#330 AI or Human: Unraveling Content Authenticity with Jon Gillham

[00:00:00]

Mehmet: Hello and welcome back to a new episode of the CTO show with Mehmet. Today I'm very pleased joining me Jon Gillham who's from Originality.Ai. Jon, the way I love to do it is I keep it to my guests to introduce themselves because I believe [00:01:00] No one can introduce someone else better than themselves. So tell us a little bit about you and what you're currently up to.

Jon: Yeah, sounds good. Yeah. Happy to be here. Thanks for having me, Mehmet. Um, yeah, so my background, um, mechanical engineer worked in oil and gas industry, built up, um, some online businesses, all focused around content marketing, um, where producing content, publishing on Google, getting traffic. Um, and then that led to starting an agency, which then saw sort of the wave of generative AI coming and, uh, we built, and we were struggling to answer the question when, uh, customers were asking us, um, you know, was this AI created or human created and how are you controlling for that?

Jon: Uh, and that was when we decided to build originality. ai. Which predated, um, chat GPT. We launched on the Friday of, uh, and chat GPT came out on basically the Monday. So then, uh, you know, year and a half since, [00:02:00] and been, uh, been an interesting ride.

Mehmet: Indeed. It's an interesting ride because, you know, since chat GPT was released, like I can say it took the world by a storm, uh, especially in the content space, Jon.

Mehmet: You mentioned like what were the main challenges for people to know like, uh, They want to know if this content is generated by AI and or is generated by I mean, is it written by by actual humans? so So what do you think the real challenge behind the content that is generated by AI? Like, is it like, is it because of the authenticity?

Mehmet: Is it because it's affecting some other areas? Why people are, you know, interested to detect if the content is generated by AI or it's a human?

Jon: Yeah, no, it's a, it's a, it's kind of a key [00:03:00] question. It's a good question. I mean, I mean, there's kind of two, There's two, two reasons to it. Um, the one reason is sort of, there's a fairness component to that.

Jon: You know, if somebody is happy to pay a writer a hundred dollars or a thousand dollars for an article, they're not super happy when that article got copied and pasted out of chat GPT in, in five minutes. Um, so there's sort of that, like that's part one, sort of the fairness piece. Um, but then the second part is, uh, the world, the way the world is viewing AI content is, is evolving.

Jon: Um, and. There's risk associated with it. There's reputational risk associated with it. If you publish factually inaccurate information and then Google, um, Is facing an existential threat if their entire search and as a result of that existential threat, um, are taking steps to fight a spam. And so, you know, it's fairly simple to think about.

Jon: Whereas if Google search results are filled with nothing but generated content, then people [00:04:00] wouldn't go to Google. They would just simply go to the that generated that result. And get a far more personalized answer to whatever that searcher's query was. And so if, if Google result, if the end result is that Google is totally overflown with AI generated content, um, I think it's really an existential threat for them and they are, they are fighting and fighting back, uh, strongly against AI spam.

Jon: And so there's sort of those two pieces. There's the, um, the fairness component. I'm happy to pay a hundred dollars, a thousand dollars, whatever the article cost is, um, but don't want the writer to have just copied and pasted it out of chat GPT. And then there's a second part where publishing AI content, uh, introduces risks that, uh, The publishers should be the ones that are making the decision on if they're going to accept those risks, reputational risk, Google risk, et cetera.

Mehmet: But Jon, like, and correct me if I might be wrong, but even before, I mean, so just for the audience to know before even chat GPT, there were some tools that were using the [00:05:00] chat GPT API to generate the content. There's plenty of tools out there. Um, but even before that, even before that era, uh, I remember I used to receive these spam emails mentioning, Hey, like I have bunch of articles that if you just put them in your blog, you will get the SEO.

Mehmet: And you know, when I was still new to all this, I was reading this article and I found that they also jumped. So is the AI just like amplified this problem or is it like It's creating a completely a new, I would say, set of problem that even the previous issues they were creating before.

Jon: Yeah. So, so yeah, it's, it's, it's, you know, the pre, pre chat GPT generative AI existed, right?

Jon: I think there's a lot of people, I mean, I'm sure most of your audience in the show would be well aware that like open AI had, had APIs that you could tap into with GPT 3 and GPT 4. Chats GPT was sort of a, [00:06:00] a very evolved wrapper on, on its GPT 3, its own GPT 3 API. Um, and so yeah, this problem existed pre Chats GPT, Chats GPT just exploded the problem, um, made it significantly larger.

Jon: I think we're kind of the, you know, what's there's a, there's a funny, funny kind of graph that shows the like progression of AI relative to, um, human intelligence where it's like, not that long ago, it's kind of like a cute, the monkey can do, uh, or the, you know, the robot can do monkey tricks. Um, and then now it's like.

Jon: You know, skyrocket in terms of capabilities. And I think we, we sort of went through that phase, um, at the tail end of sort of like what writing tools are capable of producing because pre pre chat GPT. And just slightly before that with some other tools like Jasper, the content. Um, that predated those really wasn't worthwhile.

Jon: You know, it's kind of similar to the stuff that you were getting in those like big bulk [00:07:00] packages of like, Hey, here's a bunch of content. It's like, it was just not good. And Google was really effective at identifying bad content and punishing it. Um, what has changed is that this sort of auto generated.

Jon: What Google views as spam content reads really well is, is really well written good content that is harder to tell the difference between now sort of spam generate content that is grammatically well written, um, compared to compared to what predated this, this problem.

Mehmet: Absolutely. And there's another aspect, Jon, here, which is.

Mehmet: Some people, they say, okay, like part of the problem is not only Google, like Google and other search engines who hardened their search search engine optimization in a way that forced people to, you know, style their writing in a [00:08:00] way and, you know, focusing on the keywords. And so some people actually, they blame.

Mehmet: Uh, the search engine companies and mainly Google, of course, and a little bit Bing or Microsoft for having created this. And now they're saying, okay, like ChatGPT is not only for, I mean, writing blogs, you can do it, you can do with it multiple things. So they said, okay, why we don't leverage it to just like, Make google happy.

Mehmet: So they are actually blaming the search engine companies that it's your fault that you forced us to do this Do you agree with that?

Jon: I think there's certainly it's I mean, it's a hard it's a hard problem to solve. I mean whatever there's such an economic incentive to be able to um, Win on google that that people will always try and manipulate it.

Jon: Um, And so I think there's certainly, you know, Google will always do things that are not the, what they do and what they say will not always be 100 percent aligned. Um, and what they are trying to do [00:09:00] and versus what they, they say they're doing is also not always aligned. Um, so I, I do think. Google was constantly be dialing, turning, turning the knobs, the knobs on their, on their algorithm.

Jon: And then SEOs who are producing content will always be trying to optimize against that. And I think there will be times where the, the effects of that will be like, you know, the B, if you right now, or what has ranked for a long time, we wouldn't like call it the world of recipes, was like a 2, 000 word story about their life before they got into the recipe.

Jon: Um, and that's not what people want. If we wanted to get into the recipe, that's what they were there for. And so I think I, but that was not what Google was rewarding. And so people followed what Google was rewarding. And so I think, you know, I don't think there was a, um, I don't think there was a malicious intent from Google to sort of incentivize behavior that they didn't want.

Jon: Um, but I think the, the result is that, uh, Google tries to. Deliver what people want. Um, [00:10:00] and people will always try to, um, optimize for what is currently working

Mehmet: fine. Now Google themselves. They are into the eye game as well. So What is now Gemini used to be called BARD. So they have it. So, and I'm using Gemini sometimes because if I want something which is like recent, like I don't want to rely on the knowledge base in the large language model.

Mehmet: So do you think that they can tackle this problem with Gemini? So maybe Gemini will become the interface of, The Google search as we know it, because, you know, uh, perplexity, which is another startup, actually, they are trying to do this. Uh, so are you seeing, you know, from search slash AI perspective, we're going to go to see this mix because honestly, I also, I, one of my guests, I, some of them, I guess I asked them, did your behavior change after chat GPT?

Mehmet: They said, yes, [00:11:00] we, we go less to Google now and we go to chat GPT. So how do you see that they are trying to. Tackle this, uh, the, the AI from, yeah, it's,

Jon: it's tough. So I think incredibly hard situation for them. And I think they are well aware of that. They, they don't want to be a legacy. You know, the, they don't want to be the codec that let digital cameras sort of pass them, pass them by.

Jon: So they need to be AI forward, which they are certainly attempting to be. Um, but then they can't just totally crush their cash cow and legacy business around search and display ads. Um, and so trying to balance that is, It's probably going to be one of the, you know, can they successfully navigate that will, will be a defining moment for them.

Jon: Um, so to answer the question, you know, do it, what do I personally think will happen? I don't know. Um, it is, I think, I think there's no scenario where AI. Enriched AI search will not be the future. Do I think it will be sort of the SGE, like the, the summary, um, [00:12:00] at the top? I'm not sure. I think most people that are using that, like to sort of blow past that and to get to the results if they're in, if they're in Google, they're trying to like find the best answer from the, from a site, um, I think, uh, I think it will be some kind of hybrid, um, I think I can't imagine where it's not some kind of hybrid where it's, um, AI summarized AI information that is customized to, to you.

Jon: And then the, the source, you know, I think perplexity is taking a really interesting approach that, that I, that I appreciate and I think will be useful. Um, and then I think, you know, chat GPT could, could easily sort of integrate, not easily, but it could, it could also integrate in, um, you know, an increased number of links and sourcing and.

Jon: And maybe that becomes the way that people access the world's information.

Mehmet: Yeah, fine. I agree with you on this, Jon. Now back to the detection part. So personally, even [00:13:00] I tried, I am a transparent person. People knows, you know, if they follow the show, I always share what I do. Now, sometimes it's very LinkedIn or other social media platform.

Mehmet: Now everyone knows if this piece of content was written by. AI mainly chat gpt because people after one year I would say or maybe one year and couple of months now They know how chat gpt will write a piece of content to you, you know from the styling from the wording But i've seen something interesting which personally I tried so basically a couple of times what I tried to do is to do Or to use Let's say Gemini and ChatGPT and I start to do this mix and the final one I look at it You know and it doesn't look like an AI Written a thing.

Mehmet: So is this still something that you think that it's considered as a spam? Is it like but when we think about it from a you know what [00:14:00] the work that you do now to make you know sure that there's integrity and there is also like You know, protection for the publishers. Do you think this will make things harder to detect in the future?

Jon: So, um, I think there's a question, you know, when GPT 10 comes out, well, will we still be effective at detecting it? Um, hard to bet against the progress of AI. Um, the flip side of that is that every new model that has come out, we have been, our gap to detection capabilities has been smaller than the previous model.

Jon: So our detector has been getting more effective at identifying AI. Okay. Then detectors have been, or then writers, um, being created that become undetectable. Now, I think there's a couple of reasons for that. Um, the sort of foundational technology transformers, the hardware that they're being trained on, the, um, the data sets that are being used all have a share, a lot of similarities in their training.

Jon: And so the. [00:15:00] Um, differences in their styles from, you know, cloud three to GPT four turbo to grok to mix row, you know, to any of these models, um, aren't, aren't aren't too different from one another from a, from a capability detection standpoint. And so, you know, the, the mixing example that you provided, uh, where, you know, some has written in, in GPT four.

Jon: Some is written in, in Gemini. Um, they share some common similarities that are able to be identified by detectors. Um, and so, you know, I think there will always be an adversarial component to trying to bypass detectors. And we have a, you know, a red team that's constantly working on, on attempting to, uh, to beat our own detector.

Jon: using strategies like mixing. Um, and so that's, that's, that's the aim of what we're, I guess, that's how we're thinking about the problem of, you know, will, how will the world [00:16:00] evolve from a undetectability and these adversarial attempts?

Mehmet: Absolutely. Now, one thing also about the challenge of detection, OpenAI themselves, they had a tool to detect and they had to shut it down because they figured out that it's not accurate.

Mehmet: Uh, and it's a hard thing. Now, one time and correct me if I'm wrong, Jon, because I want to a little bit, I have nothing too deep technical, but one time. What I understood that if OpenAI, Google, anyone, uh, Anthropaic, like all these companies that they have these models. So each time you ask for something to be written.

Mehmet: So it has a token and this token actually allows you to track it later to see if it was, you know, generated by this model or not, but as long as, and because there's privacy, Part in it. So it makes hard for these companies to let everyone, you know, know [00:17:00] that this is generated by AI or not Do you agree with me that also the privacy part of you know, these companies that they cannot disclose With the rest of the world like who did what on our platform make the situation of detection More hard and why do you think even themselves, they fail to do that to, to, to even detect that that's written by AI or not?

Jon: Yeah. So, so their, their model, um, they were in a really tricky spot because they're viewed as the authority on, um, if, if they have a detector and that detector says that something was AI, it's really, really challenging for somebody to say that, no, that was a false positive, because this is actually the company that, that produced the content.

Jon: Yeah. Um, and so detectors are classifiers and sort of like, to name them correctly, um, are able to provide a probability on whether or not a piece of [00:18:00] content was AI generated or human generated, similar to kind of a weather prediction. It's AI that will sometimes get it right, sometimes get it wrong, most of the time get it right, sometimes get it wrong.

Jon: Um, and you have to tune it. So whether you're sort of focusing on detection or focusing on sort of catching the AI or focusing on making sure you call human content, human content, and so you're always needing to sort of like tune between those two settings in open AI's case, because they get viewed as they're the authority.

Jon: How could they get it wrong? It's they're the ones that actually created the content. Their classifier was very, very, very heavily tuned to providing as low a false positive count as possible, but they still couldn't get it to zero because nobody can with a, with a classifier, get it to perform absolutely perfectly across a giant, you know, giant unseen data set.

Jon: Um, And they had it so tuned to reducing false positives, but still couldn't get to zero that their [00:19:00] accuracy rate was crap. Um, and so they had to, they were for, they, they had some logical constraints based on being viewed as the authority to set up their detector and their classifier in the way that they did, uh, which is why they had to shut it down.

Jon: Um, because it was still produced false positives and was pretty useless at detecting AI content. So that, that was their situation. Um, and then in terms of. Uh, so this is the second part of the question was, um, um, sorry, uh, yeah, what was, what was the second part of the question?

Mehmet: The second part of the question was like about, like, how the problem can be solved, actually, like, I mean, the challenge of actually really, detecting this because of the privacy part as well.

Mehmet: That's what I was asking.

Jon: Yeah, the privacy part. So, yeah, so, so I think watermarks, watermarks were sort of held up as the sort of, um, [00:20:00] initial hope around like there's significant societal consequences to when you're no longer able to tell the difference between what's human and what's, what's AI. Um, you know, mass, mass propaganda, um, significant spam.

Jon: And so the hope was that watermarking, um, that every LLM would be able to produce a, um, a cryptographic, um, uh, cryptographic communication around what, what, um, has, has, has what they've created and what, what other LLMs have created, um, I think that was possible when in sort of a monopolistic world where.

Jon: All content was being created by JGPT in a world where we have open source models. Um, and there will be simple adversarial strategies to remove watermarking. Um, I think it's become pretty clear that sort of. Absolute provability, um, that this content was created by an LLM through the use of watermarking will not be an effective [00:21:00] means of detection and solving sort of the societal problems with the not being able to tell the difference between human and AI content.

Mehmet: So this reminds me too much of the, you know, cat and mouse game that we talk about, you know, in, in, for example, cybersecurity, where you have the hackers and then you have the defenses, and then people will try to remove, like same thing applies even to the deep fake. Now it applies to the voice as well.

Mehmet: Now, one part, which very interesting is that I know Jon, you and your team, you have done some, some studies about the reviews. Um, you know, on different platforms, like Amazon reviews, like, uh, different other platforms, like software review platforms. Tell me about, you know, what actually these reviews revealed to you and, uh, What are like something that, you know, you think it's really, we should stop and think about it.[00:22:00]

Jon: Yeah. Yeah. No reviews. Reviews has been an interesting one. So, I mean, we sort of had the hypothesis that, you know, since, since the release of chat GPT, how many reviews, um, were, were AI generated or not. And so we went and, and scraped a bunch of reviews across a bunch of different platforms, um, and then ran them through a detector to look at.

Jon: The, um, sort of percent of reviews that were AI generated, um, kind of all of our studies start off with sort of like a two and a half percent, um, probability of the, of the reviews being AI generated. Um, that sort of falls in line with our false positive rate, um, of, of what we, what our model detects. Um, and then what we've seen is that we've seen some sites get up to as high as 30 percent of their reviews are AI generated.

Jon: And these, you know, popular, popular go to review sites, um, and it really raises a question that we all need to continue to think about in terms of like, where do we want a content? And where do we not want [00:23:00] it? And I think. Personally, I want, if I'm reading review, I want that review. I want to know that review came from a human.

Jon: Um, and I don't want to have to feel like I'm doing a Turing test every time I read a review and trying to tell, am I interacting with the, with the machine or interacting with the, with the human? Um, and so, yeah, no, it's been, it's been interesting and, and. Sort of continuation of the question without a obvious answer.

Jon: I think it's 1 of those things that society is going to have to continue to wrestle with on how it on on where it's sort of. Ethically, morally, potentially, and ending up being legally, um, allowed to use AI generated content and where, where do you use it need to actually be the, that human's voice, or does it move to a world where we put a lot more, um, the, you know, talk to the privacy where, where sort of anonymity will be, um, reduced because you want to have trust that the person that wrote this, you can connect [00:24:00] back to them and, and they're standing behind those words, um, so no, it's been a.

Jon: It's raised an interesting question on, on how, how we want as a, as society to, to handle, um, AI versus human content. And I think reviews is this place where I, you know, I would rather read, read a review from a human.

Mehmet: Yeah. And I know that Jon, you have an opinion about, you know, like how, of course, in maybe a metaphoric way, uh, AI is, is actually unintentionally doing harm to, to people.

Mehmet: Um, and maybe you share with me the story about. You know, the foraging books, you know? Yes. Right. And how do you believe, you know, after you tell us this, how, how we can address this from ethical consideration?

Jon: Yeah, no doubt. So, yeah, we, we worked with, uh, the Garrett New York times and then worked with Guardian to look at some books that were on, on Amazon.

Jon: Um, one of the books that we're identified as being AI generated, [00:25:00] um, It was a book on foraging for mushrooms, and as, you know, LLMs are prone to do whoever had been using that LLM had not proofread the book clearly well enough, um, and it had suggested that to determine if mushroom was safe to taste it, um, you know, which could obviously lead to, to death, um, and incredible, incredible, you know, harm to that individual.

Jon: Um, and so that was sort of one of, like, one of those times where it's like, this is, it can often feel a little bit theoretical about the, when we're talking about the, the harm that, that AI can produce. And just to be clear, like, I'm, I, I use AI all the time. I think it's awesome. I think there's just a, we need to sort of.

Jon: Understand when it's being used and when it's not being used. And so since the second part, um, you know, what do I think the world's going to look like, uh, and how are we going to sort of ethically handle? Um, when we don't know if content has been AI generated, human generated, I think it's going to come [00:26:00] down to a lot more, um, a lot more responsibility being put on the, um, person, putting their name behind the content.

Jon: Um, you know, with when, you know, if you host a podcast with an AI trained voice, um, if, if you put your name behind that, um, in the end, it might not matter as, as long as sort of you're not trading your, Um, trust with your audience, um, if it's, if it's great and entertaining and useful and educational, um, that could be great.

Jon: Um, but I think it's going to, the world is going to come down to the author behind pieces of content, having a lot more responsibility when they hit publish.

Mehmet: Absolutely. And the last thing I would think about is to replace myself with an AI avatar. I would not do this. I would prefer, even if I do mistakes, even if I have some technical difficulties, I would prefer to still do this, this way.

Mehmet: [00:27:00] Now, uh, Jon, tell me a little bit, you know, because I know you've built something really special and You know, you're trying to differentiate yourself by some of, you know, the technologies and some of the methodologies that you are using. So if you can like a little bit, a little bit shed some light, you know, on, on what make originality different than, you know, maybe the other tools out there and, you know, why you think it really can make some impact on getting back to what we were talking first.

Mehmet: Stop people being getting hurt by I. And of course, maybe help also in in in in detecting when it's appropriate, appropriate not to use a I

Jon: yeah, no sense. Yeah, we've got. So I think the first thing that we do that is a little bit different is that we're built exclusively for copy editors. So people that are [00:28:00] Writing content that are published in the content on the web.

Jon: Um, we don't like to be our tool being used within academia. We're not trained on academic tool set, uh, data sets. We're trained on sort of web publishing, uh, data sets. So that's sort of, that's who we're building for. And then as a result of that, um, we have some unique features that other detectors don't have.

Jon: One. Um, our efficacy level, our, our accuracy rate, true positive rate is higher, um, than, than other tools and we go to some pretty extreme steps to provide transparency, including all of our data sets that we use to sort of. Back up our claims around efficacy, both data sets that we've created, and then a bunch of data sets that we'll, we'll source from other locations and then publish all of that, that research.

Jon: Um, so that's, that's kind of the big thing around efficacy and we've added some features that are, um, unique. So we will, both positives do happen. They suck. Um, but we've provided a Chrome extension that allows [00:29:00] people to visualize the creation process of a Google document. And that really helps, um, if a false positive does happen, allows writers to show their work and show their creation process to be able to say, you know, this was a, this was a false positive does happen, you know, two and a half percent of the time, um, and, and it gives them the ability to, to sort of communicate that.

Jon: And then last sort of feature that we've, um, released is a fact checking tool. Um, you know, whether or not people are happy to publish AI generated content. There's logical reasons to logical reasons not to, I think no one says I want to publish factually inaccurate information. And so that was another feature that we added.

Jon: Um, it's still in beta. It's not as good as we want it better than any sort of out of the box LLM with the way that we do it with some, some reg. Um, but it's, uh, it's still not, it's still not where we want it.

Mehmet: That's, you know, good to know. And, you know, cool features also as well. And I believe, you know, [00:30:00] in, in, um, The necessity, especially because, you know, especially the example you gave about the, uh, mushroom forging and other, you know, we, we need, yeah, we, we need someone.

Mehmet: And honestly, we need to rely on AI to do this, uh, because we will not be able to go over all the content that is spread all around, uh, the internet now. And there is also some other aspect, Jon, and I'm not sure how this will be tackled. Now we're using the AI tools, CGPT and others for now, quite some time, as we were saying, and we know that usually they train their data based on what they find on the Internet.

Mehmet: So and to the term that majority of my Uh, guest used like junk in junk out. So we're gonna, do you think we're gonna have like more junk, you know, coming out because of what happened during all this, uh, [00:31:00] couple of months that we've been using now? Yeah,

Jon: it's kind of a, it's, it's for the rest of humanity, the only known, the only really known trusted human created data sets have already been created.

Jon: Um, you know, it's, it's impossible to find a piece of content online and say with certainty, was this AI generated or human generated? Um, so it's a pretty sort of incredible thing to think about that for the rest of rest of humanity. It's really, um, the largest human known, known human. English language created data sets have already been created.

Jon: Um, and then that leads to what will models train on in the future? Um, will it be synthetic data? Will it be, you know, the garbage in garbage out, um, the snake that eats its own tail, um, analogy will, will synthetic data eating its own tail lead to model collapse? I don't, I don't think so is, is, as what I would say, isn't that.

Jon: [00:32:00] I think that is a known problem, and there are some strategies that at some of these sort of foundational model companies to, um, address that concern. I think there's plenty of examples of if that concern isn't addressed, where the output of these LLMs continues to trend towards the middle. So we looked at, um, we ran another study where we looked at.

Jon: Um, the sentiment on a bunch of human written created pieces of content and then scored those from sort of highly negative to highly positive asked chat, and some other models to rewrite them seeing and what we're trying to look at was saying, which foundational model is sort of tends towards positivity, which tends towards negativity.

Jon: They all just converge in the middle, um, just sort of vanilla, vanilla content. Um, Without a significant bias towards positivity or negativity. Um, and so I think if we do run the risk around. [00:33:00] Model collapse or snake eating its own tail. If it keeps getting trained on, on its own output, that will just end up with a whole bunch of, uh, sort of average content across the board.

Jon: Always.

Mehmet: We're going to see what the future is hiding to us in, in, in, in that domain, just in route of. Maybe a little bit sense of humor because you said like about English. So maybe humans will need to invent new language. That's still like, uh, these AI models, they cannot understand. And then you can judge if it is, uh, it's AI generated or not.

Mehmet: Um, like I speak multiple languages and I can say, for example, when I tried it with something other than English, It's still, it's still very obvious that it's, you know, like you, even you don't need something else, but when it comes to English, yeah, like for me, even sometime after I use it, I say it doesn't look really like I was just giving an example.

Mehmet: And of course, I believe like with time, when you start, and this [00:34:00] is something ChargPT, they do. So they start the way, I mean, OpenAI through ChargPT, they understand, you know, your, interaction with the tool and then with time, of course, they are training the model on your own data as well. So it start to understand more and start, you know, to have the tone that usually I provided, which is for me, it's not a problem, but yeah, like if someone takes this and maybe he, he, or she quoted me, I would not know that.

Mehmet: Ah, really? Did I say this? You know, like this is this is really a challenge, I would say. Um, so what do you think, Jon? You know, the vision for the future for the Internet to maintain its integrity and transparency with all the things that are going on. So will it be like a collaborative effort? Do you think that we need to put some regulations?

Mehmet: Transcribed People usually [00:35:00] they are going against regulations because they said regulation might also stop, you know, some other useful use cases of chat GPT and these other AI tools. So what do you think, you know, the future of the Internet would be? Uh, to maintain it in with the integrity and transparency as I was asking.

Jon: Yeah. So I think I think regulations may come in the form of other, um, forms of media than text. I think, um, societal harm associated with, uh, Written word is, is, is, is not insignificant, but it's not the same as, um, uh, as images or video. And so I, I, you know, I think this lines up with sort of the direction from like the White House letter with, with a bunch of open AI or a bunch of AI companies that I think there will be, um, watermarking attempts or some [00:36:00] regulation around attempting to understand, um, if images, video, potentially audio was generated.

Jon: I don't think text is going to fall into that camp because one, I don't think it's, I think it's harder to, to watermark than, than the others. Um, And so I think, you know, where's the internet going to end up? I don't know exactly what it's going to look like, but I think there's going to be a lot more, uh, responsibility and trust put on who the author is, um, behind a piece of content, whether it be a podcast, a video, um, written text.

Jon: I think that's going to ultimately be the only. Um, the only path kind of through this that maintains integrity where you know, the author, and then there's some amount of sort of, uh, proof that that was the author, the author is who they said they were, um, uh, you know, I think that's ultimately what's going to, uh, where the internet's going to land on how it manages, um, [00:37:00] to, um, continue to be trusted, um, in, in a world where there's AI and human content that's so easily created.

Mehmet: Interesting times ahead. For me, another concern to be honest with you, Jon, is besides also like, you know, the content and I mean, of course the, the harm that the content can claim is that someone who And with all due respect, of course, who's not from a domain, all of a sudden, they can come out and say, Hey, see, I have a book, I have this much of reviews about this, and then they become an authority.

Mehmet: And maybe they would not use AI in the future. But AI allowed them to claim something which is not true. Like for me, it's a concern, because, and actually, of course, there are some other factors. And this is why, you know, at some stage, when you were talking about the reviews, and all this, We have some other factors which people, uh, they don't think [00:38:00] about much.

Mehmet: What social media and mainly, you know, and again, no offense to anyone. I use them to promote sometime, you know, some, some reels and, you know, and some short videos, you know, all what's happened on Instagram and TikTok and, you know, these social media platforms allowed people to think, Hey, I can, you know, get hundreds of thousands of followers if I just claim, Hey, I'm an author, right?

Mehmet: I have the bestseller on Amazon or whatever. I have like hundreds of reviews and see what I can do. And then they will get a continued. You know, doing what they do usually there. So for me, this is a big concern in addition to the other concerns that you brought up, Jon. So just, you know, as an, as my thought on this topic now, as we close to an end, Jon, like final thoughts from you, any final things you want to leave us with today for the audience and where people can find [00:39:00] more about you and about originals.

Jon: Yeah, no, I've enjoyed the, enjoy the talk. Um, you know, I think the, uh, I think there's a lot of interest and it's so new that there's a lot of sort of, um, potential misunderstanding in the space around AI content creation, AI detection. Um, so we've tried to make a really in depth guide around understanding efficacy and where it's effective.

Jon: So if sort of anyone's interested in understanding more about. Um, uh, A. I. Detection. We've got a post called originally the A. I. And it's that A. I. A. I. Detection accuracy. Um, and you can find online, but that sort of dives deep into nerding out about, um, sort of the limitations of when these tools are useful and not useful.

Jon: Um, but no, find me online. Um, Jon at originality dot A. I. Is my email. Happy to. Talk to talk to anyone.

Mehmet: Great. Thank you very much, Jon. And I really appreciate the time today. And, you know, it's an important topic, indeed, an important topic [00:40:00] because it has a social impact. Um, you know, maybe people, they don't think about it and thank you for coming today and sharing, you know, uh, These insights with us, especially the research that you and your team you have done And you know the consequences that you know can affect on our life So again, thank you for the time and this is for the audience If you just discovered this podcast by luck, thank you for passing by.

Mehmet: I hope you enjoyed it If you did so, please don't forget to subscribe. We are available on all the podcasting platforms And we are available on youtube also as well so please subscribe share the podcast with your friends and colleagues and if you are one of the loyal followers that keep coming and keep sending me their messages and recommendations.

Mehmet: Thank you for doing this. Keep doing that. Thank you very much. And if you are interested to be on the show, don't hesitate to reach out to me. If you have an important topic, such the one that we discussed today with Jon, don't hesitate to reach out to me. I'm always keen to listen to new [00:41:00] ideas, new thoughts, new ideas.

Mehmet: New opinions about like trending topics in tech and entrepreneurship and startups as usual So don't hesitate to reach out to me. I would love to talk to you Thank you very much for tuning in today, and we will be again very soon. Thank you. Bye. Bye