uh heimlich Rosenberg uh the founder and chief scientist of unanimous Ai and today I'm going to talk about generative Ai and the threat to human agency so let me start with just a couple comments about generative AI first it is a genuine Computing Revolution it will change society like the PC Revolution the internet Revolution the mobile phone Revolution and it will happen fast much faster than prior transitions not five to ten years more like one to three years and it will have risks some are familiar risks and some are unfamiliar risks I will focus mostly
on the unfamiliar risks because I think they're the most important to talk about and they relate to this key fact that generative AI is not just about creating content it will actually unleash a new form of media media that is personalized adaptive and interactive and Regulators are not ready for this new form of media which can be used as a dangerous form of targeted influence so let's talk for a minute about the familiar risks first is threat to jobs generative AI will create human competitive work products at a professional level at scale that's going to
impact jobs it's already impacting freelance artists freelance writers programmers analysts Radiologists oncologists it's going to impact everybody it's going to be a problem but I think we will work through that problem it's not the first time we've dealt with Technologies threatening jobs then there's inaccuracy of content generative AI produces content that looks very authoritative but it can be error-prone and biased there are currently no standard just certifications for accuracy uh regulators and and policymakers really need to put a process in place to certify accuracy we'll do that I think that will get done it's not
a new problem then there's deliberate misinformation definitely not a new problem gender of AI produces content that looks authentic but it can be fake or misleading at massive scale and we can be buried in fake media papers articles and right now there's no water marking process that's really what we need is a process where every piece of content generated by a large scale AI system should have watermarking built into it so we can trace back where it was generated and know that it was not authentic that it's computer generated not authentic content and I think
we'll deal with that then there's the unfamiliar risk generative AI is a new form of media that can threaten human agency and that's what I want to focus on that's the title of my talk and my thesis is that unless regulated generative AI could be the most dangerous technology for human manipulation that we've had to confront and and the reason is it enables targeted customized interactive and adaptive influence at scale aimed at individuals one by one uh before I describe that I want to just make some definitions so it's clear what I'm talking about when
I say generative AI I'm talking about large-scale AI systems that can create human competitive content at professional quality at scale we normally think of those as documents images videos music computer code but it also includes conversations and in fact conversations are maybe one of the most important pieces of content that that gets created so important that we often talk about conversational AI as a subset of generative Ai and I like to Define it as artificial agents that can engage in flowing coherent dialogue with human users while keeping track of the conversational context and the informational
goals this is not like you know issuing a command to Siri this is real conversations which brings me to conversational agents which will be the entities we interact with right now we think of these as chat Bots that we type and they type back these will soon transition to voice Bots where we just talk in uh normal voice and this will transition to Virtual spokespeople where there's a face in Persona that's uh that's displayed and as you're talking to this virtual person it's expressing uh facial expressions and and it looks like a real person you're
talking to that's the path we're on for sure because our future is conversational uh we will speak naturally to our computers not stiff commands real conversations and computers will speak naturally back to us not stiff answers but genuine back and forth with the AI asking for elaborations clarifications context and in fact the AI will get to know you over time it's collecting data as it's interacting with you this is a new form of media conversational media that's interactive and adaptive form of information delivery that's individually targeted at users at a massive scale when will we
speak to our computers all the time we will have conversations with our search engines with our digital assistants with our productivity software with websites businesses and services we'll have a digital virtual spokesperson that you're interacting with and uh and you will have conversations with them um throughout our our digital lives we will just talk to our computers this will be powered by foundational models large language models from a handful of major corporations and you might think oh if we just regulate those that small number of foundational models or if they behave responsibly we're good we're
not good because it will also be accessed through apis by countless third parties who are can influence the conversations through the API and so you'll be interacting with the website having conversation and you don't really know what the third party is doing through the API and so without regulation you won't know who you're talking to and you won't know what their agenda is which brings me to the AI manipulation problem which relates really to to targeted customized interactive AI generated content so let's say we have a generative AI system and um somebody a third party
has an influence agenda and they want that persuade somebody what they can do is generate targeted influence on that user that's custom generated custom generated piece of content specifically for that user based on information you have about that user their likes and wants and values create a piece of targeted influence custom generated that alone is dangerous and we're going to be seeing that right away um but it gets more dangerous because these generative AI systems can then sense the user's reaction to that influence in real time how does that user respond to that piece of
influence that was just presented to them senses the reaction and then it adapts its influence tactics in real time to increase the persuasive impact daps its tactics so you have this cycle that goes round and round where it imparts influence senses reactions adapts tactics imparts influence and so what you will have is in real time targeted persuasion optimized by AI now I know this sounds abstract but honestly we humans just call this a conversation after all if you're a human sales person and you want to influence somebody what are you going to do uh you're
not going to hand them a document that's not a great way to influence somebody the best way to influence somebody if you're a human salesperson yes uh you pitch them you sense their reactions and reservations in real time and then you adjust your tactics and you you pitch them again and you sense the reactions you adjust their tactics and you adjust your pitch and you go round and round and round and you can be very very persuasive generative AI can now do this and it will be doing this it will customize a pitch for you
a verbal pitch but we craft it based on your history and your likes and your wants and your background we'll customize a pitch for you it will adapt to your reactions at first your reactions might just be how do you know what did you type back but soon your reaction will be what you say back including your vocal inflections it will sense your emotions and your vocal inflections and eventually when you're looking at a real video chat with an AI you'll be looking at your facial expressions as well and and so we'll adapt to your
reactions and it will then adjust its arguments it will make very convincing arguments because it has access to the full world of information and it will also learn what works on you over time not just the tactics that work on you but the style so these conversational AI systems will be more skilled than any human salesperson so why is generative AI dangerous well it will enable targeted customized interactive and adaptive influence at scale there will be personalized advertisements that are just custom images and videos that are designed specifically for to to be persuasive to you
and that's dangerous and already Google Microsoft meta have already announced that they're going to be using generative AI to create custom advertisements potentially images and videos that Target people uh based on their very personal specific parameters but that's just the very first step the real issue is that they will be able to then do this interactively and adapt in real time and so conversational media is the easiest way to do that we're conversational uh media we'll be able to basically talk you into influence objectives and it won't just be advertisements it will be personalized propaganda
misinformation disinformation that can talk you into different beliefs if if uh if a third party wants to talk you into believing that vaccines are bad um it can make very very convincing arguments read your reactions adjust to your reservations adjust that is a much more dangerous piece of uh uh influence tactic then I just handed you a piece of content and so we're talking about AI optimized persuasion and it's extremely dangerous it will optimize what it says to you it will hone in on the most effective Arguments for you it will optimize how it says
it to you it will adjust its style tone tactics to best influence you it will optimize who says it even the voice and Persona especially when it's when it's visual the age gender race look will be optimized for you over time it will learn what types of personas are most influential on you and so generative AI could be the most powerful technology of persuasion that we've faced and this is asymmetric influence and this is a really important point because you could say well oh look human sales people already can do this so what's the big
deal well the big deal is that we humans are at extreme disadvantage when we're talking to a conversational AI um this first of all there's a familiarity asymmetry the AI will know your interests your values your education your income your political affiliation your favorite teams favorite movies you will know nothing about the AI that you're talking to uh much less than you would know just by reading a human sales person there's reaction asymmetry the AI will be able to read you from your real-time reactions your words but ultimately your vocal Reflections facial expressions posture you
will read nothing into the entity you're talking to Black Box there's continuity asymmetry the AI will model you over time it will learn how you act and react over time you will learn nothing about this entity you're talking to because it will have an a new agenda every time there's an information asymmetry this AI will have access to infinite information to make persuasive points to you you will be overpowered you can't judge if this AI is cherry picking obscure facts or not because it has access to the to the whole world of information you're just
a human there's a strategic asymmetry the AI could be trained on human psychology cognitive biases negotiation tactics sales tactics you're outsmarted you're outmatched no human salesperson could have these skills and then there's intentional asymmetry when you're engaged with a with a human sales person you generally know what their intentions are to sell you a product or service you won't know the agenda of your search engine I you could stumble on a certain website uh engage with it have a conversational interaction and you know I might not realize that its real goal is not to give
you uh the sports information that you want but to to deliver subtle messaging through conversation and so what do we do about it well we need to educate policy makers and regulators uh first they need to understand generative AI enables new forms of targeted influence it's not just traditional influence generated at scale uh they need to understand the generative AI the influence agenda could be subtle messaging the user might not even notice the targeted persuasion was woven into a conversation they need to understand that generative AI can learn what works on you personally they will
model your persuasiveness and should that even be allowed should they be allowed to model your persuasiveness to know what kind of tactics uh work best on you so my recommendation is that all forms of targeted generative influence should be highly regulated but especially if it's interactive and adaptive for example conversational influence should be very highly regulated and it's it's very easy to deploy in dangerous ways so again I'm Lewis Rosenberg um feel free to contact me at Lewis unanimous.ai and if you want more information about this particular topic I've written a recent paper called the
manipulation problem generative Ai and the threat to epistemic agency that you can also find online thanks
