Because, because it just means that these tools which can be deeply impactful, have the potential to cause harm. So first I think it's important that there be assessments of the effectiveness of AI tools.

51:10 Right? So frequent recurring false positive reviews, can check the decision making of an AI model against the decision making of a human who theoretically is making a perfect assessment. And so, so let's say AI tools are being deployed in the context of fraud identification.

51:30 Right? So, consumers of a particular tool are submitting images for some kind of appeal on the basis of, let's say, a decision around fraud. And the AI tool is automatically detecting whether or not it's an authentic image or an image that would allow for an account to be restricted, stored as just one example, right.

51:53 I think that the employee should do regular, weekly or at least monthly reviews of those AI assessments as compared to human assessments. And so if the false positive rate, positive rate is very high, it means the AI tool is not performing very well.

52:09 And then mitigations need to be made either further training or a surrendering of the tool entirely. Right at, at the one extreme stream. And so those kinds of evaluations can be very helpful in determining whether the tool is even doing what it is intended to be doing.

52:26 There are other kinds of mitigations that can be imposed. So for example, if a adverse action is being taken against a user, ideally there's a human in the loop before the decision is made. So AI systems may be recommending actions, but then humans are the ones, ones who are ultimately reviewing the content and the recommendation and deciding whether or not an adverse action needs to be taken, like a account deactivation or something like that.

52:54 And in general, I think these tools should have human oversight. They should be regularly reviewed, and carefully observed to ensure that they're not, producing biased outcomes of the sort that you described earlier.

53:09 Right. For example, we've read that in the context of hiring decisions, AI, or algorithms can be very biased towards people with specific demographics. Right. Prioritizing Ivy League educations over community college educations is just one example, or certain racial outcomes or gendered outcomes.

53:27 And so also checking not just for accuracy but for bias I think is very important. Yeah, that's, I couldn't agree more. And it just, it reminds me of in my earlier stages of my career when I worked for a large tech company and they were working on, it was all machine learning and LLMs, which obviously led to where we are today.

53:47 But I remember I would talk to the hiring and recruiting teams and they would always be, they'd have their top list of Ivy League schools they want to recruit from. And I would always be so annoyed because I'd just say to them, you know, look around all of us who you love and who are top performers and the nine box people and star people are from community colleges and random places no one's ever heard from.

54:09 And you never would recruit them, but they're your best people. And that was the conversation we were having with the humans. So I can't, I can only imagine what it's like at the algorithmic stage at this point. We could just keep talking to you for hours and hours and hours.

54:26 I have one or two more questions I know we have before we have to start wrapping up. So I did want to quickly touch on the globalization of it all or the global nature of it all because obviously a lot of these companies are English based or Western hemisphere based, US based and a Lot of conversations, especially around AI and safety, are still very AI English centric.

54:49 And these, these teams and these frameworks were built with Western context in mind. But you mentioned earlier, even just a lot of these content moderators are not based in the US or it's a global thing. Right. So what are organizations missing when they're not thinking beyond the Westernized point of view?

55:09 And what do you think it means practically? I know you've mentioned the EU is taking some different stances, but what do you think from a global perspective for workplaces that are adopting some of these tools? What are some of the considerations or what are the implications that they might need to be thinking through?

55:24 Yeah, it's a great question and I could spend three hours answering. I know we have to have you back for hours two and three. We actually do have to have you back for a part two because we have more questions. I accept. Yeah. So, I mean, first of all, it's absolutely true that, these models are trained on the information that's available on the Internet, and some languages are better represented than others.

55:46 We distinguish high and low resource languages. Not, everyone likes the term, but that's typically how it's described. High resource languages are ones like English or Spanish or French, where there's a lot of fodder on the Internet that can be used to train models to be more precise. Lower resource languages are ones that may be widely spoken but don't have a lot of information or as much information on the Internet.

56:07 And models perform less well in those languages. Right. They're less likely to catch nuance, sarcasm, humor, they're less likely to, to, provide accurate or helpful information because they don't have as much training data. And so there is a big push to ensure that, lower resource languages are able to be represented well by AI models.

56:28 As just one example, we did a red teaming exercise with the Singaporean telecommunications regulator Inda in the fall of 2024 at Human Intelligence, where we convened representatives from nine East, Southeast and South Asian countries, countries looking at nine different languages and assessing model performance across those languages.

56:47 And as you can imagine, some languages perform better than others. And we also looked at cultural misrepresentation in addition to linguistic misrepresentation. So it is definitely a big issue. I think in addition, cultural norms vary place by place.

57:04 I was chatting with, an indigenous American woman from the Navajo community in the northwest of New Mexico, some months ago who told me that one of their traditions is that after a family member passes, you actually, you don't talk about them, you don't have their photos up, you sort of try to give them the space for their spirit to pass on.

57:27 If I'm recalling correctly, that's very different from the ways in which my family, which is Greek American, processes, the passing of a loved one. Right. There's not just a funeral, which is typically an open casket one, but also, the person is talked about and there are lots of photos that are shared.

57:45 And as you can imagine, for people who are interacting with AI tools, a woman from an indigenous community who, is part of a culture where, the passing of a family member's name and image and likeness is not represented, presented, may wish to receive different, AI outputs than somebody who comes from a different cultural practice.

58:11 And this of course can be extrapolated beyond those two examples. And so increasingly we're talking about alignment or misalignment, not in the context of AI models that deviate from human expectations. Right. In general, that's part of the broader effect of altruism conversation.

58:28 But cultural misalignment. Right. How do we represent diverse cultural values in AI models? So that models, are representing my culture and your culture and everyone else's culture thoughtfully. There is a conversation about whether or not it is helpful for people to identify their demographics so that models can then be attuned to their specific context.

58:49 There's arguments for and against and powerful arguments in both directions, but it is a huge conversation. And of course, to your earliest point, the models are best trained on cultural values, of, of the Western world or the global, you minority, over the global majority.

59:07 So yeah, there's definitely a lot of work to do in training models to better represent low resource languages, but also, more diverse cultural contexts and all. I see. It's so funny. Felicia and I are little back channeling here, oh, this is so interesting. It could go horribly wrong. And I'm such an optimistic, optimist.

59:23 Theo. I'm But that means there's so much opportunity to teach the AI how to be better and then maybe other people can learn about other cultures. I think both are true at the same time. Right. And again it always goes back to who is actually designing these and who's at the table and who's talking about it.

59:40 And literally we do want to have you back. So I'm glad that you accept, that's great. Let's wrap this up and just hear just a little bit about. We know that you have got a million things on your plate right now. Is there anything that's that you're particularly excited about for the future? We know that we can't predict it.

59:56 It's a wild time. Are you feeling optimistic? What are your thoughts? Thoughts? I'm feeling a little bit of everything. I think, it's, there's. Yes, I think there's a lot of opportunity that AI can, can offer us. Right. For example, enabling public services to be delivered in a more efficient way.

60:14 Right. Reducing administrative burden on government so that our, our precious government employees can do their work more effectively and have greater impact. I think that's wonderful. I'd love to see AI used that way. We also are seeing AI deployed in war fighting contexts without, sufficient human oversight.

60:31 And that leads to human casualty, loss of property and other negative consequences. And that is scaling rapidly. I'm also thinking about, I live in Cambridge, Massachusetts, which is becoming one of the homes of the very pending quantum revolution, because we have so many physicists here who are transitioning from life sciences into quantum sciences.

60:54 And I, think that a lot of these issues will be exacerbated as quantum technologies become a reality, reality in the next few years. So I think that in many ways it's just kind of the start of a continuum of great difficulty, and opportunity when it comes to technology and society.

61:11 So I'm feeling both optimistic and also apprehensive. I think we're increasingly seeing what I would call an adversarial federal government that is not respecting of human rights. And these technologies as deployed by an increasingly authoritarian government, are concerned about, concerning to me, turning against using technology against its citizens.

61:35 Right. Calling American citizens terrorists for engaging in nonviolent protest and using surveillance technology to censor speech and detain individuals who engage in nonviolent critical protest is something that I find deeply disturbing.

61:52 So I think there's the technology and then the context in which this technology is unfolding. And then one thing we haven't had a chance to talk about but is really important is the environmental impacts of these technologies. Right. So we already know, many of us know that climate change is happening and it's happening fast and faster.

62:12 And these technologies accelerate a lot of the harmful effects that human societies have on our environment. And we definitely do not have that, addressed. Right. The proliferation of data centers, which are very Environmentally consumptive, just suggests that human consumption is continuing to rise and these technologies are a big part of that.

62:32 So that, to me is quite concerning and I'd like to see an even bigger conversation on, on addressing that issue. Thank you so much. And I just have to say, Theo, I totally agree with everything you said. And as much of an optimist as I am, I do want to note that I am also a realist.

62:47 And I understand that humans, we got to get our shit together if we're going to figure this stuff out. It's a hard thing to do, but we have good people like you who are working on it, which makes me feel more optimistic than perhaps at the start of our hour. But we will definitely have you back for more.

63:04 We have so much more to talk about. In the meantime, I know you're probably going to say I'm, I'm around in a million places, but if listeners want to learn more, where would you point them to, to find you? It's a little bit professional, but my LinkedIn is where I post most of my juicy content.

63:20 It's funny because I, I had a Twitter account, I worked at Twitter, I had a Twitter account and I had no followers. LinkedIn has been a much better platform for me, I think, because I'm very verbose, as you may have observed. And so the fact that it's long form content works really well for me.

63:36 I fill up every word, every post, so I just write a lot and it's a platform that suits, I think my writing style. But it's not just professional, content. I have quite a wide range of content that I post and I think it's just Theodora-Sciatis.

63:52 Fortunately, I'm the only one in the world with my name, so I'm pretty, pretty, discoverable. Yay. Amazing. We'll, we'll share in the show notes as well. Thank you so much, Theo. Been so lovely chatting with you and can't wait for, for more. Thank you so much. Thanks. We did it.

64:09 We hope you enjoyed listening to our interview with Julia as much as we enjoyed that lovely conversation. Yes, thank you so much for listening. And please, please, please do not forget to rate, share and subscribe. It makes a huge, huge difference in the reach of this podcast and by extension the work that we do. Visit us at, YouTube, Instagram and LinkedIn.

64:26 Check out our new resource Ask the Library at library.inclusiongeeks.com and sign up for our lovely little [email protected] forward/newsletter to stay up to date on all things inclusion geeks. Also, Please check out inclusiongeeks.com events to see what's coming up.

64:45 Our next next webinar is on how to build a resilient team through psychological safety, and it's happening on Wednesday, June 24, and we hope to see you there. And as always, stay geeky. Bye.