 Thank you for the the lead-in to what I'm about to talk about. I'm gonna go slightly off script and I apologize to the organizers but you know I was gonna talk today and I will talk today about public accountability and the work I'm doing on red teaming but I actually wanted to start off with like what I was woken up to woken up to this article that published in Politico about the extent to which the existential risk movement and effect of altruism is starting to bleed into DC and I don't know how much folks here are paying attention to what's going on politically but Professor Allen was absolutely correct right there there is a war happening right now and it's actually the war for the soul of how technology is being used and if we think about AI as the next iteration of the great pulling apart it is a battle we're currently losing because for all the hundreds of millions of dollars that the UK government spent in the last few years on responsible AI effect of altruism the manner of a few years has come in and captured the entire government so myself and probably some folks in this room will be at the UK AI summit in a couple of weeks and it's going to be a very different tone and tenor than what we are used to seeing in the UK frankly when I built my practice at Accenture in 2017 even though I lived in San Francisco I set my hub in London because it was the center at which there was applied algorithmic ethics the first Accenture about the first tool that we built at Accenture the fairness tool which is now an entire industry of bias detection and mitigation technologies was built with the Turing Institute and now folks like myself the folks of the Turing Institute said we are scrambling to be relevant anymore and this article that I woke up to this morning was about how their funded internships of effective altruism all over DC and I can tell you I've been aware of this for months we know that they're very they are well funded and they are well organized but so are we right so how are we shaping ourselves to have an affirmative vision what do we stand for and what are we about what is our goal what are we telling children when we work with them when we do projects with them what are we telling them that their goal and their mission is and I think that sometimes we are a little bit confused right so we we are certainly the field and responsible AI certainly is the field that is very very good at pointing out problems and things I hope that my career has been built on making solutions as well and while harm mitigation and risk mitigation is a good thing to do and we certainly should be thinking about it we also need to think about what we are building towards right so Professor Allen had an amazing example of her grandparents who actually worked towards an affirmative vision they wanted something and you can mobilize people when you want a thing and you make them want that thing as much as you want it to so my ask to you is to think through what is it what what is it that we want how can we put in one line the thing we will tell an undergraduate a high school student a PhD student a member of Congress this is what I want and this is how you can help me get it because it really is we have to make our language that that simple so I want to talk a little bit about what I'm working on lately which is building out public accountability and my little slice my answer of the you know my the little thing that I'm trying to solve here is how do we get better structured public feedback because the gap that I saw is that in industry because I you know came from industry in this field it's very well funded right so responsible AI teams exist there's a lot of work happening it's been happening for years government is increasingly aware right so as I mentioned the UK government had spent hundreds of millions of dollars building out all sorts of institutes that actually were focused on responsible AI the only thing I saw consistently missing and I still do see missing is the role of public feedback now we have methods in the US government various governments of public feedback right so you do RFIs you know like people provide commentary and I will talk you through the ways in which all of those simply don't work today so when NTIA put out their RFI they received I believe about 2,500 submissions now how are they supposed to parse through 2,500 five to seven page long essays about people's perspectives and opinions all of which I'm sure are very very well written I'm sure lots of people in this room contributed as well I did to but it is not a tenable way to truly get public opinion another one that we see we see constantly is sort of public commentary on some sort of a site etc. before I worked in responsible AI I taught data science at this at a boot camp and one of my students who actually now is at ProPublica or was at ProPublica did a project where he demonstrated that much of the public cap with the quote public commentary that was provided on net neutrality was actually developed by bots so people created informational bot rings to make it look as if the public thought a certain thing the fundamental problem is even with public feedback it's not always clear what people want in aggregate so the political scientist I can tell you there is no quote people there is no quote public right there is no singular voice so how do we get the kind of feedback we need to improve AI systems and I appreciate using the term human flourishing because I use that as well I mean it's it's a term from Aristotle it sounds very abstract and very strange like what does it mean to flourish I don't know if I know what the answer to that is but I do know that it is critically important that we figure out what that means for us so if the current feedback loop is broken right so right now the way the public is heard is something happens you go viral on social media a journalist picks it up maybe and there's maybe some sort of an outrage about it that is a broken loop and I'll I'll tell you why because as somebody whose job it was to try to fix these things at companies it is very difficult to deal with a contextless screenshot and people being mad like I wouldn't know what to do with that information how to even start with inspection how to actually meaningfully reach out and understand what's happening because now there's an entire you know outraged zeitgeist around whatever it is that people think happened and we don't actually know what the facts are so how do we fix this feedback loop right so red teaming is a is a practice that actually started from the military and it is used in the information security field and I've actually learned a lot reaching out to friends and colleagues that work in InfoSec and what I admire about what they've built is they have built institutions that enable them to remain independent but influential and they've built these institutions within governments within industries and they've always done it collaboratively so one of the themes that we have at this event is like how can we think about collaboration and building collaboratively and I think sometimes it sounds a bit distasteful for those of us who work in in a field in which our job is to speak for more moral decision-making to be working with individuals that maybe we perceive as to be non-moral institutions or non-moral individuals but the reality of the situation is these institutions and these individuals simply hold the kind of access and power we need to achieve our vision and our goals right it is not pretty it is certainly ugly but it is certainly true and what InfoSec has done very very well is managed to create a world in which they are in collaboration with but not beholden to these organizations and that that's the part that I think we have not yet solved and I would I would love to work with you to solve that so red teaming is a practice that exists in InfoSec and what it enables is an independent world of third-party assessors and auditors so for example as my team and I get fired from Twitter where do we go we have nowhere to go the only place we can go is back to industry and that's where most of my team is gone most of my team really wants to work on responsible AI but the best way to do the best way to get your hands on models and fix code and tell people how to improve things and have those and have those changes implemented right now only exists in industry I've been happy to work with government agencies for example the European Center for Algorithmic Transparency and trying to get things like the e-cat off the ground and right now the e-cats actually doing its job you may be following what's happening with X and well with between Elon Musk and Terry Breton and their discussion on how they're auditing their models based on misinformation and they're actually able to get into the room because legislation has been passed to enable people within government to do the kind of assessment auditing now how do we build that world outside of industry and outside of government so I think we will all in this room agree that a good ecosystem involves empowered independent actors so what I did with a couple of friends is pull together the largest ever generative AI red teaming challenge and BD right like no big deal and what we did was something that I don't know if I've seen in our industry before and and I'm very proud of the work we did we managed to bring together every single major large language model company as well as some smaller ones so eight different companies were engaged and involved this was sponsored by the White House specifically the Office of Science Technology and Policy and NIST and every two weeks and actually towards the end every single week I ran a meeting that actually is still run today where it was every single member every single lead of the red team at all of these companies NIST OSTP and civil society organizations including Tera's Black Tech Street and Avid and of course Humane Intelligence and AI Village and what we did was craft a challenge and a competition like it was actually designed as a competition to make it fun and interesting for people but the secret mission within that is to actually build the networks and collaboration in a way that we've not really seen before I was told by the red team leads at these companies that they've actually never talked to their counterparts because they never had a room in which they could safely do that I'm seeing nods from people who work in industry it's very hard like you know industry people don't always have a venue to talk to each other creating that and then creating that an environment where government is in the room and civil society is in the room it's very critical that we got together and we in a sense put aside our individual identities and our affiliations and we decided the purpose of this work was to create challenges that actually fundamentally got at the problems of large language models and generative AI and our role was to understand from the public from regular people how it can be better improved and fixed now of course there are many things about the design of the challenge that could certainly be improved Defconn of course is a competition that that is sorry Defconn is a is a conference that draws folks from cybersecurity so it is overwhelmingly a particular audience of people but it was fascinating to see in that room the hope and the ambition and I was told by more than one person that you know in InfoSec conference it's not it's a very cynical place right these are not people these are people who've seen the worst of everything they assume the worst of humanity right because that's their job and somebody pointed out to me that it had been a long time before since they had since they had seen optimism in the room and I that was probably the best compliment I got that entire weekend so that that that meeting that I'm talking about continues today so what we're building is the is a report that will be published in February and what that report will outline is not just what we found but also what was done about it and my hope is that with that little project we're moving towards closing that feedback loop so today when we get public feedback often it goes nowhere and people will say it feels useless it feels like we're shouting into the wind my hope is that we're actually able to give structured feedback to say hey you told us this and by the way this is what we did now this is a one-off project that was built on good will my goal with my nonprofit human intelligence is to make this something more sustainable and lasting and I'm going to give you some numbers from the DEF CON challenge right so as I mentioned we had eight large language mobile companies it's every company you know in a few that you don't we had 20 hours of people coming in and entering this competition we had 21 different challenges a couple of them were around hacking but most of them were about the concept of what I call embedded harms ways in which large language models can surface incorrect biased or misleading information that could be harmful in society including misinformation you know incorrect refusals meaning you know just a discriminatory output based on how the model is deciding to communicate and talk we also had 2200 people show up and take the competition and that was having a line that was over an hour long out the door we were actually not able to bring in everybody who wanted to be part of the competition and it's something we're constantly being asked to continue today and the legacy does continue we're hosting our next red teaming event on October 25th I started off this talk by talking about what's happening in the UK in a few weeks so my hope is that my little contribution is doing a red teaming exercise with the Royal Society on COVID miss and disinformation where we're getting together COVID experts and climate experts to work with large language models not just to demonstrate the kind of miss and disinformation that can and will be produced but how these models can be used to create at scale miss and disinformation campaigns and you know we we are working on llama 2 and I hope to have the support of meta in that competition and my hope is also that when I'm in Bletchley Park with the small small group of people who have been invited there myself folks like on the cock Andrew straight do she Marta who have been working in this field for a very long time who managed to snag a golden ticket which is what we're calling it can be a contingency to push back against the narrative that I see being very very pervasive in this field so what's next my hope is that we grow this concept of red teaming because structured public feedback is done well when we get the right people in the room who have the right lived experience to tell us how how language models and artificial intelligence can be improved to enable human flourishing what's next is that I would love to work with folks in this room to figure out how we have a coordinated affirmative vision for what we're going to do because as I've mentioned there's a war being fought and right now we are losing thank you