In the past few years, artificial intelligence has made great strides in its ability to generate content by itself. This is thanks to advances in natural language processing, which is the subfield of AI that deals with understanding and generating human language. There are now numerous applications of this technology, from chatbots that can carry on a conversation with a human, to systems that can generate news articles or even poetry. In this article, we'll explore how AI is able to generate content, and look at some of the applications of this technology.
OK, so let’s pause right there. Do you have any opinions on the intro to the article? Boring, too neutral, good – or even plain wrong? Well, as it was written by AI algorithm GPT-3 from platform OpenAI (based on the prompt "write an intro to an article on ai-generated content"), please send your feedback to Elon Musk rather than me. By now, you probably understand what I’m getting at with the headline of this article. Because content, and how it's produced, will likely transform fundamentally in the near future. The flood gates for automated and AI-generated content are opening up and you will probably find that you can’t afford to stand on the sides. That’s all good and well, but without setting a solid foundation of quality standards, you might soon end up a contributor of the ever-increasing volumes of crap that clutters the Internet.
Every day, enormous volumes of data are added to the untended wilderness that is the Internet, making it overflow with everything from society critical services and news to cat videos and cleverly SEO-engineered click-farms. Just for fun I tried to find out how much content we have online today, and the closest answer I could find was 2.5 quintillon bytes. Not sure exactly how much that is, but I guess no one would mind me saying that it is “a lot”. A more tangible number is 90%. As in how much of the world’s data that was created in the last two years. Yikes. Mind blown!
And sure, while cynical, systematic misuses for personal gain sometimes make us think of the Internet as a cesspool of everything unholy, most stuff out there is still there to inform, educate and entertain. But it’s that untamed nature that still makes it such an intriguing place to explore and innovate.
The keyword for content has almost always been quantity. It’s all about how often you can publish, how much content we can create on our product pages, corporate blog etc for SEO reasons. Don’t get me wrong. I know that search engine algorithms still very much reward quantity and that you can’t expect to be visible anywhere unless you create lots of content. But how can that logic, where the one who creates the most content will be rewarded, be sustained in a future where content can be automatically produced, by itself? How can quantity be a factor for visibility when the press of a button can produce all the content you could ever need?
Well, of course Google and others have also identified this risk. Google is, after all, one of the leading players in AI and know better than anyone what this will mean. Their algorithm development is already aimed at what they call EAT and YMYL. Or in plain english – quality content. And content producers must change accordingly: just as tools to create unlimited amount of content are appearing, the need for quality is at the same time on the rise.
So far, we’ve only been talking about text, but there are many uses of AI that will be applicable for creating content. You have all heard of so called deepfakes, video and images depicting real people or situations but are, in fact, computer-generated. Many of the uses are, so far, surrounded with controversy, but now we are also seeing artistic uses of the technology such as Alper Yesiltas's brilliant “As if nothing happened” series, showing that AI content also could be used for beauty. And, of course, also for commercial content creation.
Content creation in practice
All right, enough with the philosophy. Let’s try out an actual e-commerce example. Of course, a very simple one. But bear with me as we try something just for fun - the content needed for a fictitious product page.
I started with OpenAI, providing it with the task of producing “A product description about a tight-fitting red dress”, and set the length to 256 characters. Here’s the result:
“Looking for something sexy and elegant all at once? This figure-hugging red dress is perfect for a night out on the town! The fitted silhouette will show off all your best assets, while the ruching detail on the side gives the dress a touch of glamour. Pair it with some strappy heels and you're sure to turn heads all night long!”
Not bad at all! Not sure about the “ruching details”, and maybe bordering on sexist? Well, I can’t say I browse for women’s clothing a lot, but my bet is that it’s not that far from what is really out there. But all in all – I have seen many much worse product descriptions out there. So, creating the texts seemed like no problem at all.
So, now for images. This turned out to be way trickier, and of course you would never let your AI design the actual dress for you. But humor me here as the point is to show possibilities rather than a ready-to-use approach.
First, I tried a service called Craiyon. My first prompt read “A woman in a red dress” and after about a minute’s wait, I was served the following visuals:
OK, so the first attempt produced one scary lady. Not only is her basic physique questionable (is that a napkin instead of an arm...? 😱), her face is also the stuff of nightmares. I’m sure this is extremely impressive code - but the end results are simply unusable. Obviously, a human face is what is most challenging for AI, so the second time I tried doing only the dress, with better results. I got a tight-fitting red dress, but for some reason (maybe it heard my mocking remarks on its earlier attempt) it avoided the end of the garment where the head would go. So, still pretty unusable even in this improbable “I have no actual product”-scenario. It is a red dress, but only parts of it.
The Dall-E algorithm from OpenAi definitively delivered higher quality. This tool also had problems with human hands. The hands look more like blurry lumps rather than the sleek hands you would expect on a fashion model. But the dress and an otherwise female physique is there. A really impressive feature offers you “variations” of the image rendered, so you could basically get the same model in the same outfit, doing different poses. Furthermore, my backend developer colleague Adil assured me that it was also absolutely possible to tinker with the tool using code that would enable us to better control what poses and variations we would receive.
The Lexica.art Stable Diffusion engine also delivered more promising results. Just like Dall-E it obviously has problems with human physique. The dress - no problems, the faces of the women generated could be maybe be described as... blurry, if we are being kind.
The bottom line for AI-generated content
So, no more playing around, let’s get to the point. AI-generated content is already here. It’s no longer a matter of when it will be used, it already is. Not only are massive amounts of AI-generated copy/texts being used online already, more bespoke visual solutions (that are way more accurate as they are specialized for a certain type of image, like clothing models) like ZMO.ai and others are already live and producing high-fidelity imagery for budget-conscious e-commerce players like SHEIN. And Swedish e-commerce suppliers Optimizely have recently launched an AI-powered image generator for their Content Marketing tool.
Still – AI in its current state is building upon the foundation of massive amounts of data that it collects to learn and improve. And as such it will not initially strive to be edgy or even unique, but rather merely try to pass itself off as human. In a sense, it will mainly aim to be average rather than a superstar. And as a human content creator, that feels, this is more of a relief than a threat. If we can devote our time to being creative, make our respective brands stand out and leave the task of describing 200 different types of men’s low-price underwear to a machine, I am more than OK with that. And if it can do so even better than a human, all the better as it will help our SEO - and hopefully conversions.
AI-generated content might very well be the next big thing, and I would highly recommend you to go out there and try all the tools that you can find. But for now - learning from a woman with napkin arms – I’d still recommend proofreading. 😉
(And oh - wondering about the strange Mona Lisa lego art used for the article? It was produced by the AI-powered image generator Lexica.art, using the prompt "A product photo from the lego toy of Mona Lisa")