How to design unscalable AI

The tools we use and how we use them are changing. Or as John Culkin in 1967 mentioned, “We shape our tools and, thereafter, our tools shape us.”

The rate of change in these tools and their ability to mold to each of us presents unfounded potential, but also a darker path. A major risk being that users will miscalculate how to appropriately use the new tools given to them. Like a caveman with a chainsaw - a user might harm themselves, unaware of limitations or the repercussions of using the tool or service. Fast forward to a world where most digital tools are now super-powered with intelligent system mechanics. E.g. machine learning. The benefits and risks are both astounding.

What happens when a tool has a mind of it’s own? Making decisions on behalf of you, decisions for you, which you might or might not be aware of.

Relationships

Of computers and design - UX builds a relationship between man and machine. UX, a story-telling mechanism of interaction between these two forces. Designers relate this to bridging the gap, making man and machine seamless in their connection. In the early years, at the inception of user experience design, the goal was to make human-computer experiences more natural. A mighty goal - to create experiences that felt more human in their nature of operation. An optimization that would increase system effectiveness and user satisfaction. Most businesses saw this as a win-win. Though user experience design wasn't the first to try and make these optimizations. Pre-dating UX was the creation of HCI (human-computer interaction). Many of the most influential developments came out of Xerox’s PARC research center, such as the graphical user interface and the mouse (circa 1970). These were important milestones as they paved the way for early principals in the AI and ML design space. Xerox might not be a big name today in the consumer space, but they still produce great research in the domain of contextual systems and automated agent design. One of my favorite quotes relating to automated contextually aware systems ...

"There is simply too much variability given different conditions for designers to successfully model appropriate outcomes in advance." - Bellotti, 2001

(Intelligibility and Accountability: Human Considerations in Context Aware Systems)

Conversations

Great HCI and user experience denote a conversation. As it is with conversing between to living beings, there are good conversations, and conversations we’d rather not conduct. The "conversation" we have with our tools and products has an ideal state. The most comfortable state that most users seek will be similar to previous tools and products they’ve interacted with. Users seek these familiar conversations, interacting with their tools in familiar ways. Often the word used in these circumstances is, intuitive. Having expectations for how we should interact with our tools and products is natural. They are built upon previous relationships with the products we’ve already had time to adapt to. We call these affordances - A compelling indicator as to how an item operates. Affordances include both its perceived and actual functions. Users never approach a new tool or service with zero expectations. There will always be mental model associations from previous interactions of the next most similar equipment or tool.

Users have a constant and ongoing relationship with everything. From the tools in their tool belt (hammer and chisel), to the tasks completed using sophisticated software (email and Tiktok).

Users of physical or digital tools form relationships and associations as soon as they make contact with the product.

Whether a tool is used for a few seconds or many years, users develop expectations about how it should work based on its design, its behavior, and its response to input. Consider how slowly the average consumer has embraced virtual reality; this technology requires a completely new way of interacting. Now, think about the transition from using paper and pencil to typing on a modern QWERTY keyboard, which was introduced around 1878. This change was significant, yet now just as obvious as mastering a basic tool like a screwdriver. Examine toddlers playing with smartphones and you'll get the picture.

Dynamic Function

Historically, tools like the screwdriver, paper and pencil, or the keyboard, have had a consistent function. They didn’t adapt or change their purpose for each user. A screwdriver didn’t become a different tool for each person who used it, nor did paper and pencil offer different features to different writers. These tools remained unchanged, no matter who used them. This consistency is what made these products reliable and trustworthy throughout their use.

A magazine that would tell you what's on TV, and recommend you new content to watch. A little different than your Netflix home feed...

Now the relationship between man and machine has changed. No longer does the tool or product produce the same response for each user.

Digital products have the ability to morph and change after each use.

In the ever-evolving landscape of digital products, a remarkable transformation is underway. Our digital tools once perceived as uniform and monolithic, now possess the remarkable ability to tailor their output uniquely for each instance, while still maintaining their collective identity. This paradigm shift challenges traditional notions of design and user experience, prompting us to rewrite the instruction manual for navigating this new frontier. How do we design for this?

The Emergence of Dynamic Responses

Imagine a vast library of responses (responses from a product or system), each akin to a brushstroke on a canvas. These responses, though part of a larger whole, possess an inherent dynamism—an ability to adapt, evolve, and resonate with individual users. No longer confined to rigid templates, our digital tools now weave intricate narratives, responding to context, tone, action, and inaction. They create bespoke states and interactions within a standardized framework.

As users engage with AI-driven tools, they encounter a delightful paradox: the more they interact, the more personalized the experience becomes. The more the machine's model expands and trains. Each query, each prompt, contributes to a rich tapestry representing the user, personifying them in ones and zeros. The system learns, refines, and crafts bespoke outputs that reflect the user’s nuances. No screwdriver or typewriter ever had those superpowers.

So, how do users educate and protect their own agency in this era of dynamic responses? One guideline for traversing this uncharted territory: Recognize that AI responses are not static monoliths. They flow, adapt, and surprise. Therefore the interface and UX must support this new modulating foundation in the product's core design. Creators of modern software must relinquish preconceptions and embrace the fluidity of this new landscape.

As designers, PMs, or engineers we must design unique safeguards which embrace the concept of dynamism in product uniqueness (a unique product experience for each user) - Products with Personalities.

Modern AI researchers argue the move away from anthropomorphizing AI - allowing the public to heavily associate human like qualities onto AI products. For the general public I agree this makes the most sense. On the other side of the aisle, for creators of this software, we ought to do the opposite. Embrace the humanizing functionality of these new tools and services. Only from that vantage point, from that view might we begin to fully understand how to design for the relationship between AI products and users.

Designing for the AI and Machine Learning Renaissance

As designers, we must rethink our approach. No longer confined to rigid scripts, we craft experiences that blend a system's ability for inconsistency and surprise. One way is through XAI: Designing explanation interfaces that allow for dynamic responses and unique user sessions. - Keep following 90% Human for more on this topic ??

As we rewrite the instruction manual, let’s celebrate the delicate dance between uniformity and uniqueness. This is both a power and a threat. A threat to UX fundamentals. In UX there exists a core principle: Always allow the user to maintain a sense of control with the product. But there exists no principle to account for products with dynamic responses that harbor personality or intent. Products have personalities, and they cater differently to each one of us.

Creating safe tools and products once started with laborus manuels and training. We are entering a new domain where the concept of instructions is opaque. Is it feasible to have an instruction manuels for products with dynamic personalities?

At the crowning of UX, user experience was to craft the relationship based on fixed points in the product experience. Points where user and machine met, to exchange information or produce a set of responses or actions.

What if these anchor points of machine<>user interaction were never fixed in one place? What if the system's responses could not be easily reproduced? What if they were constantly changing, constantly shifting? How would our relationship with these same tools and products change? If output of our products, tools, and services cannot be expected, if the variability of system/computer response is infinite and constantly molded to each human, how can we design for a safe user experience?

Information, sharing, and education

Sharing information on what the system is doing and why. Educating humans on how they are changing the machine and how the machine is changing them.

In next month's article, we will answer product and engineering's most perplexing arguments.

"Our product doesn't have that big of an impact."

and... "Our users don't care to know about how the system functions."

Stay tuned for the next edition of 90% Human to learn more.

Konrad is currently writing the penultimate book on ML, UX design, and business. To follow along with the writing and research, subscribe to his updates on Medium, Linkedin, and the book website https://knpdesign.co/gdml.html

#digitalservicesact #aiact #90%human #gdml #machinelearning #ux #AIdesign