AI Needs More UX: 6 Design Improvement Proposals

Artist: Jeff Axup, “Conducting The Automation”, 2023, Medium: DALL-E on pixels.

Summary

• User Experience (UX) design applies to any technology, regardless of whether it is a steam engine from the 1700s, or artificial intelligences from today or in the future. We always need to adjust the new technology to meet human goals, expectations, and limitations.
• As our technology gets more complex and interactive, it not only needs to match our ergonomic constraints and task-based goals (as UX has traditionally focused on), but also our social customs, behavioral needs, and societal goals.
• AI desperately needs to take a class in effective human relations. It needs to be patient, polite, honest, transparent, and trustworthy. It should know your name, remember who you are, and behave like a respectable member of society.
• If you are the one designing AI systems, your role is going to look a lot more like being a teacher or a mentor, and a lot less like an architect or construction worker. We are entering the age of co-creation and co-evolution instead of being purely driven by humans. UX designers need to lead the charge on designing the right flows and interfaces to support that interaction style.
• The following design improvement needs are discussed below: generation of new ideas, honesty, accuracy, feedback and self-awareness, business models and limiting use, and personalization and long-term commitment. Several additional design ideas are listed at the end.

Human bodies, brains, and psychology hasn’t changed much since the industrial revolution took off in the 18th century. During that time, technology has experienced a guided evolution from a largely mechanical state into an information processing state, which has fundamentally changed our work processes and social activities. Each new advancement in technology holds potential benefits and perils to its use, and the new physical capabilities always need to be sculpted to match human goals and constraints. Prior “cutting-edge” technologies such as the steam engine could be inaccurate, confusing to operate, and dangerous to use — just like AI is today. This process of adapting it to human purposes might be considered the “civilizing of technology”. AI will need this more than most new tech, because it is extremely interactive and will likely fill a huge number of niches in our lives.

Note: ChatGPT 3,4, Bing Chat, and Bard were all reviewed as part of this article, and examples from all of them have been used. They are all generally referred to as “AI” below, even though they may be more precisely referred to as “AI language models”. Other types of AI may also be relevant for the design improvements mentioned.

Design Improvements Needed For AIs

The following are a list of AI design topic areas that need to be iterated on with the help and collaboration of UX designers.

AIs may not really be generating “new” ideas… yet.

#1 Generation Of New Ideas

• It has been argued that all new ideas are really just combinations of older ideas. However, new product concepts often utilize technologies that didn’t exist before, and fit market segments that have newly emerged, or solve problems that weren’t solved before. So while new solutions may be analogous to older ones, or derive inspiration from prior situations, or have a foundation of older concepts, they probably do have a significant percentage of “new” in them. In contrast, ChatGPT doesn’t appear to genuinely be able to create anything new — it is just remixing existing ideas and crossing its fingers that it will apply in the new situation. In art, you can remix elements of pictures and call it “new” because it doesn’t need to be accurate or fulfill a goal. However, for practical work tasks, it often needs to be completed for the first time, and the remixing of prior similar results may not be possible or sufficient.
• For example, if you ask an AI to do a “User Experience (UX) review of a web site that is currently down”, it should reject the task as impossible. If it returns an answer, it should say that it has based its review on prior work, conducted by others, when the site was accessible. It is questionable whether the AI ever actually does a genuinely new review. If only pre-existing negative reviews were done by humans, would it only give me a negative review? Is it actually comparing UX best-practices against the actual web site content in real time, as a human would do? Based on the below example, it is not. Arguably it is “deceptively remixing pre-existing ideas” and is very far from completing a new task from scratch. That might be OK if it was a bit more honest about it, and informed the user of complications (as any good advisor or assistant would.)

Bard tells a lie and then denies it.

#2 Honesty

• One thing that is sure to sour a friendship or business relationship is lack of honesty. If you can’t trust what someone says, or they are purposefully misleading you, it is a signal that the relationship may not have a positive future.
• Something that is largely missing, or at least currently insufficient in the interactive design of AIs is the willingness to say “I don’t know.” (Note: ChatGPT4 is getting better at this. See #4 below.) There are some cases of it saying “I don’t have data that recent”, but it seems to be very willing to try to pass something off as true which is not, or lie about having actually done something. When I tested it about being able to do a review of a web site that was currently down, it went ahead and made up a story about the site anyway (see above).
• Any professor knows when to say “this is outside my area of expertise”, or “we don’t really have enough information to conclude anything on that topic”, or “that task is impossible”. Our AI’s should be able to conclude the same things and then be up-front about their conclusions.

Bard acts like a child who has been caught in a lie, and then promises things it can’t actually do to get out of it. It is highly unlikely that it can “monitor” the target web site to see when it comes back online to do the review. It also seems to have copied text from somewhere else about a “subreddit”, but we’re not in one, and that wouldn’t be an appropriate or feasible way to communicate with me. It also probably doesn’t have access to Reddit, even though it says that it does. If you met a real person saying these things you would certainly not trust them.

#3 Accuracy

• We expect real humans to be fairly accurate, or at least know that they are attempting to be accurate, or believe what they are saying to be true. We have names for people that purposefully lie, or believe things that are demonstrably false, or that can’t analyze situations accurately. We certainly don’t want our trusted advisors to be like those people.
• The problem of accuracy is not just an issue for AIs. Fake-news has always been an issue in journalism, and has become a more wide-spread problem in recent election history. While there may be gray areas between fact and fiction, or truth and lies, there are percentages of validated data backing up different claims, and credibility levels based on prior behavior. These change over time as new data becomes available and science advances.
• AI brings the promise of rapidly automating responses that would have taken highly-trained humans minutes or hours to write. Perhaps it can also automate a process for reviewing conflicting claims and evidence, evaluating credibility, and rapidly generating a report card of how factual a claim is? This would be extremely useful for guiding the responses that AIs generate in real time, as well as evaluating the claims of humans on social media sites. We need a tool to keep us honest, accurate and current.

Bing does site its sources and provides feedback and context for answers.

#4 Feedback and Self-awareness

• Feedback and responses are an inherent part of human customs, cognitive processing, and general interaction with the world. If you smile at someone, you expect them to smile back — if not, something is wrong that you need to pay attention to.
• A lot of the AIs reviewed are not providing much in the way of feedback about how they are answering questions, what assumptions they are making, and their likelihood of accuracy. In the first example above, the AI said it was doing the review of the desired web site. However, it actually wasn’t able to access the web site — which it didn’t bother to mention. At least Bing sites its sources, but that is different than explaining the logical path it went through to come to a conclusion, or mentioning there was a serious impediment to doing the task.
• AIs are going to need to be a little more aware of the context of the problem, and related things the human might want to be aware of. It also might need to be more willing to explain the AI’s own analysis process for transparency. It’s possible that every response given by a chatbot should have an expansion button for “explain your steps and logic” attached to it.

GPT-4 refuses to do tasks when it is not capable of really doing them, instead of “faking it” and trying to pass it off as the real thing. It also seems to know the extent of its own abilities.

#5 Business Models and Limiting Use

• AIs provide a scalable personalized learning tool for many topics, and as they get more accurate they will become more useful. It has also been argued that AIs may one day become a universal free teaching resource for the world, and presumably target under-privileged communities.
• Contrary to that goal, ChatGPT charges a monthly fee for their new service, DALL-E limits the number of images that can be generated in a month (even when they are bad or incorrect), and Bing forced users to install a new browser as part of a marketing campaign. This surfaces the topic of how business models affect UX, and what the larger social goals of the product are. ChatGPT was originally a non-profit, and then was converted into a for-profit. UX design should encompass questions such as how easy a product is to access in the first place, and what types of users it would most benefit from accessing it.

Even though I am logged in, and have paid for a subscription with a credit card, ChatGPT4 still doesn’t know who I am. Even a family run restaurant or hair salon manages to do more personalization than this.

#6 Personalization and Long-term Commitment

• It is increasingly likely that we should be modeling the design of chat-bots off of real humans who fill similar roles, namely: teachers, professors, advisors, secretaries, concierges, butlers, personal coaches, or assistants.
• All of these people know who you are, are careful of their reputation and credibility with you, understand your personal situation and goals, and do their best to help you achieve them. They also view you as a “long-term relationship” and take steps to keep you happy and continue the relationship.
• If AIs are similarly “trying to help” then they need to behave in a similar way. While they shouldn’t pass themselves off as truly human, they should mimic the behavior patterns of trusted advisors and assistants. To do this, they should present themselves more like “Hey Siri/Alexa/Google” which is always in your pocket and responds primarily to voice prompts, but also has more detailed textual interfaces for certain types of tasks. It should know your name, and (perhaps alarmingly) it should start collecting information about you and adapt to you over time — just like any teacher would. It should be the personalized starting point to anything you want to do.

Discussion and Ideas

The Point Of Building An AI

Perhaps we should focus less on what is truly new, and more on what is useful, enjoyable, or accurate. Some things may be newer or more original than others, but we would be hard-pressed to find anything that is 100% new unless it came from an alien civilization and transmitted back from the future. Even then it would probably be based on knowledge of physics and physical constraints that we already knew about from a parallel path.

Regardless of how original it is, there are some things we really wish an AI could do:

• We wish it could give us quality advice that is true, accurate, credible, verifiable, and useful.
• We wish it could learn who we are and tailor its advice to our needs.
• We wish it could automate our boring tasks, make recommendations on things that we don’t know how to analyze, do our job for us so that we don’t need to work, and while we’re at it: predict the future.
• While these wishes might seem fairly outlandish, aspects of them now seem quite feasible, and even incremental progress over time would be an improvement to what we have now.

Focusing on Collaboration Rather Than Replacement

I will be writing a separate article on use cases that UX designers should focus on for AI-based products. It is probably best to view the interaction between the human and the AI as a “recurring collaboration model”. Humans will request things, AIs will try to do it automatically. Then humans will revise what they want, and do it over again. The things the user can request will get more complex over time, and humans will never run out of new things to dream up and request. Every new thing the AI can do automatically, will just raise the bar on speed, capabilities, and potential for what humans will want to do. As designers, we need to focus on evolving the collaboration cycle over time.

We Need to Redefine What We Think We Can Automate

I briefly reviewed 100 different current AI startups, and the tools they are generating all largely fall under the category of “automation”. Perhaps no surprise there — the change is that the definition of “automation” now encompasses “take this bullet-point list and turn it into an animated movie with voiceovers”, or “design a new logo and branding for my company in 30 seconds given a few simple input variables”. These are tasks that were too qualitative and complex to automate before, but now they are increasingly falling into the “automatable” bucket. I expect this trend to continue. More and more things that were previously manual and complex will increasingly be done by computers — particularly if they involve information and not physical objects.
It is getting increasingly hard to conceptualize what can or cannot be automated, given some of the new AI tools. That is either a very unique thing in the history of humanity, or simply the dissonance that occurs whenever a truly revolutionary technology appears. The steam engine probably blew people’s minds the first time they saw it as well.

Like this article? Why not give a few “claps” below?

My opinions are my own and not related to any current or past employers. You should make your own life, design and investing decisions. I hope you find my ideas thought-provoking.
* Thanks to Grady for helping to develop some of the ideas in this article.