Generative AI is very good at predicting derivative work. Almost too good! A huge problem is it consistently plagiarises training data and source material.
Simple derivative AI is easy and kinda boring
What if you own the source material? And you want generative AI models to produce derivative work? I certainly struggle to write a compelling post excerpt after drafting 1,000 thoughtful words for Searching for Answers. Here, I want generative AI to produce something specifically derivative of my content.
Today, I want to explore what happens when you go beyond simple derivative work such as summarization. I want to introduce the idea of “derivative plus” as the next big generative AI opportunity: this is where the AI goes beyond what you can do manually.
Derivative plus and going wildly beyond what’s humanly possible
There is the opportunity to make an outstanding quality output. The perfect meta description or the perfect headline to A/B test! We’ll call this “derivative plus”. This is where we go beyond what’s humanly possible.
I mean “humanly possible” quite literally: the leverage of AI that I find exciting is where we do things that were impossible before. Today, let’s focus on making use of AI’s ability to ingest a phenomenal amount of information when generating an output.
ChatGPT is a good example of “zero shot” generative AI: you ask it a question and it gives you the output. You don’t have to worry about training or directing it. Zero shot is considered really hard to do, and pre-ChatGPT one would have relied on “few shot” AI, where you offer some examples of what you’re looking for. Pre-ChatGPT, you did this because otherwise you got a low quality and unwieldy output.
In a ChatGPT world, though, few shot prompt engineering is really powerful and it’s how we open the door to derivative plus.
Take a few shots and get wildly better results
Consider this: if you want to generate a meta description for a piece of content, you can give the AI your content and ask it for a meta description. You might even ask, “Make it really good,” and tell it how many characters to use. This is still zero shot, though, and it’s up to the AI to follow your instructions. Gen AI is getting better and better at following instructions, but it’s still imperfect.
A few shot alternative is to provide your content, provide instructions, and then load in additional context or examples. For our meta description, we might dynamically load in the full blog posts and accompanying meta descriptions from our 10 most successful pieces of content. You don’t need to tell the AI what you like about them – just let it figure it out*. Do that at scale, and you’ve gone wildly beyond what was humanly possible.
Now we’re getting somewhere: instead of a mediocre meta description, the output is now inspired by our best meta descriptions. Even better, why stop at your own content? You could mix in inspiration from meta descriptions of the top-performing competing content in Google! And copy not their content but what makes them successful! This is “derivative plus” and the untapped opportunity.
If you’re working on conversion rate optimisation and want a perfect headline, again, you could feed in the surrounding copy and context and get a decent headline back. Or, if we embrace derivative plus: copy from a dozen top-performing competitors and the headlines they use as examples, and then feed in your copy and you’ll get an output in line with that top performance.
*I can’t help myself adding more nuance: the AI is not going to “figure it out”. It doesn’t understand and thus cannot follow a trail of logic to understand. Instead, it’s going to use its training data and the examples you’ve given it to produce something that looks like the examples you’ve provided. The outcome is the same: a good output in line with the examples you’ve provided.
Derivative plus in practice
ChatGPT plugins or custom GPTs are a good way of using a derivative plus approach. With these, you can build workflows for specific use cases and bake in the context.
But, there’s something ironic about using AI for efficiency in a way that requires you to go out of your way to use it. I prefer to make AI applications available where people are already working. You’ll need to go through the API in order to do this, but this is remarkably doable. I’d recommend Zapier, Make, and Pipedream (listed in order of complexity) for easy-to-use solutions for interacting with the API for non-developers and developers alike.
Derivative usage of generative AI has no moat. If you’re using simple ChatGPT, your competitors – or even just your customers – could be doing the same. Derivative plus is so interesting because it adds moat and value in how the AI is set up: you can bake your expertise into the approach and then reap the rewards as you get wildly improved results.
This is the use of AI I get really excited about because it unlocks a better future. I think adoption of this is still behind the curve, and thus, there’s a competitive advantage to be had from moving quickly. Now’s the time to build and do better-than-ever work; all the while, your competitors will be searching for answers.