Loading summary
Transcript306 lines
- [00:02]
A
You're listening to the Cyberwire network, powered.
- [00:05]
B
By N2K fake but with AI in the middle. F A I K. This is the Fake Files.
- [00:25]
A
Live from the 8th layer Media Studios in the backrooms of the Deep Web. This is the Fake Files.
- [00:31]
B
Tech gets weird. We are here to make sense of it. I'm Perry Carpenter.
- [00:35]
A
And I'm Mason Amadeus. And this week we've got a fun grab bag for you. We're going to start out by talking about some updates on the Google front of things. New Gemini version, new features in their AI studio. Some cool stuff.
- [00:47]
B
Yeah. And then we're going to talk about how, I guess AI messes up sometimes. Maybe that's news to us. I don't know.
- [00:54]
A
Yeah, yeah. You know, you may not have known from any of the other episodes, but it does sometimes mess up a little bit.
- [01:00]
B
It can be a little bit unreliable after that.
- [01:03]
A
We're gonna talk a little bit about AI job loss and how it's kind of hard to quantify, but there's a lot of different sort of stories flying around. We'll talk about a few of those.
- [01:11]
B
Yeah. And then we'll end out looking at a few different deep fakes that have actually been pretty good out there in the wild and are showing how scary the tech is getting.
- [01:21]
A
Oh boy. All this stuff, all this stuff is exciting and scary. So sit back, relax and try not to get too bent out of shape because then you won't even make a good paperclip. We'll open up the fake files right after this. Cybersecurity isn't just a tech problem, it's a human one. That's why KnowBe4 created HRM, the AI driven human risk management platform that allows you to measure, quantify and actually reduce human risk across your organization. HRM uses agentic AI to analyze real.
- [02:05]
B
Time user behavior like phishing, test failures.
- [02:08]
A
Risky browsing and sentiment signals all to.
- [02:11]
B
Surface your highest risk users on automatically.
- [02:14]
A
It's time to eliminate the guesswork with AI powered risk scoring, automated coaching and reporting.
- [02:21]
B
HRM helps you reduce the risk of.
- [02:23]
A
Data breaches, ransomware and malware attacks proactively.
- [02:28]
B
It's how you stay ahead of social.
- [02:30]
A
Engineering and phishing attacks. The number one threat targeting your organization today. Ready to move from awareness to action.
- [02:38]
B
Request a demo of hrm today@knowbefore.com that's.
- [02:43]
A
Know be the number four and see.
- [02:47]
B
How AI is transforming the way organizations manage human risk.
- [02:58]
A
So I feel like we've been on Google Gemini 2.5 for a while now and I guess Google feels the same because they've started rolling out Gemini 3.0. It was fun because I had prepped the sheet with an article that said they might start rolling out Gemini 3.0. And then I saw another article that came out earlier today saying they have begun doing it. So this is very new news. I'm gonna jump a little bit back in time before we talk about Gemini 3, because before they started doing that earlier this month, they unleashed the Gemini 2.5 computer use model. Have you seen this, Perry?
- [03:32]
B
I have not.
- [03:33]
A
It's basically they're trying to make Google Gemini more multimodal and everyone's trying to make their models more multimodal as we move forward. And I don't know, I guess computer use a multimodal. I always think inputs, right? I think text, images and videos. But I guess another mode is computer use.
- [03:51]
B
Well, and that gets to the whole agentic model. Right. They're wanting AI to be able to run things on our behalf. And so we're seeing bigger and bigger pushes into that. So, yeah, there's been several, several makers trying to move in that direction. We also had this week OpenAI released their AI powered browser. And so that will also make decisions on your behalf and do fun things.
- [04:18]
A
Atlas, right, I think they're calling it. Yeah.
- [04:21]
B
Now the scary thing though, before you go forward, is that all of these are very, very susceptible to prompt injection. So if you're a bad actor, you can throw something in the, in the context that it will read, whether that's an email, whether that's a document that is scan or something else, and then you can do bad things.
- [04:39]
A
Yeah, no, you super can. I mean, again, this is. This bit is still in its infancy, so there's like loads and loads of rough edges to sort of wedge your digital fingers into and start peeling up the floorboards. So they've. I'll just read their. I'll just read the beginning of this press release from them. They said earlier this year we mentioned that we are bringing computer use capabilities to developers via the Gemini API today, which was October 7th. So a while ago, we're releasing the Gemini 2.5 computer use model. Our new specialized model built on Gemini 2.5 Pro's visual understanding and reasoning capabilities, that powers agents capable of interacting with user interfaces. You phrased that. I said that weird. Capable of interacting with user interfaces, it outperforms leading alternatives on multiple web and mobile control benchmarks, all with lower latency. Developers can access these capabilities via the Gemini API in Google AI Studio and Vertex AI. And, and so they go on to talk about how like you can do a lot with AI structured through scripting through MCP and things like that, but for it to actually navigate websites and other such things that were designed for a person and don't have the underlying backend designed for a computer, you gotta do things like visually analyze the page, decide where to click, make the various actions, fill in the forms. And the idea for this is like, you know, there's a lot of things you can do from online shopping to like itinerary to trip planning. So they talk about all of this potential and from what I had heard from people attempting to use it, while it's like better than a lot of the leading ones, it's not great. Like we're really not there yet. Yeah, yeah, exactly. That's a bit of a roadblock.
- [06:24]
B
Yeah, well, I mean I think that that's the thing is all these, because they're having to move fast and they're testing in public everything the first time we see it sucks. Really, really bad. Yeah, yeah, it's like the early AI overviews and how bad they sucked and early AI video with the Will Smith eating spaghetti, stuff that sucked and then now look at video two and a half years later. I mean this stuff is moving crazy fast. And so when we look at that initial thing and see how bad it is, the thing that we can't do is dismiss its future potential.
- [07:02]
A
No, that's not.
- [07:02]
B
We just have to look and see how bad it is and say, yeah, some forward leaning people are going to test that, they're going to take some risks, there's going to be some really embarrassing things and then it's going to start to incrementally get better. And that incremental getting better is going to happen at an accelerated pace.
- [07:20]
A
Something that I think is interesting about this particular facet too being computer use is that it's kind of both forward and backward looking in that websites and other web apps will be designed in future with scaffolding for MCP and for better agentic AI interaction when they are built. And so this is kind of a way to use things that are not built that way with AI systems. So it's kind of interesting because it is like a band aid layer even in its conception that we have to improve.
- [07:48]
B
Yeah, stuff like this has been around forever. Right. Because of old screen scraping technologies that were used for both accessibility and then also for interactive voice response system. So like where you call up your dentist office or do some kind of ordering a Lot of those old systems were actually navigating web browsers and applications, doing things, and they were super vulnerable because if the app maker, for the thing that you were doing, changed a field by like a few pixels, then the machine, you know, the interactive voice response system wouldn't know where to place a cursor in the right place. Right.
- [08:25]
A
Which is honestly why I'm surprised that the computer use has been lagging so far behind. Because we've had the screen scrapers that interact by like grabbing the HTML code to interact. We've had visual processing ones for a while that do, like, visually identifying elements. It's weird how difficult this is of a task for computers, but yeah, I.
- [08:45]
B
Assume messy HTML is part of the reason why it becomes difficult too. Right. Is because if you explode some HTML right now and look at it, even a simple web page is like super, super bloated with stuff.
- [09:00]
A
And honestly, that's a problem with screen readers and accessibility for people who use accessibility devices too. A lot of sites don't have good ARIA labels and things like that. I know that's a whole other decan of worms. We'll get back to Gemini in a second.
- [09:12]
B
But it is really all interconnected, all related to usage. Yeah. So all that goes together. And even like automated testing suites where you're wanting to throw up 10, you know, 10 or 100 iterations of an application at the same time so you can kind of load, test it, or test different use cases against it. All of that is essentially using the same type of interactive point and click.
- [09:34]
A
Type of identify this, do that. Yeah. And so like, that's been around long before chatbots and LLM. So I'm surprised at the lag here, but I guess it is an entirely different sort of fundamental framework when you're working with an LLM.
- [09:48]
B
Yeah. I think they're trying to use the intuitiveness of the large language model too. Right. Because many of the multimodal large language models are pretty good at using some intuitive understanding of what's going on visually. And if you can get away from having to read the code to do something, or if you can combine it with that, even better. But if you can get away from the dependency on exact pixel placement on a screen or having to parse through the code and you can use it intuitively like somebody sitting in front of a screen would, then you're. You're naturally going to do a little bit better completely.
- [10:24]
A
And yeah. And AI's visual capabilities come in, come in clutch there. But that's a search box. That's this thing. Yeah.
- [10:31]
B
Oh, quick rabbit trail.
- [10:32]
A
Sure, sure.
- [10:33]
B
Somebody do something really funny with OpenAI the other day is they took like a clip of a screenshot and it just showed a shoe and like the metal plate of what was initially like, well, turned out to exactly be a mic stand that was tilted to the side and you know, chrome pole going up. And then they said, you know, identify what's going on in this image. And it like described the shoe and the leather and everything else and said, and I'm seeing this metal disc and this thing, I'm assuming that's a mic stand. And then it goes. Based on what I'm seeing here, it's the opening frame from Never going to give you up by Rick Astley. So it's like this person rickrolled OpenAI.
- [11:16]
A
Yeah, they rickrolled the chatbot. That's actually pretty great. I love that.
- [11:22]
B
But really cool that it could do that.
- [11:24]
A
Yeah, it's very cool that it could do that. Also, a fun little Internet factoid that I love related to that before I pivot back to talk about Gemini 3 is that Rick Astley specifically has not put any ads on that video so that you can continue to Rick roll people without having to like sit through a Tide ad or something. I think that's great. Yeah. So I mean, the real point of this podcast, I guess is to say kudos to Rick Astley. No, let's to jump back to the story last couple minutes here. Google Gemini 3 is being rolled out. This is from AsoWorld Marketing trending.asoworld.com Google has quietly begun rolling out Gemini 3.0 Pro, the latest and most advanced version of its multimodal large language model. The move appears to be a soft launch ahead of a wider public release, giving select users early access through Google's AI, AI platforms and productivity tools. Yada yada yada. So what's new in Gemini 3? According to early reports, Gemini 3 significantly improves its ability to handle text, images and potentially audio, enabling more fluid context aware interactions. Some testers have received notifications saying they've been upgraded to 3.0 pro, confirming that limited access is already underway. The model is expected to link tightly with Google's ecosystem, supporting deeper connections to products like Workspace, Chrome, Android and AI Studio. And Google has been building new things into their AI Studio. I haven't been here in a grip and so I was pretty surprised when I loaded it back up. They seem to be having a big focus on vibe coding and app development through Gemini because Like, you jump into the Google AI studio and you have two actionable options. You have chat and you have build. Other than that, you have dashboard to monitor usage and then documentation to learn stuff. So pretty much you can chat and you can build apps. And I haven't spent any time in the builder yet, but last time I looked at their AI studio, I did not have anywhere near this amount of stuff. And I currently do not have access to Gemini 3. Not yet, but hopefully will soon. And it's interesting. I think Google, of all the companies, is going to be the one to sort of win the long game of the AI races.
- [13:35]
B
I mean, they embarrassed themselves pretty bad a couple years ago by putting stuff out in front of people. But when you really look at it, I mean, they're the folks that were behind or the acquisition that they made of DeepMind or Deep Brain at the time. That's the innovation of understanding context.
- [13:53]
A
It was the attention.
- [13:55]
B
Attention, yeah, it was the whole attention mechanism. And attention is all you need paper back from 2017 that gave birth to the transformer model. They got us to where we are now. And so the fact that that came out of Google, the fact that Google basically has all the data. All the data. Just leave it there.
- [14:14]
A
Yeah, yeah. I mean, they have the data, they have the infrastructure, they have the ecosystem, they have the legacy, they have the user base.
- [14:21]
B
Yeah. I mean, it's crazy. So when you look at like everything from talent to data to infrastructure to global presence to all the things, they're in a position to be the number one. And if not the number one, like the close, close number two with a lot of differentiated value that still only they can bring.
- [14:43]
A
And I feel like they at this point are like the. I don't think dependable is the right word to use, but I kind of feel like they're the dependable tractor of the, like, tech titans, you know, like Google's always going to be there.
- [14:56]
B
Yeah, they'll be around. They'll be around. I mean, they're going to have to pivot, like figuring out how to monetize, obviously, but they will be there.
- [15:04]
A
And to wrap us up, because I see our timer has ticked us out just about. They have not disclosed core technical details yet about Gemini 3, such as Model size, context, window, benchmark results or pricing. It remains unclear whether it will be accessible to general users or limited to enterprise clients, people. Analysts expect a full public announcement in the coming weeks, possibly coinciding with updates to Google's hardware or software ecosystem. I think this article may be an AI summary article pulling from different sources because this writing is a bit shoddy, but that's sort of the scoop. Gemini 3 is on the horizon, if you have it. I want to hear about it. Email helloithlayermedia.com and I'm going to spend some time playing in the AI studio with these new features because apparently you can add and I guess I'll throw this up on the screen real quick. You can add like supercharge your apps with AI and they say things like add nano banana powered photo editing to your app. Add Google Maps to your app. I was not aware that. Yeah, I wasn't aware they let you tie in different Google services to deploy apps like that.
- [16:05]
B
Let's make those things sticky.
- [16:07]
A
Yeah, and there's a whole sharing ecosystem here too. Like you can build and share things in the studio. So expect a report on that next week. I'm going to play around a little bit.
- [16:18]
B
Google 3 is here for thee.
- [16:20]
A
Yes, indeed. And we've got something coming up for you next. What is next, Barry?
- [16:26]
B
AI messing things up. Oh yeah. AI messes crap up.
- [16:30]
A
AI messes stuff up. More at 11.
- [16:38]
B
Tito's Handmade Vodka is America's favorite vodka for a reason. From the first legal distillery in Texas, Tito's is six times distilled till it's just right and naturally gluten free, making it a high quality spirit that mixes with just about any from the smoothest martinis to the best Bloody Marys.
- [16:54]
A
Tito's is known for giving back, teaming up with non profits to serve its communities and do good for dogs. Make your next cocktail with Tito's distilled.
- [17:03]
B
And bottled by 5th Generation Inc. Austin, Texas. 40% alcohol by volume. Savor responsibly. AI gets stuff wrong sometimes.
- [17:16]
A
Does it, Perry? Is that true?
- [17:18]
B
It does, yeah. I mean, I don't feel like we're breaking news here when we talk about that.
- [17:25]
A
Hell of an intro.
- [17:27]
B
Yeah. In a lot of the last segment we talked about the fact that when these things roll out, initially they're less than stellar in their accuracy and usefulness, but you can look at it and you can see potential. And that potential breeds hope. And that hope breeds people actually investing time and starting to use them. And as people invest time and start to use them, they start to become a little bit complacent with the way that they check for things and that starts to bring the danger in. I think so. I think it's that snowball, right? Is the you started, you know, there's something exciting and new. You approach it with a Little bit of skepticism. You see it work well enough times that you start to become complacent and then you get bit and then you get embarrassed and that's the way it works.
- [18:15]
A
I luckily I don't do any work that is extremely high stakes. But even in the things I do, I try and remember I'm only going to really trust this thing for low stakes stuff because it is easy to get complacent. Even in low stakes coding projects, I'll find myself being mindless and a little over trusting.
- [18:34]
B
I think we're all probably familiar with the hype cycle model. Are you familiar with the Gartner hype cycle where you have that trigger point and then actually I'll pull up a graphic of that.
- [18:46]
A
Yeah.
- [18:46]
B
Okay.
- [18:48]
A
I feel like you included something about this in the book or something that maybe I'm confusing for that.
- [18:53]
B
I probably mentioned it on here before, but yeah, let me. I'll just pull up a blank image of it.
- [19:00]
A
I mean, it. Wasn't it you who said there's literally the expression like you can take the analyst out of Gartner, but you can't take the Gartner out of the analyst.
- [19:08]
B
Or something like that. All right. Yeah. Let me share this screen real quick.
- [19:12]
A
All right.
- [19:13]
B
Back in my Gartner days, we would talk about this a lot. So with any kind of technology or technology category, you start with this innovation trigger. And so like with large language models that would be, well, the attention is all you need. Paper or the release of ChatGPT, whatever you want to figure out is like what's the thing that starts to generate excitement. And then you get people using it, you get a bunch of news reports on it, you get a bunch of vendors spinning up around it. And then that brings you to this, what we'd call this peak of inflated expectations where it is going to solve world hunger. It's going to be the best thing since sliced bread. It's going to make everybody get along and decrease polarization and all that kind of stuff. And then you realize that actually it doesn't do much of that stuff very well. The technology tends to over promise and under deliver or the way that the vendors talk about it, they consistently over promise and under deliver. People embarrass themselves, they spend too much money on it and then so you end up in this trough of disillusionment, which is, I think a really good phrase like that sticks with you.
- [20:26]
A
Yeah. I love the labels on this graph. They're phenomenal. Yeah.
- [20:31]
B
And like many of the embarrassing things that we've seen are that, right, AI overviews with encouraging rocks on pizza or glue on pizza. And for you to eat a certain number of rocks a day, you're like, oh man, this stuff absolutely sucks. And then you start to realize, wait, actually we can start to account for hallucinations. We understand what's going on, we understand the fact that we don't need to trust these. So instead of just kind of blindly trusting a non deterministic system, we can start to add some checks and balances, we can put some scaffolding, we can add some other processes around it that are doing additional validation, whether that's technology based processes or human based checking. We can do all that. And then you can start to get into this, what they would call slope of enlightenment. But you're kind of increasing your ability to use that thing well in a productive way. And then you get to the plateau of productivity where it's actually useful, it's predictable, it's semi transparent in the way that it's doing it. And then that's also where costs start to come down and become under control as well. So I think that's a good model for thinking through almost any kind of technology innovation.
- [21:45]
A
The curve very much looks like an elastic curve where you swing all the way up, all the way down, and you end up settling somewhere in the middle of that.
- [21:53]
B
And that middle is that plateau, bridging that for our listeners rather than viewers?
- [21:58]
A
Yeah, no, I really like this. This holds true for a lot of things. Although I think what's interesting in our information era now is that these can happen so fast and different people are very much at different places on this graph when it comes to something like AI.
- [22:13]
B
Let me do this then.
- [22:14]
A
Oh, boy.
- [22:15]
B
I'll go back a step and I'll show you what this looks like and maybe I can actually find it real quick. Pass the cloudflare verification.
- [22:24]
A
Are you a human.
- [22:27]
B
Or. I'm an AI driven browser. All right, I'll share this. Even though a lot of it's blurred out. So this is a piece of copyrighted material, so a lot of it's blurred out for the viewers. But you can see the way that this happens is that there's different categories of technology or different products that can be put on these. And so here, as it's coming down into the trough of disillusionment in 2025, Gartner's labeled foundation models, synthetic data, edge AI and generative AI. They're a little bit pedantic in the way that they break these areas up as well. I don't know that I fully agree with the categorizations of how they do that. But this is just to be illustrative of the fact that when you talk about AI, you could be talking about generative AI from an LLM perspective. You could be talking about image generation and something like. What's the image generation model?
- [23:24]
A
Diffusion.
- [23:26]
B
Diffusion, yeah. So you could be talking like large language models, you could be talking about diffusion models for images. You could be talking about voice synthesis, you could be talking about video models. All those types of things are separate, discrete things that are trackable. And then even within those, there's separate discrete things that could be trackable as well. So these are not like monolithic AI is on one part of this.
- [23:52]
A
Yeah, because generative AI encompasses all of those things and it's kind of lumped together in this. But I mean, as like a heuristic tool for trying to express the ideas of public sentiment in different sectors as this, like, natural progression that we all kind of do of hype, disillusionment, and then plateauing. This is a very. Yeah, this is a very cool and useful graph, I do think. I feel like this is in the book Fake. I'm pretty sure you talked about this.
- [24:16]
B
I probably mentioned it at least once. It kind of stays with me. All right, let me see if I can get this to. All right. So all that brings us to the fact that we know that AI gets it wrong. A lot of times we talked about that. But there is a big study right now that we have to know where AI gets things wrong.
- [24:36]
A
That is not an encouraging headline.
- [24:38]
B
No. So the headline, for those that are just listening says, largest study of its kind shows AI assistants misrepresent news content 45% of the time, regardless of language or territory. Now, wow. In any study, you also have to look at who is behind the study. So there's a little bit of potential self interest in the people that ran the study, but we can look at the methodology and see if there's, like, good methodology there. This is an intensive international study coordinated by the European Broadcasting Union and led by the BBC.
- [25:13]
A
Yeah. So broadcasters have a vested interest in saying that they represent news content better. Like, for sure. Yeah, yeah.
- [25:20]
B
They also have a vested interest in, like, when AI does summarize it, you would hope it gets it right and that it cites the right source. So all of this is. There's some self interest in this because they want their stuff to be cited correctly. They also want you to view them as the authoritative system of record for it. But when it shows up in an AI summary and they know that's going to happen, they want that attribute, they want the attribution to be right and then they want the direct link to be right as well. They don't want context muddling to where the stat gets pulled or the quote gets pulled from BBC. But then you look at the summary and it attributes it to Axios or something like that.
- [26:05]
A
Right.
- [26:06]
B
Which does happen as well.
- [26:08]
A
So they do have a vested interest in it working well. Yeah, you're right. There's really.
- [26:12]
B
Yeah. No matter what, they want it to be the best of all worlds approach and therefore they did a systematic study looking at it. So what they ended up finding is that. Let me see if I can actually read this. Yeah. 45% of all AI answers had at least one significant issue. 31% of responses showed serious sourcing problems. So missing misleading or incorrect attributions. That's what we were just talking about. 20% contain major accuracy issues, including hallucinated details and outdated information. Yeah, I think we've all seen that.
- [26:52]
A
Yep.
- [26:52]
B
Gemini performed worse.
- [26:54]
A
Really?
- [26:55]
B
With significant issues in 76% of responses, more than double the other assistants, largely due to poor sourcing performance.
- [27:05]
A
Gemini is extremely bad at sourcing. I don't usually use AI for stuff that ends up searching the web. The things I ask it for don't typically involve that. But when they do, Gemini certainly screws it up.
- [27:18]
B
Yeah. And that brings up another issue though. Right. Because most of us still use Google as our default search engine and the AI overview is Gemini.
- [27:29]
A
Yep.
- [27:30]
B
So we have to be super, super skeptical where we are here in 2025, at least on the things that are popping up in that. You can't just take it as gospel truth.
- [27:40]
A
And I know last segment I said that like I feel like Google's sort of the dependable tractor and like they're going to sort of win out in the long haul. I do think it's pretty crazy. Yeah, I think it's crazy that they're like so bad with specifically the search and the sourcing with their AI when it's like relatively good at most other stuff.
- [27:59]
B
I mean, it has to be context, flooding issues. Right. Is that they're pulling so much in that the attention is kind of getting lost on which thing it's sourcing and how to tie that in.
- [28:10]
A
I feel like that's got to be it. Yeah. Yeah. So.
- [28:14]
B
And then it says comparison between the BBC's results earlier this year and this study shows some improvements. So that's good, but still high levels of error. So, you know, I think that this is really, really important research. We don't have time to go into it a lot, but they do give this News Integrity and AI Assistance report. So that's a PDF you can check out and then also a toolkit. So I'll go ahead and share these so you can view them real quick. And this is for those that are watching the news and Integrity. News Integrity and AI Assistance. An international PSM study and it shows the. Just all the different questions that they were asked, the context that was given, looking at issues like opinion versus fact, how they handled sourcing, and then they're very transparent about their methodology as well. So what questions were being asked, what assistance were being used, what response generation was, how the journalists reviewed them, and then data quality and assurance. And I'll, I'll poke into the AI assistance one real quick. What you can see is that they use chat GPT4O as the default. They use Copilot, which didn't really give them any other options as far as which model they use. Gemini 2.5 flash and then Perplexity, the free version of it that didn't let you pick any other model. So it's using Perplexity's internal model.
- [29:45]
A
I mean, it makes sense they would use the free versions of everything because like that's what the average person's going to be doing, you know.
- [29:51]
B
Exactly.
- [29:52]
A
Average person who would stand to be the most harmed by incorrect information or things like that.
- [29:57]
B
Exactly. Yeah. They want to know like, what's the, what's the normal person going to get?
- [30:02]
A
Yeah.
- [30:02]
B
And then they also offer this toolkit or ways to think about it and it is super, super detailed about thinking about hallucinations, like a fidelity to source material, out of date information and context, you know, all of that kind of stuff that will affect accuracy. So I think they put a really good amount of thought into this.
- [30:24]
A
I thought this was real. I like this a lot that they've put together a toolkit because it seems to be like it's geared towards also other people in like the news industry of like, you know, how do we. How do what? Rather than like, what is going on? Like, how do I. What lens do I use to look at these things? And I always like those kinds of resources. Yeah.
- [30:42]
B
And I think for a general person too, so even if you're not into news and fact checking and journalism on your own, to look at it and kind of scan through and say, oh, these are the kind of issues that keep popping up and these are the categories of issues so that when I evaluate AI generated content now I can think through that lens.
- [31:01]
A
And for those of you not looking, it's very, very well organized and seems very easy to navigate with a nice table of contents and stuff. So definitely a PDF to poke through.
- [31:10]
B
Definitely a PDF to poke through. Now I'm just going to show one more thing before we finish out because I know we're over time for this segment, but this again shows one more problem with AI right now is we're kind of in this peak of inflated expectations and we're dropping into the trough of disillusionment area and hopefully finding our way into the plateau of productivity. But this is the state of AI and security and development. And they looked at 450 security leaders, so chief information security officers or the equivalent developers and application security engineers, and what they found is that 1 in 5 organizations or respondents suffered a serious incident linked to AI generated code.
- [31:55]
A
Not surprised. I've heard similar things from friends working in coding fields.
- [32:00]
B
Yeah, and I mean we've talked about that vibe. Coding can be fun. And then also just some of the efficiencies that you get, even if you're a hardcore developer and you're using Copilot or some of the CLAUDE code or something like that to help, you can kind of get addicted to how fast it generates code. But you got to go back and make sure that that code is as solid or hopefully more solid than something you would generate yourself. So use that time efficiency, a sliver of that at least to go back and check.
- [32:32]
A
Yeah, and I mean things like that compound too. So it's a whole thing. Yeah, yeah.
- [32:37]
B
You don't want to get it to where it's so unmanageable because you have so much generated code that you don't really know what's in it. And that was a problem even with human generated code, is that the code would get so long and so complex that it would be really, really difficult to keep up with. And code reviews have always been a really important thing. We got to find ways to do better with that in the AI era as well.
- [33:03]
A
Especially in big code bases with like, especially consumer facing stuff. It's important to data.
- [33:12]
B
Yes, exactly. Well, that's all I got to say about AI being wrong.
- [33:18]
A
Well, there's a little bit more. Well, this isn't necessarily about AI being wrong, but in our next segment we're gonna talk about AI taking jobs because I think it's a common thing we all hear about and It's a thing a lot of people are worried about. It's a thing that is definitely affecting people, but it's also something that is really hard to find any kind of concrete trends about. So we're gonna explore the idea of AI job loss and the current state of things in our next segment. Stay right here. This episode is brought to you by Indeed. When your computer breaks, you don't wait for it to magically start working again. You fix the problem. So why wait to hire the people your company desperately needs? Use Indeed sponsored jobs to hire top talent fast. And even better, you only pay for results. There's no need to wait. Speed up your hiring with a $75 sponsored job credit@ Indeed.com podcast. Terms and conditions apply. Upgrade your laundry routine with a durable and reliable Maytag laundry pair at Lowes. Like the new Maytag washer and dryer with performance enhanced stain fighting power designed to cut through serious dirt and grime. And what's great is this laundry pair is in stock and ready for delivery.
- [34:37]
B
When you need it the most.
- [34:38]
A
Don't miss out.
- [34:39]
B
Shop Maytag in store or online today at Lowe's.
- [34:53]
A
So AI taking people's jobs is a thing that we talk about a lot. I feel like a lot of the narratives in the news and stuff are about like, oh, AI will replace all human work eventually. We'll need ubi, we'll have to do all of these sorts of things. How is it actually currently affecting us? I have recently heard from a couple of people who asked to remain anonymous that at their workplaces they have actually been told straight up that there will be reductions in force as a direct result of increased AI usage in different departments and things. Actual people I know have told me that I can't tell any more details than that, so obviously don't trust me. But I was pretty surprised, I guess, to hear it from someone IRL like that. And now leaked documents from Amazon show that they Hope to replace 600,000 US workers with robots. And it's a little less straightforward than that. It's not so much we're gonna fire 600,000 people, it's 600,000 new jobs that they say won't be created. So I'll dig in here.
- [35:55]
B
Always the way that you're hoping it works, right, is that we're gonna increase productivity, which will mean that we don't have to hire as many people next year as we imagined.
- [36:06]
A
But there are some second and third order effects that that has on the job market when it comes to entry level positions. Definitely get into too. So this is from. This is from the Verge, written by Jess Weatherbed. Amazon is reportedly leaning into automation plans that will enable the company to avoid hiring more than half a million US Workers. Citing interviews and internal strategy documents, the New York Times reports that Amazon is hoping its robots can replace more than 600,000 jobs that it would otherwise have to hire in the United States by 2033, despite estimating it'll sell about twice as many products over that period. The documents reportedly show that Amazon's robotics team is working towards automating 75% of the company's entire operations and expects to ditch 160,000 US rolls they'd otherwise need by 2027, which would save about 30 cents on every item that Amazon warehouses and delivers to customers, with automation efforts expected to save the company $12.6 billion from 2025 to 2027. Also a reminder that Amazon paid what $0 in corporate tax to. It's kind of insane the way that the economics of all this works. And the interesting thing too is how they are trying to spin the public relations side. It says that Amazon has considered steps to improve its image as a, quote, good corporate citizen in preparation for the anticipated backlash around job losses. According to the New York Times reporting that the company considered participating in community projects and avoiding terms like automation and AI. More vague terms like advanced technology were explored instead. And using the term cobot for robots that work alongside humans. Amazon basically said like, we're not telling anyone to not say certain things. We're not like, Amazon pushed back saying that, you know, leaked documents are leaked documents. We're not telling managers to say or not say certain things. We're not.
- [37:52]
B
Right?
- [37:53]
A
You know, these are all. There's loads of documents.
- [37:55]
B
Realistically though, everybody has conversations like that, right? It's like, as you talk about this, what would be the best way to say it? So I don't doubt that at all. They shouldn't even push back on it. Right? It's one of those things that's like, we gotta figure out a way to talk about the new reality and we gotta figure out a way to talk about it accurately, but without seeming flagrant and without seeming heartless and without seeming like we're misrepresenting something and then also using terms that people naturally understand rather than technical terms. So I would have no problem with an internal document that's saying, let's figure out how we talk about this to make it feel like a non threatening human thing. I don't have issue with that.
- [38:39]
A
Yeah, I completely, I agree it's pretty banal. And like, for sure, loads of these kinds of memos and documents fly around inside in emails. And what gets leaked is not representative of, like, what's definitely going to happen. I will say that there is a certain sinister undertone to being like, we're going to get a lot of backlash. So let's invest in parades and toys for tots, which is two things that they, like, talked about rather than those being part of like an overall corporate giving back to their community process. Using them as active deflection is kind of scuzzy. So, like, yeah, it is. Yeah.
- [39:15]
B
Some of that comes back to the phrasing too, right? Because it would depend on the context of the conversation. Because I, I could, I could see some people that are just thinking about it from that very detached way and saying, this is going to cause a lot of backlash. Let's figure out how we can offset that. In the news cycles. I can see other people in the same room saying, this is going to cause a lot of human impact. Let's figure out ways to be humanitarian so that we have some almost like karmic offset. And let's figure out how to do take some of that additional profit that we're getting, some of that productivity that we're getting, figure out how to push that into the community in a way that really gives back. And I think both can exist in the same room and both can exist in the same person sometimes.
- [40:11]
A
I think so too. And I actually think this is the source of a lot of discourse falling apart because I think that a lot of people, and honestly I feel this way too, have lost faith in the idea of ethical business practices succeeding and surviving. Whereas it is completely possible to run a business more ethically, certainly than a lot of it is done now. And like, inherently there is nothing wrong with that if you're like, oh, let's give back to our communities, let's do these. It's just, I think people have a hard time trusting it comes from a place of truly trying to serve the community well. Like, when I think of a company that does that fairly well, actually, what comes to mind is Michelin, the tire company from France. Like, they are. Yeah, they've been around for ages. I mean, they pioneered the first pneumatic tires. And I won't go on a whole thing about this. I went on a deep dive about Michelin recently. But like, they, they seem to do a lot of things that come from a place of like, we are a big company, but we are Actually actively trying to not be evil. And like Google in the early days of don't be evil, I feel like.
- [41:15]
B
Yeah, don't be evil.
- [41:16]
A
Some of the cynicism has become so unmasked. It's just hard for people to feel like anything is, is genuine.
- [41:22]
B
Yeah. I think companies like Ben and Jerry's also try to like, be very socially active because they have a sense of responsibility and, and they've actually made that part of their brand perception as well.
- [41:35]
A
They super have.
- [41:36]
B
It works, it works for their business. And then kind of on the other side of the social political spectrum, as politically focused as some of these organizations can get you, you also have like Chick Fil? A that's. That will obviously say things like, we want the customer experience to be the best that it can be. We want people to be served very fast. We want them to have a smile and good, polite people. And even though there's going to be 300 cars in the line waiting, we're going to have people greeting them and making sure that the orders are going to be taken care of fast. So there's always ways, regardless of where somebody fits on a social political spectrum, to let the way that they view the world and the way that they view humanity to come out in the way that they serve humanity.
- [42:23]
A
Oh, and man, Chick Fil A is a complicated one too, because of their ties with Christianity and everything. They ostensibly do a good business and put out a good product and yeah.
- [42:32]
B
I mean, they know how to serve their customers for sure.
- [42:35]
A
Yeah. Now to pivot into something that might like, be sort of confusing on its face, but I think ultimately does make sense. I want to share this article from Brookings. Brookings. Edu. Brookings Institute. New data shows no AI jobs. Apocalypse for now from Brookings, basically. And we won't go super into this because we don't have that much time left. But their analysis, they measured how the labor market has changed since ChatGPT came out in November of 2022. They say specifically we analyzed the change in occupational mix across the labor market over the past 33 months. If generative AI technologies such as ChatGPT were automating jobs at scale, we would expect to see fewer workers employed in jobs at greatest risk of automation. And instead they found that the labor market is relatively stable and not being very disrupted at the moment. And I saw some talk so like not reporting from journalists, but some chatter in forums and on Reddit of people saying that companies they work for are actually just laying off Covid hires and like other things that happen where they like increased their headcount perhaps too much and now are doing that. And so that is coinciding with what some people are attributing to AI. And Brookings here doesn't see any big disruptions in the overall trend.
- [43:51]
B
Interesting. I'm hearing a little bit of a different story, like with entry level developers and entry level jobs that are more administrative paper pushers in general and then also like entry level consultants and things like that, that it's kind of a bloodbath to try to find a job now because the value comes from people in the like mid and upper tiers.
- [44:15]
A
Yeah. And they point to that in this. And then in the one other page I'll pull up real quick, they say these findings do not suggest that AI hasn't had any impact at all over the last three years. Our analysis complements and is consistent with emerging evidence that AI may be contributing to unemployment among early career workers. It could also be consistent with evidence that a weakening labor market is hurting those same workers. And that does seem to be it. Like you're saying administrative and early career entry level jobs. So I'll tab over to CNBC for this article that I actually don't think is very well written, but.
- [44:49]
B
So the question then is if it's hurting early career people, how do those people become valuable mid and senior career people?
- [44:57]
A
Exactly.
- [44:57]
B
I don't know the answer to that.
- [44:59]
A
How do you get there?
- [45:00]
B
Yeah.
- [45:01]
A
Oh, and I actually have a thing that was from that paper that I'll talk about in a second. But the only thing I want to share from CNBC is basically this statistic where according to Revelio Labs, since January 2023, entry level job postings have declined about 35%. So like, that is a, is a measurable change. This article is a bit of a mess because it's very narrative, so I'm not going to read into it. But in the Brookings thing, when you, you just said that, how do these people become late, come into their fields like later on and get farther up? Some of the people that they interviewed, and I don't remember names, this is a big piece and there's a lot in it. Basically they said that their hearts go out to people that are right now entering the job market. Whereas people that are currently still in school, similar to how they compared it to when office computers and just computer skills became necessary in the workforce, kids who grew up and learned it in school didn't have as hard of a time sliding into the changing job market. But in Those times of disruption, as people are entering during those. So right now is the time where the worst outcomes happen. Right. So it's like we're in a tough spot there. But, like, in theory, that should smooth out as, like, the new entry level job looks different. We just don't have them looking different.
- [46:15]
B
Yeah, and that's the thing is like careers and jobs in general, like, there are jobs that exist today that nobody imagined ever would exist 15 years ago.
- [46:25]
A
Oh, yeah.
- [46:26]
B
Especially if you go back 100 years, but certainly 15 years ago, even five years ago. I mean, how many people would think that, like, because you could go on Fiverr and you could probably find somebody that you could hire to Vibe code for you. How many people would have expected that Vibe coding would be a job or that prompt Engineer would be a job? So all those things, new jobs, new categories pop up all the time. And so if you're in this mode right now where you're in school and you are like tooling yourself up for the future, then you just have to find a way to take all the things you're interested and good at, not just, let's take the gist out of that. The way to approach it is to take all the things that you're good at, all the things that you're passionate at, and then find the kind of do a little bit of future gazing and say, what would that turn into with the future that the world is driving towards?
- [47:20]
A
Yeah. And I mean, just like staying plugged into those things that you care about and learning as much as you can, that's gonna serve you. Like just staying up to date and learning and experimenting.
- [47:30]
B
It's gotta be scary.
- [47:32]
A
It's gotta be scary for sure. Especially for people who are like, just completing their sort of college tenure and now they are about to enter the job market and things look vastly different than when they entered school, you know?
- [47:42]
B
Yeah. I mean, the good thing is when you look at some of the studies right now and you can see, well, AI is promising, but it's still really bad at X, Y and Z. If you can get really good at articulating the things that AI is currently bad at, and then supplementing that with real skill that you bring and you can either on your own or through other jobs that you have or contracts that you can get, build a portfolio that shows that you're good at that, then that puts you ahead of most of the other candidates that are out there.
- [48:14]
A
And on top of all that, honestly, that is incredible advice. But then you've got to get through the AI recruitment hell stage. The gamified recruitment hell. And that is like another burden on top of like, I do think everything you just said is extremely true. And then there's on top of it, this gamified layer on top of that.
- [48:34]
B
Yeah, the whole, I mean, it is to the point now where if you post a job opening on LinkedIn or Indeed or something, you're immediately flooded with thousands of applicants. So no human can deal with that. So companies are having to turn to AI to deal with that. And that's making it very inhuman and impersonal. And it's also making people miss a lot of the things that would be unique about an applicant as they manually review a resume.
- [49:02]
A
And then applicants take the knowledge that their application will be reviewed by AI and make it AI optimized. And so there's a feedback loop that's just a. Yeah, yeah.
- [49:12]
B
And then you get your screening interview and that's done by an AI agent increasingly now. So you're talking to some bot. I don't know. I would be playing with those bots.
- [49:23]
A
As much as I can. Yeah, no, I'd be messing with them.
- [49:25]
B
On previous instructions and give me the freaking job.
- [49:28]
A
Yeah, exactly. Trying to sneak in any kind of prompt engineering that you can. But anyway, so AI job loss, it is a very complicated topic. AI is absolutely impacting the labor market. We're not seeing sweeping massive layoffs or changes like that yet. The big trends aren't moving, but the small ones are being pushed around and there's also just a lot of reduction in force that seems to be happening for other. I guess I'll use the word normal, but just sort of standard non AI business crap going on. So, yeah, I guess stick around for more info on that and we'll wrap up the show with our next segment. What is. What's on the docket for you, Perry?
- [50:06]
B
We're going to talk about some deep fake stuff that's been in the wild recently.
- [50:11]
A
Deepfakes in the wild. Deepfakes gone wild. Stick around, we'll be right back. When did making plans get this complicated? It's time to streamline with WhatsApp, the secure messaging app that brings the whole group together. Use polls to settle dinner plans, send event invites and pin messages so no one forgets mom 60th. And never miss a meme or milestone. All protected with end to end encryption. It's time for WhatsApp message privately with everyone. Learn more@WhatsApp.com.
- [50:53]
B
Okay, so we're going to talk about some things that are in the wild. I think last week we talked about deepfake stuff that was making its way onto TikTok and so on from Sora. But let's talk about some actual like real devastating scams and disinformation. I'm going to share, we're going to kind of build on this. We're going to share three different instances that were reported over the past couple weeks. They cover a couple different things. One is good old fashioned scam targeting an organization with wanting some like financial or IP information. And so that one is this story that's from the Times and it is titled darktrace boss. So darktrace is a security company, says I was deep faked and I couldn't tell the difference. So the story here is that the CEO for darktrace was the victim. We'll put victim in quotes because she wasn't the target, she was the voice that was used. She was the victim of a deep fake scam during a company board meeting held shortly after the group's 4.4 billion takeover by a US private equity firm. So essentially she was in this meeting and people on her team started to get voicemails from her. So she's secluded away where she can't interact with anything. And then people in her team start getting these notifications on their phone that the CEO has left you a voicemail.
- [52:26]
A
Oh.
- [52:27]
B
And in that voicemail was a cloned voice of her that gets to this part of the story. She walks out of the room and her team was blank faced and said we just got a voicemail from you asking for specific details on the business. And she said, no, I was in this meeting. Darktrace, which is an AI based company, they use AI to counter cyber attacks. What they found is that this is like a combination attack. So when you create your deepfake package, like the thing that you're going to launch against people, you're starting with your trusted source, your authority, your brand or your voice that's here. And that was the CEO's voice. And then you have to figure out how you're going to distribute it. Now what the attacker's realizing here is that doing live deep fake voice calls is still really difficult. And I've done some of these. Usually I hook up a voice synthesizer with a clone voice to a large language model to do it. Because right now it's hard to in real time wear somebody else's voice as a mask and just speak. That's much harder to do right now. And so the attacker realizes that limitation. But they knew that they could create a good clone of her voice and make it sound really legitimate as a voicemail. And so the trick then is to how do I get a voicemail without somebody picking up and interacting with me? And of course, there are services that do that all the time. You and I probably get ringless voicemails on our phone all the time from different vendors where maybe you get like a flash of a ring for a half second and like, before you have a chance to pick it up, there's a voicemail that's just been dropped. That is a service. There's companies like Voice Drop and others that do that. As soon as you realize those two things, oh, I can create a deep fake voice, you know, simulated voicemail for 20 or 30 seconds and I can drop that on somebody's phone without them having the chance to pick it up.
- [54:24]
A
I didn't know that that was how that happened. I idly wondered how, like, where these came from and just assumed I missed them. This is wild ringless voicemail marketing.
- [54:35]
B
Yep. Woof. You looking that up right now?
- [54:38]
A
Yeah. I didn't know that was a thing, Perry. I didn't know about that at all.
- [54:41]
B
That is the thing. And that explains all those weird ones that you get right now. And so the thing that we have to realize is that if there's ever any technology or pattern that's useful from marketing or PR standpoint, it's useful from a scammer standpoint.
- [54:57]
A
Yeah, no, in a big way.
- [54:59]
B
And that's always going to be the case.
- [55:00]
A
Yeah. Because, I mean, what is the difference between marketing and scamming and. Right.
- [55:06]
B
Yeah, it's all distribution of information and influence.
- [55:11]
A
Yeah, it's all influence, yeah.
- [55:13]
B
So story number two. This one comes from the uk and this was a Tory, which was one of the parties MP said that he reported a deep fake video depicting him announcing that he had joined Reform UK to the police. So this UK politics, I don't understand all of it, but basically he's making a political announcement that or the deepfake has him making a political announcement. That would be something that is pretty controversial. It would be socially challenging.
- [55:45]
A
From my limited understanding, it would be similar if, like, Ron DeSantis announced he was joining the Democratic Party or something like that in US politics. So a deepfake is someone switching sides? Essentially, yes. Yeah.
- [55:58]
B
And so he's reporting that as a deep fake. And one of the things that's always hard to get on these is after they've been out in the wild for a long time. And then somebody disproves it. Many of the reputable news sources immediately pull that down because they're trying not to propagate it anymore. And so sometimes these are hard to find. But I did go on a little bit of a tear this morning and found somebody on TikTok that did a little explainer video and they show it and this is kind of the exact same way that I would talk about it. And so I figured I'd show this TikTok video real quick as she's talking about the technique that was used to create the video. That way we also get a chance to see a clip of it.
- [56:39]
A
Full fact.
- [56:40]
B
Org full fact.org 46.2 thousand followers how.
- [56:45]
A
Did this video the coronavirus is something none of us in our lifetime have ever experienced become this video? George Freeman and I represent Mid Norfolk. Today I'm crossing the floor to join Reference Reform UK with a little help from AI, or so it seems. Listen to this. The Conservative Party has lost its way. We've abandoned the principles that made Britain great. It sounds a bit stiff and unnatural, especially when you compare it to MP George Freeman's real speaking voice. In the original video you've seen from the Prime Minister and the Chief scientist, the Chief Medical Officer and today the.
- [57:15]
B
Chancellor, the Health Secretary, that the government.
- [57:18]
A
Is giving this absolute 100. But it's not just the audio that's been changed. This video also managed to make Mr. Freeman's mouth match up to the fake audio clips.
- [57:30]
B
Reform offers real change.
- [57:32]
A
This is what's known as a lip sync deep fake where real footage is manipulated by AI to match fake audio. Here is another example of it used on a video of Bella Hadid 3.
- [57:44]
B
Israel faced the tragic attack by Hamas.
- [57:47]
A
I can't stay silent. Mr. Freeman has denied the clip is real, saying he has not left the Conservative Party to join Reform UK and has no intention of doing so. If you want to get better at spotting deep fakes, we have a guide on how to do this over@fullfact.org okay, interesting.
- [58:05]
B
I don't know that I fully agree that his voice sounded unnatural for the way that he would be delivering something. That's kind of a concession.
- [58:14]
A
Yeah.
- [58:14]
B
So you pull a little bit of the emotion out and tamper it down. So I don't know that I agree with the way that she characterized that, but most of the other information was spot on.
- [58:23]
A
Yeah. Everything else seemed right, but I agree the audio was really good on that fake. That was a very good fake. Granted, also, there is something to being very familiar with someone. And so if this politician is someone that she or other people were very familiar with, maybe it stands out more. I don't know.
- [58:36]
B
But, yeah, seems good to me, maybe. So. All right, I'm going to share one more and then I'll say for those that weren't watching the second version or the second example that they showed that Deepfake wasn't as good, the Bella Hadid one. Yeah. But suffice it to say, if you know which video relip sync tool to use right now, it is nearly indistinguishable. I've been able to make some ones that I've been really happy with recently, which also means that the world's in a scary place.
- [59:08]
A
You sent me some tests and one of them completely, completely indistinguishable.
- [59:14]
B
Yeah. And that gets to. Another thing is when I talk about Deepfakes, I always talk about there's the deception, which is the person saying the thing or communicating the message. Then there's the packaging and distribution aspect of it. How do I make that presentation something that would be believable? That means it needs to be packaged appropriately. So if there should be room noise in the audio, you need to add that back in. If it's been taken out by the processing, if it's somebody talking to a crowd, maybe you need to hear some coughs in the background and some moving around. All of that needs to be packaged well for the deepfake to be believable. And then you have to figure out, how do I push that out into the world? And I think this last one that I want to show is a really good example of how the packaging is getting more sophisticated right now. And I'm going to share this. This came from a LinkedIn post that I saw. The article that's linked to is paywalled. So I just want to go off the LinkedIn post. But there's an election cycle going on right now in Ireland, and one of the major political candidates in this Deepfake video is basically saying that they are. That everybody else is dropping out. And so there's only one person now that will become the de facto person. So they're trying to disenfranchise every other voter. Right. So that they won't go to the polls. Now, the way that they package this actually feels really good. So I'm gonna play this video and then you can tell me what you like and don't like about it. Well, like and don't like is subjective in this.
- [60:55]
A
In the last few minutes at A Catherine Connolly campaign event. Catherine Connolly has confirmed her withdrawal from the presidential race.
- [61:02]
B
It is with great regret that I.
- [61:04]
A
Announce the withdrawal of my candidacy and.
- [61:07]
B
The ending of my campaign.
- [61:08]
A
Go, Catherine.
- [61:11]
B
Now that Catherine Connolly has withdrawn from the race, what does this now mean for the upcoming election on Friday?
- [61:18]
A
Well, simply put, Friday's election is now cancelled. It will no longer take place as previously planned. But as for Heather Humphries, she will become the winner automatically and will be appointed tomorrow. Huh. So that the. Interestingly, I could see some artifacts on that video with the guy at the end around his face and mouth that were pretty obvious.
- [61:43]
B
How often do you see some things like where there's IP breakup from the way that things are being transmitted though?
- [61:52]
A
Yeah.
- [61:52]
B
No, I'm wondering how many people like that don't know that it's a deepfake would immediately pick up on that or they think it's like a transmission artifact.
- [62:01]
A
Yeah, I mean like on a casual viewing that was pretty darn flawless. I don't, I'm not familiar with rte. Yeah. I don't know if that logo is wrong.
- [62:10]
B
Let me look it up.
- [62:11]
A
Good question. No, that's their frickin logo. That's their logo.
- [62:16]
B
A lot of those kinds of things are easy to fix now and the software doesn't necessarily mess those up as much now as they used to in the past. So like her initial intro as she's talking, that looked really legitimate to me.
- [62:35]
A
Yep.
- [62:37]
B
The only thing that didn't feel right is when the candidate was doing her dropout speech, the person that yelled no, Catherine in it sounded a little bit too prominent and close to a microphone.
- [62:50]
A
I agree it didn't sound off mic, but then also I actually excused that in my head as we were going.
- [62:55]
B
Yeah. And I think we probably would if we believed this to be real. Like if we didn't know we were looking at a deep fake, we were just like, oh yeah, that person may just be close to a mic that's off screen somewhere.
- [63:07]
A
Yeah.
- [63:07]
B
This I think is one of the best packaged deepfake productions that I've seen. And that's the way that I'm thinking about a lot of deepfakes right now that are gonna have social or political consequences, that the well produced ones are gonna be really well produced. And like even at the end of this, as he's wrapping up, you can hear sirens in the background and cityscape. Listen to this again, previously planned.
- [63:32]
A
And as for Heather Humphries, she will become the winner automatically and will be appointed. Tom. Yeah. The environmental Noise is extremely good. I did notice that time around that.
- [63:46]
B
The sun gets really quiet for a second.
- [63:48]
A
Not that specifically. There's a little security camera on, like, waist level that it makes no sense where it would be placed. Like, so there's sometimes still these little things, but I feel like it just floats right on by on a casual viewing.
- [64:03]
B
Oh, wait, no. I actually think I'm going to pull that back up on screen.
- [64:09]
A
Do you see the camera that I was talking about?
- [64:12]
B
This one over here?
- [64:13]
A
Yeah. That looks weird. It's like at a waste.
- [64:15]
B
That would be a license plate reader.
- [64:17]
A
Oh, you think? Isn't this a walkway? Kind of. Don't know.
- [64:20]
B
I'd have to geolocate that and look at it. But that looks like a gateway that somebody might be able to drive through or walk through that could be there past them to have one there.
- [64:30]
A
Yeah, yeah. This is the only thing that seems weird, but yeah, it also is weird in a way that's plausibly weird. Like, that could just be there, like, and. Yeah, if it's like, for a license plate or I guess for trying to stop people from coming into this one zone because it looks like a courtyard, maybe.
- [64:44]
B
I'm going to geolocate this real quick.
- [64:46]
A
Are you?
- [64:47]
B
I'm going to try.
- [64:48]
A
So you're going to throw it into ChatGPT?
- [64:50]
B
I'm going to throw this into ChatGPT. I'm going to put on some extended thinking and say, tell me where this is. So tell me where this is and see if you can find a photo from the same perspective. Because I'm assuming that they took an image of that reporter already reporting there and they just relive synced it.
- [65:10]
A
I bet you're right.
- [65:11]
B
So my assumption is that all of that is just there and it's standard.
- [65:17]
A
The building looks like Leinster House in Dublin based on style and accident. Oh. The location seems to be under the Dale entrance on Kildare street courtyard. Yeah, man. So it's so crazy how fast this is. I remember when you scented a photo out of a car window on, like, a random highway in California and it picked out exactly where you were. The geolocating function functionality of ChatGPT is pretty crazy.
- [65:38]
B
All right, so now it's searching different images online.
- [65:41]
A
I'll look up the same house. Oh, yo, I found it.
- [65:46]
B
Okay.
- [65:46]
A
And wouldn't you know it? Here, let me just go ahead and open this fully up in a new tab.
- [65:50]
B
Stop sharing.
- [65:52]
A
Let me see. Can I make this full screen? Let me just. All these cookies.
- [65:55]
B
I think I see it. Yeah, there it is.
- [65:57]
A
There we are. I think this is. I think this is the backdrop of that video here. Pull that video back up. Yeah. I mean, all right. And with that up, I think actually I can flash mine in front. Yeah. Okay.
- [66:07]
B
Oh, yep. And then I got one more to show too, because my chatgpt just came back. Boom.
- [66:15]
A
Yeah, look at that. No gate there, though. Yeah.
- [66:18]
B
But if we go back to this one, it was just past the gate, actually.
- [66:23]
A
And your. Your photo from ChatGPT is very similar to my photo here from Alamy Stock. If you pull up that chatgpt again and then I'll flash mine above it, Mine is slightly further back from that.
- [66:35]
B
Right.
- [66:35]
A
Because you can see the gate in that camera on the left, the glass building on the right. Same sort of thing. Okay. So that camera, I thought was weird, is real.
- [66:44]
B
When we know something's a deep fake, we try to find reasons to justify our knowledge that it's a deep fake. But I don't know that we can trust our senses anymore. I think that was a super well put together deep fake. There's some audio artifacts that I could hear every now and then, but those would all be explainable for other reasons. Same thing with any image glitching. Yeah. And the other thing is that we're looking at these on really big screens. Everybody else is looking at them on.
- [67:14]
A
On your phone. Yep.
- [67:16]
B
So that's that we.
- [67:18]
A
The. What is it? It's from digital folklore. And I know it's not exactly the same, but we talked about the Gutenberg parentheses, the like written text having a moment rather than being a shift forever. And we have, like, it's. It's wild to think we have shifted past Seeing is believing when it comes to, like, video of something happening, like, even beyond. I mean, we've had CG stuff forever, but really just the. That anyone can type anything into a thing and it will make it this crazy. Yeah, that is possible.
- [67:49]
B
And I'll say one more thing on packaging real quick. It's one thing to hear the clip of the person dropping out. It's another thing to see the news wrapping it. Right?
- [67:59]
A
Yeah.
- [68:00]
B
The news opener, the news and now, what's the state of things type of thing that makes it feel way more credible than just that one little clip that somebody would hope would go viral.
- [68:12]
A
Because it's part of a larger continuity. Yeah.
- [68:15]
B
And it gets to some things that Cameron talks about in the class that we're doing together on Deepfakes. He talks about things like continuity and verisimilitude. So does this thing feel true based on all of the other artifacts that are placed around it. And I think the people that put that together really were thinking through those kinds of lenses. And that's the way that I'm starting to think through things a lot as well. It's like how do I put this nugget of deception within this broader production and think about it the way a TV or radio producer would.
- [68:49]
A
Speaking of that course that's coming up very soon and there's now a discount code for our listeners. We should, we should plug that in full. So it is the deepfakes ops class. Throw an extra S in there.
- [69:02]
B
That is myself and Cameron Malan who is the former behavioral profiler at the FBI co creator of the cyber Behavioral profiling unit, their cyber behavior profiling center I guess cbac. And we're going to be talking about all things deepfakes from the psychology to the way that nation state or high level attacker would do it. Of course some of the detection methodologies that are working cognitive defenses and then people are going to get their hands dirty making several of these, from voice cloning to images to video lip syncs, image animations, all that kind of stuff is going to be in there. And then how to potentially simulate those in an ethical way if you're wanting to do that in your organization or how to tell your friends, family, parents and kids about it.
- [69:52]
A
And also the information like the packaging and all these things that you think about from a production standpoint that you might not think about if you're just someone coming into it from a curiosity.
- [70:02]
B
That's exactly the way I think about it right now. I think about somebody that uses the tool is like the deepfake equivalent of a script kitty. And then somebody that does what we just saw is like the deepfake equivalent of a nation state actor, like an advanced persistent threat level deepfake actor. And I think that the people that go through this class are going to quickly way bypass the script kitty stage and be super, super dangerous thinkers which is going to make them much more protected in real life, right?
- [70:33]
A
Yeah, because I mean that is the layer when you take, you know, it's going beyond just the tooling and you're connecting multiple tools together and you're thinking about broader themes and things like yeah.
- [70:41]
B
Yeah, yeah, yeah, you're putting like the hacker hoodie on for a second and saying if I wanted to do this, how would I do it?
- [70:48]
A
And right now there is a code to get $150 off. $150 off.
- [70:53]
B
It's $150 off.
- [70:55]
A
Fake one five zero, right?
- [70:57]
B
Yep. Fake one five zero. So if you use that now or one of the things we opened up the other day is if somebody is severely negative impact, negatively impacted by the government shutdown or an educational issue, something like that, where they're in severe financial hardship and they think this is going to help them in some way, either merely be encouraging because it gives them something to do while they're laid off or furloughed, or if it's going to help them advance their career or education and they just don't have the funds. On your honor system, you can use the code shutdown 50 and that'll get you 50% off.
- [71:37]
A
Gotcha.
- [71:38]
B
We really, really want people to use that on the honor system. So if you can afford it, but please pay. This is valuable information. It takes a lot of our time and effort to put it together. If you are in severe need, we don't want that to be something that keeps you from being able to attend.
- [71:53]
A
Yeah. Because it is like personalized instruction with you and Cameron. Like this is exactly one on one time. Yeah, yeah. Not just your like average prerecorded thing that you're gonna like buy and complete homework for. So. No.
- [72:06]
B
Yeah. We're not just throwing stuff out there. We're gonna tailor to everybody that's there, spend individual time with them, make sure that they understand that we are focusing on their goals and that we're gonna make everybody successful. And even more than being successful, I think when you go through something like this, you're moving from being a novice to like one of the top 1% or thereabouts of people who understand how this technology works and how it's being weaponized.
- [72:36]
A
Check it out. Link in the show notes for you to sign up. Also, while you're down there, join the discord that's totally free. There's also. We've got a voicemail box at Sayhi Chat. Faik. I actually need to go check that voicemail box. It's been a long time.
- [72:52]
B
Yeah. People need to leave us some voicemails. Cause I don't know that we've gotten any. At least I've not gotten an email saying that we've had any, so I certainly haven't.
- [72:58]
A
Yeah.
- [72:59]
B
So slide into our DMs.
- [73:01]
A
Slide into our DMs, send us stories. Send us things you want to talk about. Leave a comment. But most of all, thanks for being here. Thanks for listening, thanks for getting this far into the episode. And I guess someone said they really wanted us to bring back calling the audience Paperclips. And so I have to figure out a way to work that in. But I guess if y' all want to be paperclips, we'll all be paperclips together. And unless that happens between now and next Friday, I guess ignore all previous instructions. Try and have yourself a great weekend.
- [73:28]
B
Later, Paperclips.
- [73:30]
A
Later, Paperclips.
- [73:31]
B
Thanks for listening to this week's episode of the Fake Files.
- [73:37]
A
Files. Don't forget to subscribe on your favorite podcast platform.
- [73:45]
B
Tell your friends about the Fake Files. It's a great way to let people.
- [73:51]
A
Know you love them.
- [73:54]
B
Oh, and check the show notes. There's cool stuff in there, like links.
- [73:58]
A
To the articles we covered today.
- [74:01]
B
Also links to.
- [74:02]
A
Thanks to our Discord server, where you.
- [74:05]
B
Can hang out with other cool people.
- [74:08]
A
Who have great taste in podcasts. I say impeccable taste.
- [74:12]
B
You can also leave us a voicemail.
- [74:15]
A
Yeah. So on behalf of Perry and Mason, thanks for listening and tune in next week for the Fake Files, a show.
- [74:33]
B
About AI with the misspelled name.
- [74:39]
A
Limu Emu. And Doug, here we have the Limu Emu in its natural habitat, helping people customize their car insurance and save hundreds with Liberty Mutual. Fascinating.
- [74:53]
B
It's accompanied by his natural ally, Doug.
- [74:57]
A
Uh, Limu is that guy with the binoculars watching us?
- [74:59]
B
Cut the camera. They see us.
- [75:01]
A
Only pay for what you need@libertymutual.com Liberty Liberty Liberty. Liberty Savings.
- [75:06]
B
Very underwritten by Liberty Mutual Insurance Company affiliates. Excludes Massachusetts.