A New AI: Creation as Exploration and Discovery

Summary: AI transforms UX into exploration-based discovery. Users will navigate latent solution spaces rather than specifying outcomes. Creation shifts from building or describing to discovering possibilities. Future interfaces become collaborative playgrounds where users recognize solutions instead of articulating visions, augmenting human existence.

I have been too modest in my assessment of the degree to which AI changes interaction design.

In May 2024, two months after the launch of the first good AI (GPT 4), I declared that AI represented the first new interaction paradigm in 60 years, since computing shifted from batch processing to interactive commands:

• All earlier UI generations, from line-mode to the GUI, were variations of command-based interactions. The user would instruct the computer on what operations to carry out, progressing toward the goal one step at a time.

• AI is intent-based outcome specification: the user tells the AI what he or she wants to achieve, but not what steps should be carried out to reach that goal. The AI figures out the operations for itself.

Batch processing was the first paradigm for “using” computers, where individual users didn’t interact with the computer at all while it ran their “batch job” of punch cards. Since batch processing was abandoned around 1960, all new technologies during the following 60 years, from time sharing to smartphones, used the interaction paradigm of command processing. (GPT Image-1)

Intent-based outcome specification is a significant usability advance over command-based interaction, but it doesn’t eliminate usability considerations. In fact, AI introduces a new usability problem, the articulation barrier. It is difficult for most users (especially lower-literacy individuals) to articulate in prose what they want.

UI solutions have emerged over the past two years since I articulated the articulation barrier, such as a wide range of prompt augmentation and aided prompt understanding designs. However, these new designs only alleviate the augmentation barrier; they don’t eliminate it.

I now have a more radical vision for the future of AI UX: we will switch from users telling the computer what they want (whether through commands or intent specification) to users discovering what they need, through exploring a latent solutions space created by AI.

Discovery through navigating latent space to explore adjacent solutions to your current idea. (GPT Image-1)

Creativity switches from making to discovery. This will be a huge usability advance, because it is much easier to recognize that something fulfills your vision than it is to specify that vision. (This is a parallel to the well-known usability heuristic 6, Recognition Rather than Recall, but one level up in user empowerment, since the heuristic was concerned with recognizing commands and features, not outcomes.)

Enter the garden of discovery: a likely metaphor for the future of user experience, which will entail exploring a space of solutions where each user will pick what he or she likes the best. (GPT Image-1)

The 3 Eras of UX Goals

User experience design has progressed through three major eras, with the main goal of design shifting dramatically from productivity to influencing, and now to creating.

• Productivity: The UX Goal in Era 1, Business Computing (1960–1995)

• Influencing: The UX Goal in Era 2, Internet (1995–2025)

• Creating: The UX Goal in Era 3, AI (2026–)

UX has had three main goals during its history. (GPT Image-1)

Let’s discuss each of these eras and how the goals of UX work have shifted with the times.

Productivity: The UX Goal in Era 1, Business Computing (1960–1995)

User experience design started around 1960, with early projects like the design of the touch-tone telephone keypad at Bell Labs. This was the age of business computing, initially based on mainframes and later dominated by personal computers. Whatever the technology, the dominant applications were business tasks, such as accounting, payroll processing, document editing, and slide presentations.

The main goal of UX design in this period was productivity. We aimed to make software easier to learn: as I used to say, a company’s training budget is a juicy pork chop ready to be eaten by usability, because using our methods can cut the training time for any software application at least in half. Similarly, good interaction design could probably increase employee output by anywhere between 20% and 40% after they have learned the system.

In the productivity era of UX, enterprise training budgets were like juicy pork chops ready to be cut in half by usability. (GPT Image-1)

(Note that learnability usually has greater potential for usability improvements than efficiency of use. It’s very hard to increase throughput by as much as we can reduce training time. However, the user error rate is the usability metric with the largest potential, since we can often cut errors by a factor of 10 by designing less error-prone user interfaces, following usability heuristic number 5, error prevention.)

Influencing: The UX Goal in Era 2, Internet (1995–2025)

The rise of the Internet shifted the goal of user experience design to influencing users, typically to either make a purchase or stay on a social media platform to “doom scroll” indefinitely. In fact, this era gave rise to such infamous design patterns as the infinite scroll, which has no user benefits but often entices people to look at more postings than they really want.

Infinite scroll may not be a dark pattern per se, but it is a usability abomination that pulls people into spending more time than they intend. (GPT Image-1)

Designing to influence users follows a continuum from the benign, such as showing clear and interesting product photos, over acceptable implementations of Robert Cialdini’s influencing principles, to positively evil dark design patterns.

Caldini’s Seven Principles of Influence

1.      Reciprocity: Humans feel compelled to return favors, gifts, or concessions. This aversion to indebtedness creates strong social obligations, making reciprocity highly effective in negotiation and marketing. Example: Offering free trials or downloadable resources before requesting email signup.

Caldini’s reciprocity principle influences people by giving them something. Then they feel obliged to do something for you in return. (GPT Image-1)

2.      Commitment/Consistency: People strive to align actions with previous commitments, especially public ones. Small initial agreements often lead to larger concessions, as individuals maintain behavioral consistency to preserve their public persona. Example: Progress bars in checkout flows that show users they're “almost done.”

You’ve already made it so far. Why not finish? (GPT Image-1)

3.      Social Proof: We shape decisions by observing others, particularly during uncertainty. Collective behavior serves as evidence for effective action, making testimonials and peer endorsements powerful persuasive tools. Example: “2,847 people bought this today” or star ratings with review counts.

Social proof: if a lot of other people did something, they must (?) be right, so you should do it as well. (GPT Image-1)

4.      Authority: People yield to those perceived as knowledgeable or credible. Titles, credentials, and professional appearance establish legitimacy and amplify persuasive power. Example: Security badges and certifications on landing pages.

We tend to believe sources with perceived authority, for example, as evidenced by visible credentials. (GPT Image-1)

5.      Liking: We're more easily influenced by those we find attractive, familiar, or similar. Emotional affinity and perceived connection foster openness to persuasion. Example: Website showing photos of smiling team members with brief personal bios, making visitors feel connected to real people.

If someone is good-looking, you will be predisposed to like them, trust them, and follow their recommendations. (GPT Image-1)

6.      Scarcity: Limited availability increases desirability. Fear of missing out, driven by temporal or quantitative restrictions, motivates action by elevating perceived value and urgency. Example: “Only 3 left in stock” warnings.

Caldini’s scarcity principle is often verifiable in the physical world, and it might feel good to snag the only available item for sale, meaning that it’s not just a sales trick but also can be helpful for customers. The scarcity principle also translates well to the virtual world, but unfortunately, sometimes turns into a dark pattern if sites claim scarcity when there’s abundance. (GPT Image-1)

7.      Unity: Shared identity or kinship enhances influence. Appeals to common belonging, identity, or purpose trigger social motivators and deepen trust and compliance. Example: “Join 50,000+ entrepreneurs” message.

If you can find a way to create commonality or unity with certain users, they will be more likely to be influenced by your design. (GPT Image-1)

As the examples show, not all influence-oriented designs are sinister or unethical. In fact, progress bars were a recommended usability element even back in the days when UX aimed to support users and increase their productivity.

There is a continuum from honest UI design to questionable influencing to outright dark design. Sometimes the same design idea can change colors like a chameleon and land on different spots within this spectrum, depending on exactly how it’s used. (GPT Image-1)

However, I feel that Internet-era design has strayed too far from my original passion for helping users. As the old saying goes, if you’re not paying, you are not the customer; you are the product.

If you’re not paying, you are not the customer; you are the product, and the user experience will never be great. (GPT Image-1)

Creating: The UX Goal in Era 3, AI (2026–)

We are about to switch the goals of UX design once again. The original productivity goals of my youth are becoming less relevant in the AI era, because it is unlikely that users will be performing very many work tasks. Almost all the tasks human employees used to perform will very shortly become the job of superintelligent AI, leaving humans with only three activities: agency, judgment, and persuasion.

Once AI does all the traditional work (around 2030), humans will still be in charge of agency, judgment, and persuasion of other humans. (GPT Image-1)

Similarly, the influencing goals of the AI era will become less relevant as AI agents take over from human users. With no humans seeing your website, there will be no benefits to designing according to  Caldini’s Influence Principles. I have used the phrase “No More UI” to describe this upcoming Internet.

In the future, most Internet services will only interact with AI agents. While human users probably won’t be outright banned, people will prefer to interact with their personal agent and shun brand websites. This marks the end of persuasion and human influencing principles as the primary driver of web design. (GPT Image-1)

On the other hand, it is possible (even likely) that new influencing principles will emerge for online services to curry favor with AI agents, similar to current trends that utilize GEO (Generative Engine Optimization) to increase the presence of a website in AI chatbot answers. However, influencing AI agents will be an entirely different beast than influencing humans, with our evolution-determined weaknesses.

The goal of GEO is to be nice to the AI so that it will quote you. We need to do much more work on how to influence AI, but doing so will be harder than influencing humans, who were shaped by evolution to be susceptible to many well-documented influencing techniques. (GPT Image-1)

All is not lost for UX. While we won’t need to spend much time on task usability or influencing, a new need is arising: that of helping users create and discover new things.

Creation and discovery represent fundamentally different human activities from productivity or commerce. When AI handles the mundane execution of tasks, humans are freed to focus on imagination, curiosity, and the deeply personal act of making meaning. Ultimately, meaning is the one thing that AI can never do: only the overlords (us) can decide what’s meaningful to us.

The UX of the AI era must therefore shift from removing friction in predetermined paths to opening up possibilities we haven’t yet imagined. Creation-focused UX asks, “What might the user want to bring into existence that has never existed before?”

This is not about making better content creation tools, though those will certainly evolve. It’s about designing systems that augment human creativity itself: interfaces that help us explore the adjacent possible, that surface unexpected combinations, that translate half-formed intuitions into tangible expressions. Where productivity UX minimized cognitive load, creation UX might deliberately introduce productive friction, the kind that sparks new thinking.

Mix, match, and get surprised when you discover something better than what you could have thought of yourself. (GPT Image-1)

Discovery, too, takes on new dimensions when freed from the constraints of task completion. Instead of search results optimized for quick answers, we need experiences that foster serendipity and wonder. Instead of recommendation algorithms that narrow our choices based on past behavior, we need systems that help us surprise ourselves, that lead us down paths we didn't know we wanted to explore.

UX metrics will need to evolve accordingly. We won’t measure efficiency or conversion rates, but perhaps the novelty of what users create, the diversity of their discoveries, or the satisfaction they derive from the creative process itself. Success might look like a user spending three hours in flow state, not three seconds completing a transaction.

When creativity and discovery become the goal of UX design, our metrics have to evolve accordingly. For example, it may be positive if the user spends more time on a task, if that time is mainly in a flow state, which is an enjoyable experience. Getting a birthday card designed fast is a goal if you’re a professional designer, but for an amateur, “the journey is the reward.” (Or at least it can be, with a well-designed creation system.) (GPT-Image 1)

This represents a return, in some ways, to the UX field’s earliest aspirations, as stated by Doug Engelbart’s goal to “augment the human intellect.” However, his phrasing was still clouded by the productivity thinking necessitated by the business use of computers during his time.

Today, I proclaim the new goal of UX to be “augmenting human existence.”

With AI and the resulting shift of UX to supporting creation and discovery more than productivity or persuasion, the mission of UX professionals should be upgraded from Doug Engelbart’s pioneering exhortation to augment the human intellect to my new vision of augmenting human existence. From a mouse to an elephant! (GPT Image-1)

The UX designer’s role transforms from eliminating friction to creating generative spaces, shifting from guiding users to predetermined outcomes to helping them discover unexpected outcomes.

In this new paradigm, the best interfaces might feel less like tools and more like collaborators, less like pathways and more like playgrounds. They will need to balance structure with openness, guidance with freedom, familiarity with surprise.

Discovering Solutions in Latent Space

As generative AI floods the digital world with content, products, and solutions, the challenge shifts from a scarcity of options to an overwhelming abundance. If an AI can generate a thousand competent solutions in a minute, the primary user need is no longer the production of the solution, but the discovery of the best one.

But what does that even mean? Cristóbal Valenzuela (CEO of video-creation company Runway) gives us one likely direction that aligns with my thinking: AI use becomes aimed at exploring the latent space of possible solutions to see what we discover and like, as opposed to the old approach of having a solution in mind and trying to wrangle the computer into fulfilling that solution.

This is a profound shift from all previous forms of creation, which were fundamentally about assembly: arranging smaller units (paint, words) into larger units (paintings, novels). AI creation is about evocation: coaxing form out of a compressed universe of possibilities.

We are rapidly progressing through three approaches to creation, whether of content or action objects such as software and products:

• Building: the pre-AI era, where the way to get a desired outcome was to build it yourself, whether painting with a brush or a series of Photoshop tools, and whether writing with a quill feather pen or Microsoft Word. For action objects, this was the era of traditional software engineering.

• Describing: the current era of generative AI, where we can use intent-based outcome specification to tell the AI what we want, after which it builds it as requested. For action objects, this is the era of vibe coding and current vibe design.

• Discovery: we don’t have to know what we want in advance, because we set out on a journey of exploration through an AI-created space of possibilities. When we come across something we like, we recognize that it’s good: then we have our result. For action objects, this will be the era of next-generation vibe design.

The three ways to create. (GPT Image-1)

Since the Enlightenment in the 18^th century, discovery has been prized. Now, AI is bringing discovery to the masses by allowing them to explore latent idea space. (GPT Image-1)

Grok “Imagine” Points the Way

Grok’s “Imagine” mode represents an early but significant step toward understanding AI as its own medium. Unlike traditional image generation tools that position themselves as replacements for cameras or brushes, Imagine mode embraces the unique affordances of AI creation.

The interface itself acknowledges that creation happens through navigation rather than construction. Users don’t build images but explore possibilities through conversational interaction. This seemingly simple shift has profound implications. The conversation becomes part of the creative act, not just instructions for it. Each exchange refines the navigation coordinates, with the AI serving as both vehicle and co-navigator through latent space.

Using Grok Imagine to explore the latent space of images where a robot serves burgers in a diner. We started with a “1950s robot,” but I navigated latent space in the “time” dimension by editing the text to request a “futuristic” robot instead. Since I liked the look of the upper left panel, I clicked it to move in the “color” dimension toward the last screenshot filled with neon-lit pink and cyan. (These color descriptors were added to the text prompt by the system — I just clicked a color scheme I liked.) After exploring these directions, I prefer the 1950s robot, so if I needed the image for a project (and not just for this example), I would backtrack and continue exploring in new directions.

What's particularly revealing is how Imagine mode handles iteration. Traditional tools frame iteration as “editing,” i.e., fixing problems, adjusting mistakes. Imagine mode frames iteration as exploration: each new variation isn’t a correction but a different path through possibility space. This reframing is essential to understanding AI creation. There are no mistakes in latent space, only different destinations.

I can see Socrates agreeing with me that there are no mistakes in intellectual exploration. (GPT Image-1)

The mode’s approach to style is equally telling. Rather than offering preset filters or effects, it allows semantic style description: “make it more melancholic” or “add a sense of urgency.” This isn’t just natural language processing; it's recognition that in AI creation, meaning and aesthetics are inseparable. The latent space doesn’t distinguish between what something looks like and what it means since form and content exist in the same mathematical manifold.

Our creations will adapt their look as we explore latent space with semantic descriptors, such as mood or occasion. Arriving at a great look is a journey of discovery. (GPT Image-1)

Most significantly, Imagine mode’s real-time generation makes visible what has always been true about AI creation: it’s a performance, not a production. Watching possibilities emerge, selecting paths in the moment, responding to what appears: this is closer to improvisation than traditional creative work. The tool acknowledges that AI creation is synchronous with AI generation, that the act of creation and the act of experiencing creation can’t be separated.

Creation and performance were always unified in ballet (an art form that only exists while being performed), but AI is now bringing an improvisational aspect, closer to jazz, to all creation. (GPT Image-1)

Controversially, at least among the more prudish AI influencers, Grok includes a so-called “Spicy” mode that moves in an erotically-charged dimension of the latent space. For sure, this mode can generate what’s euphemistically called NSFW, so use at your own risk. (Though even the “spiciest” results are softcore compared with what you can easily find on the Internet if you care to look for such things.)

To maintain my “safe for work” rating, I’m sharing one of Grok Imagine’s tamer creations from Spicy mode: the beginning, middle, and end of a video that was part of my exploration journey to find a K-pop avatar to perform one of my songs. (Grok)

For certain creative projects, Spicy mode is a godsend, and even for more staid projects, exploring in that direction can unlock new ideas, even if you may then have to backtrack towards the “safe” dimension within the space.

A key UX reason for Grok Imagine’s success is its fast response times, which encourage users to keep exploring. Quantity begets quality in discovery-based creation, and Grok offers quantity in abundance.

The UX of Exploration

Much as I like Grok’s Imagine mode as a current step toward exploration-based discovery, it is a baby step in that direction and lacks most of the features we would want to support users.

Proper exploration of the latent space is not linear; it is a forking process with many attempts to explore down a path that later is abandoned. Let’s call it pathcraft: the art of choosing and navigating exploration paths through the possible solutions.

The lack of UI support for pathcraft in current AI systems makes it feel more like you are jumping obstacles to get through a maze than as if you are freely navigating and exploring the latent space of design possibilities. We can do better. (GPT Image-1)

Currently, the main support for pathcraft, whether in traditional generative AI such as Midjourney or Grok Imagine, is based on the old-school Web browser’s Back feature: sometimes literally so (by clicking Back), and sometimes by scrolling back through an endless set of previous results, presented in sequence. The main shortcut offered by Grok is that one can “favorite” some results before moving on, and then scroll through a shorter list of thumbnails of past favorites.

Linear, linear, linear! A drag!

Current AI user interfaces are too linear without proper support for navigating the forkpath. (GPT Image-1)

ChatGPT uses its language capabilities to auto-name past results in the library and in the linear(!) list of past chats. These names provide a small usability assistance but fundamentally don’t solve the problem of navigating and exploring a multi-dimensional design space.

We need support for the forkgraph and for collapsing explorations into grouped categories. Branching and triage must become first-class UI features.

A second need that’s slowly being introduced, especially in video creation engines, is the Look Lock: the ability to specify that certain parts of the current creation should be kept invariant during moves to other regions of the latent space. This is highly requested for the use case of creating advertisements, where the product being advertised should always look like the real thing, even if it’s being rotated or held by a new model under different lighting conditions.

Similarly, if creating a wedding invitation, the bridal couple should always look like themselves. But Look Lock should also apply to different aspects of the latent space, such as the mood or a visual style that you want to retain while exploring other dimensions. An extremely primitive form of Look Lock is offered by the start frame and end frame features of almost all AI video makers.

Conclusion: A New AI

Early filmmakers famously placed a static camera in front of a stage, capturing a theatrical performance from “the best seat in the house.” Similarly, today’s creators often treat AI as a hyper-efficient illustrator or writer, providing directions for the AI to execute.

I have argued that proper use of AI as a creative medium requires us to move to a new paradigm. (Just like good filmmaking required a moving camera and the use of cuts and montage, replacing the initial single long stream.)

The true creative revolution lies not in refining prompts, but in reconceptualizing the act of creation itself. The core of AI creation is not a command-line chat interface but a vast, structured, and navigable universe of possibilities: a semantic map where every possible creation exists in latent space and only needs to be discovered.

AI isn’t here to execute directions but to be a co-creator that takes humans to new heights. (GPT Image-1)

Therefore, our paradigm should shift from one of instruction to one of exploration. The AI creator’s role is evolving from that of a client giving a brief to that of an explorer charting a new continent, navigating a universe of what can be made.

Let’s rethink AI as a new creative medium of its own, based on exploring latent space to discover our solutions. (GPT Image-1)