Generative AI: Harnessing Average

The Rise of AI: Harnessing Average

The power to create something with the generative AI technologies is incredible. Have half an idea, just give it some carefully crafted prompts and tell it to go. Presto. You’ve got something. Is it good? Is it bad? I’ll bet my bottom dollar it is neither.

Perfectly average. There’s no chance it’ll drop the f-bomb, it won’t be outstanding and it won’t be terrible. It might be wrong, especially about specific facts and it won’t let you know if it’s not too sure about something.

This is all really to be expected, it’s by design. It’s how these models work, they figure out what is the average thing to write next based on the inputs (prompts) and it’s training (parameters).

At this point, I’ve stopped and asked the Notion AI to “Continue writing”. And the following is the output on the 3rd try of “continue writing”. Probably could have said it better myself, but this’ll do.

So, what's the point of using LLM/Stable Diffusion technologies if it's just going to create average content? Well, sometimes average is all you need. For example, if you're creating a large amount of content that doesn't need to be particularly high quality, it can save a lot of time and effort. Additionally, it can be a starting point for more creative endeavours, giving you a foundation to build upon and improve.

What do you want?

It all begs the question: what are you trying to achieve? What is the outcome you want?

Recently at work we needed the ability to programatically read from a mailbox and asses the contents of this. A quick question to the all powerful ChatGPT and there’s a quick and dirty script to do just that. “Extract structured data from x”. Presto. Does this need to be good? Not really. Average is fine. A few iterations, some poking and prodding and this part of the job is done perfectly well.

Take a look at the car industry, the difference between each of the different shapes across manufacturers is neglible. A hatch-back is a hatch-back and a SUV is a SUV. They all approach the mean.

Yet, the success of each manufacturer doesn’t revolve around the average. Success is at the margins. How well can we execute this design? How well can it be manufactured? How reliable is it so we can offer a 10 year warranty at a low price?

Just like our mailbox script, it’s just a piece of the actual work.


To go back to the generated text above, the focus on how these technologies apply to the creative industries has got my spidey-sense tingling. There’s a reason I think the majority of Taylor Swift’s music is pretty crap. It’s samey - as samey as fuck. But that album, Folklore? That’s incredible.

It was a creative risk and it paid off - it's a critical and commercial success. Could an AI have come up with something like that? Sure, with the right prompts and training data, an AI could have generated a Folklore-esque album. But would it have taken the same creative risks that Taylor Swift did? Would it have had the same emotional depth and personal resonance? Probably not. AI can generate content, but it can't replicate the human experience that makes art truly great.

And in a lot of ways, this homogeneity of creativity has got little to do with generative AI and more to do with the way we consume media. Just open up and scroll Instagram for a while, there’s a fair chance there’s nothing there that’s groundbreaking or amazing. It’s perfectly average, there’s some ups and downs but on the whole it’s unlikely to change your world.

So, is there a role for this in creative work?

Perhaps there is, but it’s all about finding where the difference is. Can we drop Mona Lisa’s eyes into a generated lady and expect greatness? Probably not, her eyes are not all that defines here, but perhaps, creating time and space with some useful inpainting with Stable Diffusion might help with productivity and time constraints for people producing artworks.


There’s also the possibility of training specific models based on a the art-style of an agency or a brand. This is certainly the kind of thing that langchain is enabling in a text-based world, I presume there’s a similar approach in images.

This type of customisation pulls us away from the world-wide “average” and into our local average.

It’s exactly this that Github Copilot and its ilk are doing. They find the local average for code based questions. It turns out using age-old tools such as tests and specifications helps us use it to fill in the spaces between the details.

There’s another vector of customisation happening, and that’s plugins. These will fill the void of capabilities where average simply isn’t good enough. Maths is a perfect example of this wherein ChatGPT will just simply be wrong: “The product of 483,838 multiplied by 222,344 is 107,553,872,672.” it is actually “107,578,476,272” – kinda close. The plugin should be able to fill in the gap and actually give us the actual facts.


So these customisations, either private or public will certainly close off a number of the shortcomings. They will be trained to have more voice, the clear places where the models are wrong will be papered over with better suited models or functionality. But throughout all of that, they will remain a product of what they fundamentally are, a giant averaging machine.

This is the fundamental that we need to understand while implementing these tools. For these tools to not just drive us towards mediocrity we’ve got to better understand who we are and what we are trying to achieve.

Keep understanding what is your voice, your idea, your uniqueness and lean in. Don’t let the ease of the tool create averageness.

Michael Gall

Software Publican