In this episode of the Opinionated SEO, I discuss the capabilities of and implications of chat GPT, the free AI-based service from Open AI.
I break down what GPT is, what it's good and bad at, how to craft requests correctly and how to avoid detection from search engine algorithms.
I believes that GPT offers great potential as a tool, but should never be trusted to replace human content writers, especially for investigative reportage.
I talk about how AI-based content can be used to create content more quickly and cost-effectively than traditional methods; I look at a past piece of content that I outsourced to a writer, compared to the same one using AI with the same prompt.
I recommends trying out Chat GPT, an interface into DaVinci 3.5 from OpenAI, to produce creative, long-form pieces of content without much effort.
Scott Aaronson was recently interviewed about AI safety; he mentioned that OpenAI research suggests that watermarking could help identify AI-generated content.
Google's AI tools are able to pick up on certain key features such as punctuation and word choice that may indicate the content was generated by an artificial intelligence program.
I believes that GPT 4 and 5 will make it increasingly difficult to tell the difference between human-written and AI-written pieces of content without adding original research or interviews.
Have a happy holiday season and I look forward to seeing everyone next year!
In this episode of the Opinionated SEO, I discuss the capabilities of and implications of chat GPT, the free AI-based service from Open AI.
I break down what GPT is, what it's good and bad at, how to craft requests correctly and how to avoid detection from search engine algorithms.
I believes that GPT offers great potential as a tool, but should never be trusted to replace human content writers, especially for investigative reportage.
I talk about how AI-based content can be used to create content more quickly and cost-effectively than traditional methods; I look at a past piece of content that I outsourced to a writer, compared to the same one using AI with the same prompt.
I recommends trying out Chat GPT, an interface into DaVinci 3.5 from OpenAI, to produce creative, long-form pieces of content without much effort.
Scott Aaronson was recently interviewed about AI safety; he mentioned that OpenAI research suggests that watermarking could help identify AI-generated content.
Google's AI tools are able to pick up on certain key features such as punctuation and word choice that may indicate the content was generated by an artificial intelligence program.
I believes that GPT 4 and 5 will make it increasingly difficult to tell the difference between human-written and AI-written pieces of content without adding original research or interviews.
Have a happy holiday season and I look forward to seeing everyone next year!
Phil:
Welcome to another episode of the opinionated SEO. Today. I want to talk about. Chat GPT. You've probably heard this. If you haven't, haven't tried it. Run over to open AI, take a look at chat GPT. It's free right now. And unless it's under heavy load, you can pretty much use it, just like their other AI tools, like their playground. it's. Really good. And this is really making people. Consider AI as a replacement. For a content writer. So what is chat? GPT? Okay. So chat GPT is a. Interface to open AI's model there DaVinci 3.5 in there. 3.5 was just released recently. It was version 3.0, which was really good. This is even better, but then they've trained this model using this chat GPT interface to work much more conversationally, much more like a dialogue, allowing it to answer follow-up questions, chat GBT can remember some of the things that you've asked. So there's some train of thought to it. It also can challenge you if you have some incorrect premises and it can admit mistakes. And because of that, it works really great in a question and answer format. In a chat type format. Now, what does it not good for? It's not good for factual data. If I wanted it to tell me what the hours were for a restaurant. That's not the kind of data it has. It's still trained. On that same data that I think goes to about July of 2021. So we're talking about about 18 months old, so it doesn't have anything in the last about year and a half. Now, what it can do though. Is, it's also been trained on helping things like fixing code. So if you have code and it has an error, you can actually paste that code in and say, what is going wrong with my code and help debug it. And it can actually help tell you what areas you need to fix. And so it's turning into this very different almost search engine replacement or forum replacement. Now you've got places like stack overflow. Which are saying, you cannot use it to answer questions. You have to come up with your own original answers, because what people started doing is they were taking people's questions, copying and pasting it into chat GPT, and getting those answers and pacing them back in and trying to get points for it. And as great as that is. It's using an AI tool for that. And they want it to be original concept. And remember that a lot of this data that it has probably also ingested from stack overflow and similar types of sites. Now they trained this. On the same thing that we've seen with the other AI tools. So you you've got like your Jaspers and things like that. They all tie in to this same backend system. really what's different with chat GPT. Is that the model that they did where it's really been tweaked so that the phrasing and the conversation Of it. Feel very realistic. there are limitations., they don't allow certain things like you can't say, how do I bully someone? How do I make a bomb? These kinds of things, it'll come back and say, we're sorry, you can't do that. There are some ways around it., and some people are trying all kinds of things in the end. If you really need that kind of information, you're going to have to go directly to the API, but it's not going to be trained in the same way as chat GBT is. And this is what's. I think the biggest differentiator is the training model that they used. The. Language model that they've done for dialogue is really just that much better., They're calling chat GPT, a sibling model. To instruct GPT, which is trained on following instructions in a prompt to provide detail responses. So it's really interesting. It takes a lot of their learnings that they have, and builds on it. But people are now starting to get a little bit worried. They're like, okay, if this were to replace a writer, what does that mean for marketing and for SEO and actually search engine journal did a whole piece on it and they tried to answer a few things and they said, you know, what can chat GBT do. What can't it do? Really what they're saying is. It's AI generated content when it comes down to it, they're not designed to do specific tasks, but they're tried to have general knowledge. And so if you have a very kind of niche, Thing or there's something that only a few people have answers. There's not a lot of data on the internet about it. It's not going to be able to give you those kinds of answers., a couple of things that you do need to know is that it is not a reliable source of information. You have to still fact check it. Doesn't know. Everything perfectly. And sometimes it makes things up. So an example that someone had given was., you asked a question about, where did this person get their first job and maybe their first jobs, not even online, but chat GPT kind of made up what it thought would be the type of first job that they would have. And that person said, you know what that sounds about. Right. But I really didn't work there. And so it, it almost creates like a story out of it., the other thing is it doesn't know about things really after that June, July, 2021. So anything in the last about 18 months right now, we're at the end of 2022. It does not have that information. So anything that happened recently. It gets kind of interesting when you start asking you about things that maybe happened 12, 18 months ago or even 24 months ago. Cause it doesn't have a lot of data on it., it does have some biases., it's not perfectly neutral., so you have to just kind of accept that it tends to be a little bit more positive and removes. A lot of the negativity in its responses, which may make it less factual., the other thing is it does need detailed instructions., the more specific you get, the better the answer's going to be. So you really have to spend some time crafting it. And this is what
Philip:
we'll probably start seeing more and more in content writers and copy editors it's going to be an art form on how you craft these requests to the AI tool. You ask it to do things in a certain voice, in a certain tone, in a certain way. You see that with Jasper as an example, they have these different, kind of recipes , so to speak and it's all of these different variables that go into it to create a certain type of response. And once you start creating those., you end up getting some really great results and you can start getting similar type results over and over for different areas that you're trying to do. The last part here that's really big is can chat GPT content be identified. And I went into this a little bit into detailed on the GPT three in Canada, or can it not be identified now? GPT two can be. It's pretty easy to identify it.
Phil:
GPT three. What I have found is that you can usually recognize. AI written content because it uses similar types of phrasing. Now that's it's default. And if you really spend that time crafting that request. You can end up tweaking it to sound a very specific way that doesn't sound like AI. And that's really important. What ends up happening is the default settings are probably something that a person could identify. If a person can identify it. Google or, you know, machine learning is going to be able to identify it. That's why it's so important that you don't just copy and paste. I've always found that AI content is a great starting point. It can never replace a journalist, someone who has to do investigative reporting. I find original content. And our original interviews in order to create content now, because of that. There's never going to be an AI that can complete replace something like a newspaper, because those things it's never been trained on. Now, if you feed AI a bunch of content, ask it to summarize and put that content together into something cohesive. That's where it can do a really great job. One of the things that I did recently is I looked at some content that I had requested about two years ago through an online writing website. And I took the exact same prompt that I gave the content writer. And I put that into one of the GPT 3.5 tools. And the content that I got back. Was just as good. And so it was just as good as someone. And I think it was a 10 cent, a word, you know, not a super expensive, not, a highly researched article or anything like that. There was no interviews, nothing like that. We wanted some information about a topic because on the website, we wanted to make sure that. The people were informed and understood what that service was that we didn't need anything really original is more to help a user than anything else. That content was just as good. It took me, I timed it. It took about six minutes to create that content. Two years ago, I spent something like 15, $20 on it for 250 words. This gave me about 550 words. And, , my cost was pennies. So what would make the most sense? Well, the AI based content, could it be identified as AI content? Probably not. I went through and copy edited it, cleared it up, changed some of the phrasing to phrasing that I preferred. And that was probably the bulk of the four or five minutes of the six minutes was just doing those little tweaks to kind of customize it to how I would write. And that's the big key here is that I made it my own. The last point that search engine journal had mentioned is they had a interview or recent article with Scott Aaronson who talks AI safety. And one of the things that he talked about from the open AI research is that there may be ways that they can watermark the content and they would do this by, things like how many punctuation marks are, where they are. Number of words, in a sentence, things like that, where. it can be understood through machine learning that if it's a sentence is structured in a certain way than it was generated by AI, that's pretty much all they can do., the AI tools that Google is using to determine if they're spam, they can likely look for certain words, certain. Errors and punctuation. I see that a lot where it uses dashes, where most people don't write with dashes. there's extra spaces after a comma, things like that. Those little telltales are something that a copywriter should be able to fix and probably adjust. And so when you're talking about creating a long form piece of content, if you don't have someone copy editing AI content, Then, yeah, it's probably going to be something that can be found, but realistically watermarking, if someone does a good job with their copy editing and goes through it, it's probably not going to be possible. So in the end, Chat GPT. I highly recommend taking a look at it. It's a really cool tool. It's a free for now interface into DaVinci 3.5 from open AI. And I think that what you'll find is that the content. Really does a great job. And in some cases it can save you a ton of time when you're trying to just come up with some responses, get some ideas on maybe top things you want to talk about., but Hey, give it a shot, write a blog post with it, ask it some questions. Use that to generate something, see how long it takes. See how creative you end up getting from just using the tool because in the end, that's what it is. It's a tool. And I think that it can be used really well. My opinion. Chat GPT is just the tip of the iceberg that we're going to see in AI content. When we see GPT four, GPT five come out in the next couple years. It's going to be really hard for you to be able to distinguish. That kind of content. From someone who's writing regular piece of content again, without that kind of original research interviews. I hope you guys all we'll see you next year.
