Undecided: AI content generation

Should UXers be using this? Is it even possible?

These LLMs (large language models) are interesting tech. It’s August 2023, so I’m going to assume you have some basic familiarity with programs like ChatGPT and Perplexity, which are all the rage. It’s cool how you can prompt these programs and they’ll spit out content.

The trick is prompting them, of course. But if you get the prompts right? Neat!

This week at my workplace, we’re looking into the idea that we can use LLMs to streamline our UX content generation. I’m the lead of documentation for our design system, so “content” falls most neatly under my umbrella.

What content do we wanted generated? This sort of thing…

Success states, error states, etcetera. I can’t tell you how many times I’ve been working through wireframes and I realize I need to write something up for these. I wish we had an *official* content writer, I think.

In the last year, one of my colleagues created a guide for how to come up with the content for these modals. No, not AI — he created a document, with his own human hands, which outlined types of modals and the structure for organizing them.

Here’s a basic example (one I came up with just now — his document has much more detail):

Header: Invalid Type

Body: You entered [this]. Please enter [object] that is [that] or [the other].

Backstory: It’s because we’re looking to make this document “official” that we decided to pursue the AI question.

So we’ve thought through the prompts…

It’s not lost on me what this document is.

It’s a way to prompt humans to generate content. So, could we load this into AI and let AI generate it instead?

And more importantly than can we… should we? Will this help our workload?

(It’s important not to use something just because it’s there!)

Automation in software development

Over the weekend, my friend in software development was telling me about how he leveraged AI. He said that 80% of his job is finding code that already exists. Some of this is human-made; some is AI-made. He says the other 20% of his job is making sure the code is right, that it’s going to work. His manager is basically asking him to do UX work with this 20%. (And it’s his opinion that they should just hire someone to do that full time, since it’s not the sort of thing you want to rush.)

Do you hear that? He’s saying the UX work is not the automation. UX work is less similar to software development and more similar to curation and quality control.

That’s been my experience too.

In fact, it seems that one theory for why UX Design exists is because, as software development becomes more and more an assembly line, someone needs to look out for the user. Someone needs to make sure the software works for the context.

If we try to automate UX design, are we truly doing UX design?

What if humans are — for the foreseeable future — always superior to AI in judging what makes sense to other humans?

It’s not black-and-white, of course. We don’t need to be all-or-nothing. It may be that there are some aspects of UX Design that we can streamline.

So is AI the right tool?

Tomorrow I’m meeting with colleagues to discuss this situation further, but I have an initial thought.

I’m aware of how exact our prompts will need to be to generate 1-2 sentences of precise content. By the time we’ve figured out the prompts, are we 95% there to writing it ourselves? Is the guide that my colleague created really all we need? Is it worth expending an extra 40% of labor to generate that last 5%? (My percentages don’t add up, but I think you get me.)

It would be different if what we needed to generate was a long piece of content. But we don’t.

Hmm…

Verdict

In the end, it’s important to keep an open mind when it comes to burgeoning technologies. If the future isn’t here now, it may arrive shortly. I’ve heard that LLMs are hitting a conceptual limit, and that newer AI models may be based not on a mass conglomerate of content, but on teaching AI to think similar to how humans are able to think: by learning, by layering knowledge on top of itself.

Those models could go much further. Those models may truly offset our workload — by doing the analysis, the prompting and the content generation, all in the blink of an eye. Cool.

AI content generation? In our out?

For us, it’s TBD.