AI assisted content creation is an exciting new approach driven by algorithms and models like GPT-3, DALL-E, Stable Diffusion and more. At CodeArt we've been experimenting how to use these technologies to enhance the editorial process.
This is part 1 in a short blogging series about AI assisted content creation. Read part 2 here.
In 2021-2022 the available AI technology for creating and enhancing really useful content, including text, images and latest videos - have exploded - and especially the image generation has a lot of buzz at the moment, with contenders such as DALL-E 2 (OpenAI), Stable Diffusion and Midjourney (among others) showcasing more and more mindblowing creations and tech.
Alongside that, are the integrations and we are starting to see this technology used in everything fromgraphic design tools, website generation, social media, dynamic AR/VR, game graphics and much, much more. I've been a big fan of the possibilities for using these technologies in what I've started calling "AI Assisted Content Creation" - the keyword here is "Assisted".
Cause as powerful as it is, I still see a need for a human editor to get the general ideas and putting it all together - but all of a sudden a lot of things got so much easier - and hence your reach increases so much more.
But first things first. Let's take a look at what two of these technologies are - and what they can do:
Textual content generation / enhancement.
For example through GPT-3.
By taking a prompt where you either start on a text or write a specific wish for a text, it can generate a text as you wish. Or modify an existing one. Or translate one. Or summarize the meaning of one. Or all of the above at the same time.
This is a pretty amazing tool for helping you write a lot of copy / or just enhancing the text you already have. But for quality content it still needs supervision and proper inputs.
You could theoretically ask GPT-3 to generate an entire article about a given topic - and it would provide you with a good article, that did exactly that. However without further instruction it would probably have made up quotes, and you would have no control over the point it's trying to make.
But instead you could do something much more useful. For example you could provide a text like the paragraph in the beginning where I introduce this blog post and ask it to summarize it to a bullet list - and it would look like this:
- The available AI technology for creating and enhancing content has exploded, with image generation receiving a lot of buzz.
- Integrations of this technology are appearing in graphic design tools, website generation, social media, dynamic AR/VR, game graphics, and more.
- I see a need for a human editor to get the general ideas and putting it all together, but a lot of things have become much easier.
You could also provide a half-written blogpost and have it finish it. Perhaps just tie the different sections together in a smooth way and correct the grammer.
Sometimes I use it for brainstorming - like "Which 5 key topics could be interesting to discuss surrounding the ethics of AI assisted content creation?"
And the response:
1. The impact of AI on content creators' livelihoods.
2. The role of AI in content creation and the potential for it to replace human creators.
3. The ethical implications of AI creating or manipulating content.
4. The impact of AI on the quality and accuracy of content.
5. The potential for AI to be used for malicious purposes, such as creating propaganda or spreading false information.
All good points to explore (and I just might in future blog posts :-) ).
If you already have a lot of domain-specific content to your industry or business you could also train GPT-3 on that content and base the responses from that. Like - have it read questions and answers to/from your support hotline and based on that generate a FAQ for your website.
You can use it as much as you're comfortable with - but always remember to quality check the output - and that the messaging is what you want it to be.
This is what really tends to blow people away. Typically you 'just' provide a prompt of what you want to see generated and the AI will suggest options that match that. Potentially this means a lot of boring stock-photos on your websites and a lot more unique works.
What really seems to shock people, is that this isn't just a search in existing images - it's newly generated content based on your input. The way I like to think of it is that this is kind-of a reverse of the rather well-known image recognition where AI tells you what is on a picture - Now you tell the AI what should be on the picture and the AI creates a picture that would match that.
For example this image is an example of how AI imagines it would look if we asked Salvador Dali to draw a content editor creating content. Or like this - A space shuttle taking off as Monet would have painted it.
But we can even do more photorealism. Like the below photo where I asked DALL-E to make a macro 35mm film still of a happy dog cuddling a kitten and a duck in front of a waterfall.
But not only that - you can also take existing photos and edit them. Erase an area an ask the AI to insert something in a way where it matches the rest of the photo. As in the photo of my dog, where I asked DALL-E to insert a mountain lion.
Notice how it even gets the shades right.
Or you can extend on a photo. Add more canvas and ask the AI to fill it in. Like this:
In fact, the possibilities are quite endless - and I've started to find that DALL-E and Stable Diffusions have become my goto-tools when I need to illustrate to make a point. For example in blogposts or powerpoint presentations.
But... A word of caution should be here as well. With great power comes great responsibility and that is indeed the case here. You could easily use both text and image generation to create fake content, to scam people or worse. In my eyes it is of vital importance to always accommodate AI generated / modified images with a notice informing the viewer that what they are seeing is AI generated.
You can play around with GPT-3 and DALL-E yourself at OpenAI.com. Or explore StableDiffusion at stability.ai.
This was Part 1 in this series. Find part 2 here, and see a demo of a prototype where AI assisted content creation is used in Optimizely CMS (Episerver).