Editor’s note: WRAL TechWire contributing writer Jen McFarland has 20+ years working in IT with experiences across a range of tools and technologies. She wants to help small businesses and teams design, improve, and maintain the technology that helps them succeed. In 2022, she incorporated Marit Digital.
RESEARCH TRIANGLE PARK – ChatGPT is missing something important, something fundamental: a quality user experience.
While OpenAI may have achieved a record-breaking number of users in its first three months, I would argue that the “adoption rate” probably leaves something to be desired. How many of those accounts were created by users flush with curiosity and the desperate desire to see a computer generate a limerick about their boss? How many of those users never logged in again?
The problem is that while the tool undeniably offers what people want, getting that content is a real challenge. Whether due to the interface or the experience of interacting with the AI, there’s lots of room for frustration. And as any web user who has had to wait more than 4 seconds for a webpage to load will tell you, that doesn’t translate to a lot of return visitors.
Jakob Nielsen, a pioneer in user experience (UX) has written about AI experiences extensively. A major concern is the “articulation barrier.” It turns out that requiring users to write quality prompts for intent-based outcomes requires – unsurprisingly – really good writers. According to the Program for the International Assessment of Adult Competencies (PIAAC) more than half of the world’s population has “low literacy” making the use of prompts a challenge. Anecdotally, the people I know who have had the most success with generative AI are the people who have played with it the most – and have master’s degrees in communications.
Nielsen also writes about the fundamentals of web user interfaces (UI) that have been ignored or forgotten in the frenetic race to push generative AI in front of a salivating audience. “AI companies only need to hire a minimal UX team and conduct the cheapest qualitative usability studies with 5 users to derive a plethora of actionable insights,” Neilsen writes in a recent blog post.
Let’s go find a few.
The ChatGPT User’s Experience
To be clear, I think most of the generative AI tools are suffering from UI problems, including Bard and Midjourney, but for better or worse, everyone is talking about ChatGPT so let’s start there.
It has been 11 months since the launch of ChatGPT and the interface is nearly the same as it was when it was released.
What’s changed? Probably the biggest move is the four “floating” prompts that appear at the bottom of a new chat. The four prompts are random and change on each page reload.
In the unlikely event that one of these prompts is exactly what you need (for example, “Make up a story about Sharky, a tooth-brushing shark superhero”) you can click on it and the prompt will automatically be entered and submitted. So here is our first user experience fail: rather than allowing you to edit the sample prompt, it’s immediately submitted and you’re now on the hook to wait for the completed response, even if it wasn’t actually relevant to what you needed.
From here it’s possible to edit the initial prompt, however, that option is only made clear to you when hovering over the original prompt text. Unlike other icons that show all the time, the prompt edit icon is hidden. Further frustrating matters, none of the icons have labels, tooltips, or other descriptive text to explain what they’re for. You have to click on them to figure out what they do, which in some cases involves giving OpenAI a thumbs up or thumbs down on the chat response. And once you click on those thumbs, there’s no option to undo that feedback.
Then there’s the matter of all this content. It’s nice that ChatGPT is so verbose, but if you’re using it with any frequency you know there are times when you just want it to shut up. While there is a “Stop Generating” button there’s no way to remove or hide a prompt or its response. Short of starting a new “chat”, there’s no way to clean up your thread or designate the important points for future use.
The Bard & Midjourney Experience
I’m being hard on ChatGPT because while Bard is also not great, it is definitely better. Bard provides more starter prompts and lets you edit them before submitting. It also provides more options for working with and sharing prompts and responses, including the ability to use voice prompts and have responses read aloud. Google’s feedback and response action icons all have labels, the ability to undo actions, and use graphics that are easily recognizable from the Google Workspace suite of tools. Finally, Bard provides a way to search Google for evidence of a prompt’s content or related results with a single click, a piece of insight sorely lacking from ChatGPT.
Certainly, issues remain. There’s no way to save a prompt, or even set one or more of the provided suggestions as defaults that stay in place all the time. There’s also still no option to “close” or hide a section of a thread to navigate a long series of prompts and responses.
Midjourney is a whole other case, with complications brought on by the user interface being in Discord. For those unfamiliar with the platform, it can be unclear what options are available for prompts or how to get help. The generated images also fall victim to a classic UX issue: the button layout. In the output below, which image is #2? Which is #3?
What’s still missing?
It’s nice that generative AI is so good at talking like a human but you might have noticed that humans often aren’t that good at communicating.
Real success with generative AI requires help with the communication and that means quality prompts – a fact that’s been clear almost since ChatGPT’s arrival heralded the new job of “prompt engineer.” If real-world users are going to have day-to-day success, they’re going to need help understanding, creating, and storing functional prompts.
Further, they need good ways to utilize them. Storing a bunch of prompts in a doc or spreadsheet to copy and paste when needed is not going to meet the productivity requirements of the AI age. These tools need to offer prompt basics at a glance, options to save and share prompts among teams, and better data on the interactions.
ChatGPT especially also needs to improve its interface. The novelty of the tool is wearing off; if OpenAI wants people to engage with the product it needs to be easier to work with (especially at $20 or more a month for those paying for Plus subscriptions).
As Neilsen says in his post, “The usefulness of a product is the combination of its utility and usability.” It’s beyond question that generative AI has incredible utility, beyond almost any tool we’ve ever created. It’s for that reason that usability has been so easily overlooked on the path to millions of users. But for generative AI to become a fully realized solution across countries and demographics, developers need to start considering the users.