The Dark Mirror: Reflecting on AI's Hidden Dangers

I mean this day and age you're right I mean audio can easily be mimicked a lot of data sets don't necessarily filter out copyrighted information before all the AI stuff came you know came around there was people that would call my grandma and say that I was in jail and it's assumed that everything you tell to an AI it's going to store and remember forever and uses its own training data down the line [Music] hello everyone this is Jason bner from triple helix Corporation and welcome back to our Helix Insider podcast I'm joined in studio with two of my senior developers Sam Sheldon and Andy Webster welcome both of you so today we're going to talk to you about something that is very very interesting that's going on in the very Leading Edge of our our technology space and that is AI um we've had a few conversations related to AI already however um the AI um Paradigm is actually also has some risks and it's a concern for users of early adoption of AI that we wanted to highlight some of the

um things to be aware of when using AI I it's a great new technology but it also does pose some inherent risks um so with that uh let's start the conversation around the big topic is data harvesting and privacy concerns um many folks are concerned that the data being used to uh train these large data sets is inherently being taken possibly without even their knowledge uh and the data that is sensitive to them is actually being used in training and so uh Sam let's start with you on on that uh topic what what are your thoughts on the data and privacy concerns with using AI uh something that I've been looking into and you know is generally interested in is a lot of the you know ethical concerns around uh training data including copyrighted images that aren't necessarily used with the knowledge or consent of the Image Creators so that's actually a big thing in a lot of art communities you'll get um there are tools for example that would let you generate images based on the style of an existing

artist that is a really controversial thing a lot of those artists actually don't want images generated that way of their art um and a lot of data sets don't necessarily filter out copyrighted information for example what I know one of the big data sets I'm not going to name names because partially because I don't know if I can pronounce it correctly and partially because there's a lot of data sets out there but data sets that are just lists of URLs to images with appropriate metadata don't necessarily need to worry about copyright on the training data side because they don't have any copyrighted information but they assume that their end users will certainly filter for copyright and you know legally usable images but in practice who knows it's not clear if that's ever happening in a lot of cases so to combat that as sort of a risk in using these AI tools there are actually tools that artists use um glaze and Nightshade both come to mind that are tools for um glaze will mask your style

essentially so that's specifically to fight against the you know do give me an image of bananas and the style of a given artist it'll give you an incorrect style because they've glazed their artwork and then Nightshade is the other one that I think is probably more relevant to most people and it's a tool that essentially poisons The Well of training data so for example if you're asking for an image of a cow but the data set the AI has been trained on a data set that uses a lot of night shaded image data it might give you a picture of a handbag instead because they've because the code that they run on those images makes the AI think ah yes cows have smooth leather handles and are square that that's the that's actually the example they use on nightshades website is cows and Handbags it's the only example that I remember them using I don't know why they were so fixated on that but it's very interesting because it doesn't change the way the original images look to human but it changes the

way that the training algorithm work so there's a bit of an arms race going on there yeah and that's that's good from like uh you know people are defending themsel standpoint but from an ethical standpoint it's kind of like okay should they have to be defending themsel in this manner they're putting a lot of extra effort into it but yeah and for the people who are using the AIS for image generation it means that at some point there might be a whole Rush of unreliable images where you you know you're you're trying to generate images of cows or whatever and you just get something completely irrelevant no that's an excellent point um you know we all know of a few of the major players in AI image generation and you know because we're interested and curious about this we've invested in in some of these tools to learn about them um there was one other image generation tool and I don't have it at the top of my my head but I do recall that what was really unique about this one they're not as big

as some of the major players in that they would actually site all the sources for an image that was provided to you from who the original art creators contributed in the training set to make that image and they would they would attribute the artists and it got me thinking you know we all familiar with like music streaming services right like you know you go buy a CD that you physically own the music or you rent it so to speak from like a Spotify and the new economy for how those artists are being compensated is that they get paid fractions of a penny per play but if you have millions and millions of plays then you you make some decent money and I see an economy starting where the AI generated tools and the um data sets if it follows that same model we could see the artist and creators getting a residual you know Pennies on the generation uh payback because the the tools um and this obviously requires the whole environment to be more ethically minded but you know thinking about how an artist

could actually make a decent living and create their art that they love but then if it's used in generative training and AI they still make money off of it I think that's a real win-win for everyone some thoughts to consider I suppose all right I'm G to do a pivot now so we've actually talked about some of the the ethics and pitfalls of of you know copyrighted data and private data um let's turn to some of the other challenges we're aware of with AI is the potential for scams and misuse uh we know AI tools are actually getting quite sophisticated in that they can mimic real people uh not only in the way they sound in their tone or their tenor but actually even in video um there was a pretty famous um video of a congressman who had actually recorded uh or trained an AI in him through I think 30 hours of of his speeches and then he gave the AI script and had him saying some crazy things that he would never say um and then he played that in front of Congress and he said you guys see this

and it looks like me but I would you know me I would never say that so you know we get to some real challenges where it's like I got a video of so and so saying something and but is it real how how can we protect ourselves from that um Andy I'd like to start with your thoughts on some of those things yeah I mean I think the potential for that is is really pretty dangerous I mean like I live in the computer world so when I'm looking at like what would be a deep fake I can kind of like see like wait a minute his mouth isn't moving in a realistic way you know he might be he's not a real person but like um I can just see potential if people are uneducated in the fact that these things exist then um you could have real potential for um you know duping of the masses you know like by by making a politician say something um say something bizarre but um you know and honestly the best and the best kinds of lies are the ones that have a some truth in them or a lot of Truth and a little bit of a lie

so you know you could um I could see right now this technolog is kind of in its infancy I would say right now but in given it a couple years it would become a lot harder to uh actually determine the difference between a deep fake and uh a a real a real person speaking I I I recently actually just saw I was just going through YouTube and I I just see this ad and it's like Blake Shelton giving away free crossbows and I'm listening to it and I watch the voice so you know I know what he sounds like and so I'm just listening to it and I'm thinking that doesn't sound like him and I start looking at him and I'm like and that doesn't that doesn't really look like him either and and then I go look it up and sure lo and behold there's like actually a scam with there somebody made this scam video and legally P you know they legally purchased an advertisement space on YouTube and and you're just going through YouTube and you get an ad for this it's a complete scam you go there and they have some kind

of um it's it was some kind of classic scam that they uh they rope you in for an extra1 $200 somewhere to ship it to your house the free crossbow and and then they get you and um you know that can be dangerous for um for people and I I think yes we need to find ways to combat it directly but also just we need to really make sure people are um informed that things like this exist and that they should probably have their antennas up when they're watching literally anything on the internet and I mean it doesn't even have to be watching either because I know that chat GPT has been fighting for a while with people that use it to generate fishing email templates which is timec consuming and annoying but get the AI to do it you can send extra you can send as many fishing emails as you want with those templates and if you're if you're sending that to you know a business email list that you got by whatever means how many people are you going to catch with that because I mean I don't know about

you but we get like our emails aren't all over the place but we still get a decent amount of attempted scam emails they're usually pretty bad they're usually just people pretending to be Jason asking for our mobile numbers for some reason um but but imagine you're sitting at work and you're waiting for a document from some other company and you get a you know Microsoft looking email that says something about retrieving your documents and you click the link you go to the site put in your password and no it'ss a scam I I think you nailed it right with your first comment there andy is that awareness is key right and to to your point Sam you know we've gotten ourselves educated through repetition we see so many of those emails that come through that we immediately all right you know Jason would never send me an email asking for a Walmart gift card but it can get more subtle than that right and so a lot of you know what's out there in the technology you have to kind of say to yourself you know

believe that it could be that complex and always uh the the adage I've always been hurt is trust but verify you still have to do business you still have to work with your colleagues so you trust them but you also want to verify that is the information you're getting is real so if you know if I ever sent an email to you guys or vice versa and it was kind of a weird request to get via email you would we would call each other and say hey was this really from you and you know so many of those a scams could be prevented this way where you know the boss who might be a little bit pushy and he tells his assistant I'm walking into a meeting I need $50,000 why do this account stat don't talk to me and and they they rely on on fear and like immediacy right I I think you mentioned that on one of our earlier podcasts Andy and just the fact that this stuff is out there the key is to stay vigilant and be aware of it and and don't think just oh it's a video of course it's real it's not true anymore yeah

and also keep you know inform people you love too you know like like before all the AI stuff came you know came around there was people that would call my grandma and say that I was in jail in Canada or something and that and that they needed to wire you know bail for me to get out I they were PR pretending to be me saying that I was you know in jail in Canada and needed bail to get out of to get out of jail and uh you know my grandma did the right thing she called me and she's like [Music] Andy are you in jail I'm like no I'm sitting in my house somewh I'm fine but like uh with the ability of honestly I think a voice a voice voice is actually potentially even a little more um dangerous because you can't see the the video or the face or anything like that to it's it's not as sophisticated yet so like a but like AI generating scams through voice they could just call you on the phone and kind of uh you know have somewhat of a conversation there will probably be a delay which will be the

giveaway when it's calculating what to say next but um yeah you know inform your inform Your Love All right so let's pivot to the last topic we want to go over and that's actually um false confidence and Rel liability and AI generated data so obviously there's a lot of data in there and there's a sort of a tendency because it's so new and so interesting that anything you consume via the AI landscape is immediately factual and and valid but we all know even just with the internet like um that that old phrase um don't believe everything you saw on the internet and that quote is attributed to Abraham Lincoln and obviously not right not real but it's a funny take on that and so even with AI data you have to be sensitive to the accuracy and validity of that uh Sam talk to our listeners about what that means in our world and and the space that that you play in yeah so false confidence and just general reliability issues so I think I described it at one point as the AI can be very confidently

wrong um so there was actually an interesting case with Air Canada not too long ago where they have a chatbot on their website that you know customers they use it to get information on policies on flights whatever the case may be um basically a standin for sales rep um so someone had called had you know reached out to the chat bot and asked on information asked for information on a specific policy that they had and the chatbot was confidently wrong in its answer and gave them a completely inaccurate response it linked to the page that would have had the correct information but the user understandably took the chat bot's confident reply as oh okay this is the policy and that ended up in small claims court because as it turns out the policy was incorrect and Air Canada needed to functionally needed to honor it because it was their chat bot and it was working for them and it provided information that was wrong and it's a potentially interesting legal precedent there not in the United States

but definitely shows some of the risks of using essentially unmonitored AI as part of your business that if that AI provides incorrect information it's just like an employee essentially but you don't know what's going on in their head and you don't really have a great deal of control over their training so true so true Andy thoughts on that yeah I mean I think my my takeaway from my lesson from that story would be that AI still needs human review on on everything that it does and when you blose an AI on out like that that you kind of need to make some identifying factors that show that it is in fact an AI so that people can know okay this is an AI it might not be telling the truth it might think it is but it might not be telling the truth right and that's an important point it's not a malicious like inaccuracy it's just how it was trained and what it's presenting for sure yeah and interestingly that can also have some implications to I know that AI has been used in like is being used in

hiring and stuff it can introduce some unfortunate bias that may actually have legal ramifications because the AI might through no fault of its own discriminate because it has all this information about you know past High performers that you'd want to hire all of whom happen to be middle-aged white men and it says oh I see this is the criteria for Success let me filter out all the ones that don't meet this you know bar of success that we've listed here which is uh problematic yeah absolutely that sounds terrifying and never mind what it might be doing with the data you feed into it it's gonna learn from that too you know we've gone through quite a few of the interesting um pitfalls and things to be aware of with AI and I guess you know at this point I'd like to kind of wrap up but I think it's probably safe to say that you know with AI and how interesting and um possibly terrifying it is it's all about striking the right balance and so you know recognizing a new um just very very on the

Leading Edge technology also has inherent pitfalls that you know Eyes Wide Open when using it right so you know final thoughts um I'll start with you Andy and then I'll wrap up with you s yeah I would just say um for one because of data being harvested by AI just if you want to be cautious then take down any public information you wouldn't want an AI to be scraping I mean it probably are it may it may have already done that but um in that sense you know and um and like Sam was saying you know about artists with art taking counter measures to um protect your work because it really is the wild west with AI out there right now so you kind of have to defend yourself cuz there isn't really any any protection in place yet good point excellent point Sam final thoughts yeah so I I'd say be wary be be wary of the technology that you can't look under the hood very easily it's assume that everything you tell to an AI it's going to store and remember forever and uses its own training data down the

line don't nothing is secret when you talk to the AI and you it is it it doesn't have it's not actually intelligent it just not it just knows what what the next you know most common word or pixel is going to be it doesn't know why excellent and I guess I goes right back to that that main point is the Striking the balance be be aware of what it is you're using so okay well that is all the time we have today and I wanted to thank my special guests Andy Webster Sam Sheldon my two senior developers uh for being in the podcast with me today and until next time this has been the Helix Insider podcast and if you like the uh podcast please be sure to like share and subscribe down in the comments below and until then we'll see you next time thanks everyone bye-bye