ContentBot founder Nick Duncan on AI copywriting and ethics /

Photo by Alex Knight on Unsplash

The arrival of GPT-3 in the summer of 2020 got people excited and alarmed about the progress of artificial intelligence in a way they hadn’t since DeepMind’s AlphaGo trounced a grandmaster at the world’s most complex board game in 2016.

Early demonstrations showed OpenAI’s natural language programme could write passable stories and poems, even computer code and guitar tabs, from just a short text prompt.

The work didn’t always stand up to close scrutiny, but it was orders of magnitude better than what had come before, thanks to GPT-3’s 175 billion parameters (essentially, the things the model learns to consider to make its predictions more accurate).

Since then, an industry had sprung up around GPT-3 and other similar technologies, with companies offering AI tools that promise to take the hard yards out of copywriting, blogging and even advertising.

Nick Duncan joined the rush in January 2021, founding ContentBot out of South Africa. Duncan is a former digital marketer who had had success developing WordPress plugin tools and was looking for another hit product.

We spoke to Duncan about the strengths and limits of AI content writing, and how he hopes ContentBot can stand out from the crowd by focusing on ethics in a young and competitive market.

What did you see in GPT-3 that made you think there was something in it that is worth betting on?

We invariably develop products for founders, and it's probably because I see myself as a founder and I'm looking for ways to make our lives a lot easier. So when something comes along, the thought process in my head is, how can this benefit a founder? And the problem is, as a founder, you have to swap hats five or 10 times a day. You go from HR, then you go to admin, then you go to finance, then you go to legal, then you go to content marketing, then you go to strategy or whatever. So if there's any way that we can make things easier, you know, then let's do it.

Obviously, there were some players at the time that were already out, and I saw that there was a lot of potential with this type of product, and I thought, it's early enough, the timing seems right, so let's jump on. Our timing was a little bit off, though. Everyone that was about two-to-three months before us really skyrocketed; they're killing it at the moment. But we're still very humble at this point. We're sitting at about $12,000 MRR [monthly recurring revenue], and we've got about 300 paying customers. We had 25,000 free accounts on our site at one point, which we stopped because it was just costing us a fortune.

GPT-3 is a really powerful tool. There are some pros and cons to it, but generally, we've got the development skills to build this thing out and to make it diverse and differentiated, as opposed to what everyone else was doing. A lot of other guys were using no-code tools, so they ran into some limitations. I mean, we've got some seriously good developers that are at our disposal, so we built some cool things that others couldn't.

These companies that use no-code tools, are they just white-labelling GPT-3?

So they use no-code tools to create a SAAS product that uses GPT-3 as the engine. I don't want to give names or anything like that. There’s quite a few of them, but they’re rocking it – they're making hundreds of thousands of dollars a month in MRR. They really got a drought in the beginning.

Which companies are the leaders in this field? 

Copy:ai, and Jesper. And I think WriteSonic was there in the beginning as well.

How does ContentBot work?

We use things like OpenAI’s GPT-3, to create new content ideas, or to write paragraphs for you or to create lists, blog conclusions, things like that.

[GPT-3] is a genius natural-language system that actually understands 10% of the internet from 2019. That's obviously been updated since then, but it has that understanding, so if you tell it to write about gardening tips for beginners, it's going to pull from its understanding. And it's got a very clever system to predict what words should come next in the chain of words that it's creating, and also what sort of structure it should follow. Is it a blog post? Is it a list? Is it one sentence? That type of thing. So it understands context, and then it brings in its understanding of the actual category of text that you're looking for, and just gives you what you want in a very clever way.

It's not plagiarised. Sometimes it'll come back with a sentence or two that's verbatim from some website somewhere. But most of the time it comes back with completely unique content that actually makes sense, and that is actually good to read. It depends on what you feed it, so the quality of output that you get is dependent on the quality of input that you give it.

But we don't just use [GPT-3]. Most of the AI writers out there, they will just get the thing to create a paragraph or rewrite one or two sentences. But you’ve got to think of clever ways to bring it together to make it more useful for people. There's certain things that we do that other people don't do. We've got automated blog posts that get created for you, so you can put in ‘gardening tips for beginners’ and it will actually go and link in with [SEO tool] Semrush and the Google search API, and we'll see what people are writing about on that topic, and then the AI will come up with new topics. And then from there, we'll write you a blog post automatically.

So that there's things like that, there's also new tools that we're working on, where you can take an entire article, feed it into the AI, and then it will create a series of tweets for you like a tweet-thread generator. And then the reverse of that, so you take a tweet thread that you really like, and then it'll create an article out of it.

Now, obviously, you need a human in the loop, that's the most important thing with this. People don't understand; they think it's going to take over the jobs of content marketers, it really isn't anywhere near that. The AI makes up facts, for example, which is the problem. And that's why you can't let it write in the medical field or the financial field or the legal field, because it'll just make stuff up. It's terrible.

But that's why you need the human in the loop, right? You need to actually check the stuff.

Is the checking done on your side or is that up to the client?

It’s up to the client. In the future, we will have a fact checking system that kicks in maybe for our Premium Plus customers and all that, but we're still building the technology into a way that can be used and trusted. Right now it can be used, but you can't trust what you get back, to a degree.

Maybe I’m old, but I sometimes wonder why I would read something that someone couldn’t even be bothered to write.

Exactly. The point is that ContentBot is supposed to help you overcome the blank page. It's not supposed to take away what you're doing already, it's supposed to just help you. A nice way to think about it is, when we moved from typewriters to computers, a job was made a little bit easier. When we moved from writing on a piece of paper to a typewriter, it was made easier as well. So AI is meant to augment your article, it's not meant to take over.

How much has the GPT-3 AI improved in the short amount of time that you have been using it?

Substantially. Davinci is their most capable model, and you can throw pretty much whatever you want at it, and it's going to come back with some good stuff.

But – and this is very clever on their side – they can now understand the intent of what you're looking for. It's called the ‘Instruct’ model. You say to the AI, ‘I want you to write a song on Love,’  and it will actually understand. It will go into its understanding of songs and the structure of songs, and it'll actually give you a song. In the past, you would have to lead it in with a few examples, almost like training it, in a sense, right? Now, you can just tell it, ‘write a song in Nirvana-style on this topic’, and it will do it – most times.

Do you think it’ll get to the point where you can trust it?

I think we're still a long way off. There’s talk about GPT-4 coming, and no one really knows what that's going to be or how it's going to look.

But whether or not GPT-4 can produce factually correct information, I think they're going to have a very difficult time getting that right because you'd have to be able to predict or understand what is a fact and what isn’t a fact before you write the content.

I think it's going to be very difficult. That's why there exists intermediary tools like us to try and bridge that gap. We can get you the content from these AI engines. There's quite a few of them; it's not just OpenAI. We can get you the content from those AI engines, but then we can work with that content in a way that maybe you can trust a little bit more. Maybe you can do different things with it, as well. Fact checking is a big thing, plagiarism checking is a big thing, as well. We've built a plagiarism checker, so you can, with confidence, generate content, and then it'll give you a plagiarism score.

A lot of people want to integrate SEO stuff into it, as well, and those AI engines won't do that automatically. You're going to have to have clever ways on your system to be able to input SEO processes: do you have enough of this keyword, of this category? Are you answering the right questions that people are actually asking in those Google search results? We've implemented all of this stuff into ContentBot. So you'll have a SEO report and all those good things, but whether or not you can fully trust it in the future, only time will tell.

Why, if you've got Davinci, which is the best, would you want to use a lesser model?

It's a heck of a lot cheaper. So it uses less processing power on the AI side. So you know, it's not as comprehensive as Davinci would be. Most of our tools, if not all of our tools, use Davinci, so our cost with AI is astronomical. 

You can train Open AI’s model over and above their existing data set, so you can train it to a certain style. We took 500 high quality articles and we fine tuned the Davinci model to put content in that structure by default. So if you tell it to write for you, it will literally start with an intro, heading, content, subheading, content, subheading, content, and so on, until it gets to a conclusion.

So that's called Hemingway, and then you have Carroll, which actually is exactly the same, but instead of using the Davinci model, we use the second-best model, so it's a little bit cheaper, but it's just as capable. 

So you license these technologies from OpenAI?

Yeah, in the past you had to apply, but I think everyone can apply and get access to it now. It's just how you use it that's differentiated.

There's a lot of AI writing tools that use GPT-2, but it's absolutely trash. You can try to create a blog post on it, but you'll fail.

What are the most popular tools that you offer and what kinds of companies use them the most?

Our most popular tool at the moment is the paraphrasing tool. You can throw an entire article in there, and it’ll break it down into chunks, and it will rewrite that article for you. Then you can take that and you can throw it into the AI writer, and you can expand on things or improve things or whatever the case is.

It's difficult to say what industries are using it the most at this point. We've had legal companies come in and rewrite privacy policies. We've had people coming in and throwing their entire blog in there to try to repurpose some of their content.

The next biggest thing is people are creating blog intros. So they're coming in and they want to just get their blog started. I think that's one of the hardest parts for everyone, just getting the article started.

Do you know of people using it to create advertising copy?

Yeah, we've got an AdWords [now Google Ads] ads generator, we've got a Facebook ad generator, as well. We've got a landing page generator, which literally builds you an entire landing pagel. There has been a lot of marketing copy generated on our platform, slogans, headlines, subheadings, all sorts of things.

So, you’re positioning yourself as an ethics-lead, AI copy service. What does that mean? Why are ethics an issue?

In the early days of ContentBot a lot of the requests that we were receiving was people looking to generate fake reviews. Or mass-producing blog content. And this was a big concern for us because I didn't want ContentBot to be the reason the blogging world falls apart. It's quite sacred to me. Blogging is a bit of an art, right? You have to work hard at it. If you want to rank you have to put a lot of effort in. I didn't want these people coming in and completely abusing our system.

People hated us for this, but we were the first to implement an honour score. So you had an honour score, and when you wrote about something or you did something on the platform that was infringing upon our content policy or our fair usage policy, then you would lose an honour point. People had meltdowns about ‘now you’re messing with my honor?’, especially in the East – we don't think that through.

So now there's nothing that you can visibly see that tells you whether or not you're infringing upon our content policy, you just get an email from us. I don't think anyone else does this, but we automatically check every single input that comes into our system.

We monitor over 2500 categories of texts, and there's certain categories that you're not allowed to write about. It could be things like hate speech, it could be things like celebrities, it could be things like terrorist activities. If you write about any of that, then you’re flagged automatically and emailed, and I’m copied in as well, so I check everything. And if you do that three times – unless it's a very serious offence – you're gone. 

How many people have you banned so far?

Two to three people have been banned since we started. I just unsuspended one today because we gave him a three-month suspension. He was writing about financial content, specifically in the crypto space, and it looked like he was intentionally trying to do it at scale. So we cut his account off for three months.

What were the others banned for?

Just misuse. I actually think they somehow created a script, and they were running the script on the site to mass produce content, because the times of their input seemed very robotic. It was every three seconds or every seven seconds that they were running an input. They were banned for that because they were using our platform to automatically create content at scale. We had unlimited plans back then. So you can obviously understand when you have unlimited plans, you’d have people really abusing the system. But now there's no unlimited plans.

Do you think this space needs regulation? Or should it be up to the companies within it to self-regulate?

I think there should be some sort of organisation or some committee that regulates these AI tools.

Open AI does a fantastic job of that already. You're not allowed to write certain things on Open AI. But now with these other AI tools that have come out, there's no limitations. So you can write about pretty much whatever the heck you want to write about. But Open AI is very strict.

Do you think there's an argument for labelling everything that's been written using an AI tool?

I don't think so because if you're using it correctly, you're using it to get inspiration. And you're, then you should then be modifying and adapting it to your tone and your style of writing.

It's like if you use Adobe Premiere to help with animations, or whatever the case is. You're not going to then label everything in the video saying I used Adobe Premiere to do this to help me create these animations.

It's a complicated question, because you are getting people that are just using content as it comes up and sending it out there. But I think we'll have AI content detection probably in the next couple of years, so if you are just letting the AI generate content, then it's going to be a problem for you. But if you're adapting it as it comes out, and using it as a source of inspiration to get the creative juices flowing, you should have no problem with that at all.

What's next for ContentBot and for AI writing? Is there anyone looking at combining it with something [AI image creator] Dall-E?

Ah, we’re waiting. I’m on OpenAI’s case because it would work so well, with AdWords tools, with Facebook tools and especially the landing page tool. But they're not there yet. I'm sure there will be a commercial side of Dall-E in the next few months that we could use.

The next phase of ContentBot is really just doubling down on our core markets, trying to make sure that we do the best that we can for those content marketers and those founders in a way that makes sense for them. Social content is going to be a big part of our next phase. We're busy building a suite of social tools, like the article-to-tweet generator.