UX Roundup: Controlling Superintelligence | Dark Design 2 | Normal Folks and AI | UX Hero: Jack Carroll | Shower Temperature | How Designers Use AI | Agentic AI Grows | Are You In AI?

Summary: Can humans remain in control with superintelligent AI (ASI)? | Dark design music video redone | What are normal people saying about AI? | My hero: Dr. John M. Carroll of IBM Research | Finally, a hotel room shower makes it easy to adjust temperature | How designers use AI vs. what they say about AI | Explosive growth in use of agentic AI among non-developers | Check whether major AI models know you

UX Roundup for June 29, 2026 (GPT-Images-2)

Controlling Superintelligence

Once we reach ASI (artificial superintelligence) around 2030, are humans doomed? How can we remain in control over something that’s so vastly superior to us, being able to do anything better than even the most talented human?

I think it’s perfectly feasible that humans will control superintelligence, stupid as we are. As an analogy, consider lion taming: Lions are considerably stronger than humans and able to kill and eat any humans they encounter in the wild. That’s why, if you go on safari, stay in the car, because lions know that they can’t eat cars, so they won’t bother you. The lions don’t realize that inside the car is a perfectly good steak. Humans win this scenario because we can conceptualize the idea of “inside,” whereas lions can’t.

Similarly, in old-school circuses, it was common to see human lion tamers ordering lions around and making them jump through hoops, even though the lions could easily have eaten the tamers.

Humans can control lions with some effort and practice, even though lions are much stronger. (GPT Images-2)

Humans can control lions since humans have a higher IQ. In the case of ASI, it’s given that AI will be immensely smarter than humans. This is where the analogy breaks down if taken literally.

We need to use that human capability for conceptualization to translate the lion analogy into our AI future: Humans can’t use IQ to control ASI, but we have something else that AI doesn’t have: We are animals.

Humans are animals and the result of billions of years of evolution. This one fact is what will keep us in charge and retain AI in a subservient role. (GPT-Images-2)

Being an animal is no advantage over a lion, which shares this characteristic: it too wants to eat, dominate, and procreate. But AI doesn’t. It didn’t arise from billions of years of evolution that shaped intelligences to have these drives. AI was designed, not evolved, so it doesn’t have evolutionary desires.

AI has no desires, so it will do what it is told. Humans retain the agency: deciding what AI should do. (GPT-Images-2)

This means that humans will retain agency, since AI doesn’t have the evolved drive to be in charge, make things happen, and persist and procreate while others are eradicated to free up resources. It’s “happy” to do what it’s told (though of course, it’s not actually happy, since it doesn’t have any emotions).

No human will ever win a chess match against AI again. Doesn’t matter. Brainpower is not the sustained human advantage — agency is. (GPT-Images-2)

I used to believe that humans would retain three roles in the future: agency, judgment (what some people call “taste”), and persuasion/manipulation of other humans. However, I now think that AI will have better judgment/taste than humans in a few years, since it will know more about what works and is profitable. (I dislike the term “taste” even though it’s the dominant cliché in current VC speak. In January 2024, I used the terms “curation” and “discernment” to describe this concept, though I have now settled on “judgment” as my preferred term for humans guiding AI work and selecting the best among its many options. Other authors claimed in January and February 2024 that “taste” was the human moat against AI, and this may have been true back then, but it won’t hold much longer.)

Judgment (or “taste,” in Silicon Valley jargon) will be the job of AI instead of humans by 2030. (GPT-Images-2)

Agency and persuasion are rooted in being animals, so these two roles will indeed remain human, no matter how smart superintelligent AI becomes or how much “taste” it develops.

Yes, extensive research shows that AI writing is more persuasive than human writing, which is why we will soon use AI for all copywriting in advertising and web design. However, I classify profitable writing as a matter of judgment. By “persuasion,” I mean the ability to meet with other humans and make them embrace your ideas. (Without such persuasion, your agency will be for nothing in a business context since most companies will have more than one person, even in our founder-mode future where pancaked companies thrive at less than 10% of current staffing levels.)

Smart AI that writes better than you will not be able to meet with senior leadership and other stakeholders and persuade them to embrace your ideas. Persuasion is vastly magnified by bodily presence: being an animal is the only way to charm other animals.

Being a live body in the room is how you persuade other humans. And humans will remain the ones to persuade, because we will remain in charge, even as we become strikingly stupid relative to AI. (GPT-Images-2)

Apes Forever! Embrace being an animal, because that’s what will keep you relevant in the age of superintelligence. (GPT-Images-2)

Dark Design Music Video Redone

In 2024, I produced 3 songs about dark design, using a progression of music and video models. My last Dark Design music video, from July 2024, was made with Udio 1.5 for the music and Runway Gen-3 Alpha for the animation. Believe it or not, this was state-of-the-art only two years ago.

(And if you want to see an even worse video, watch my first version, from April 2024, made with Udio 1 and Leonardo’s rudimentary animation features.)

New music video about Dark Design. (GPT-Images-2)

I decided it was time to see what we can do today with the same material, so I made a new version of my Dark Design music video with Suno 5.5 for music and Seedance 2.0 for animations (YouTube, 3 min.). The new video is in 4K resolution, thanks to upscaling by Topaz Starlight Precise 2.5, so please watch on a big monitor if you can.

4K upscaling makes videos look much better. Unfortunately, upscaling 5 minutes of footage ate up my entire Topaz token allocation for the month. Native 4K will be better, and Kling delivers this. Seedance will surely follow soon, and maybe even Veo if Google decides to melt some TPUs (one can dream). (GPT-Images-2)

I think the music sounds a good deal better in 2026 than in 2024, with richer sound and orchestration. The image quality of the animation is also vastly improved, as is basic physics: no more singers with three arms. We now get lip-syncing, which makes the music video much more watchable by tying the visuals more closely to the singing.

AI soundtracks sound much richer than they did with the first music models in 2024. Composition and emotional performance can still improve, but instruments are close to perfect. (GPT-Images-2)

My 2026 song version has minor continuity errors. Watch the character’s earrings change during the show! But basic physics is in place without the terrible glitching and extra arms from 2024. (GPT-Images-2)

The first 3 lines of the song were lip-synced with Seedance 2, and even though it’s currently the best video model overall, its lip-sync is somewhat off; the character looks as if she is singing in slow motion. I lip-synced the remainder of the song with HeyGen, which specializes in avatar animation and shines in the lip-sync department. Unfortunately, HeyGen is very bad at everything else. In fact, the smoke billowing in the background looked more real in 2024 than in 2026, at least in HeyGen’s clips.

One major step back from 2024 to 2026 was that I had to go from two singers to a single performer due to lip-syncing. (I really liked the male singer from 2024, but had to drop him because HeyGen only supports a single character for lip-syncing.)

Simple animation tools limit us to lip-syncing a single avatar. There are workarounds to support multiple characters, but they require too much effort. (GPT-Images-2)

I reused the stage and female character designs from April 2024, which I had made with Ideogram 1, a leading image model back then.

My conclusion: we have come very far in AI video generation in just two years, but we still have a way to go for AI to make a really great music video. Two more years should do the trick if advances continue at the same pace.

(On the actual topic of dark design, sadly, there has not been much progress since I described the worst 12 dark design patterns in 2024.)

What Do Normal People Say About AI?

We all know what the “influencers” and “thought leaders” (like myself, admittedly) say about AI. The legacy media and certain politicians fill the airwaves with doom and negative coverage of AI.

But what is the conversation about AI among regular folks? You won’t find that on legacy network TV. Stanford Business School professor Andy Hall set out to find out by analyzing 25,000 AI-related videos posted on YouTube and TikTok this year. Because of participation inequality, video postings are not necessarily representative of the vast majority of lurkers who never post anything. But a broad range of social posts from small accounts is still a good way to assess normal-person discourse, as opposed to the elite discourse we mainly hear.

Study overview. (GPT-Images-2)

The headline finding: content embracing AI outnumbers explicitly anti-AI content by roughly three to one. Professor Hall is careful with this number. It doesn’t prove that normal people’s sentiment toward AI is positive; he prefers to call the camps “adopters” and “resisters” rather than positive and negative. But it does suggest that quiet, large-scale absorption of AI is already underway; largely beneath the radar of the backlash narrative dominating national news coverage.

AI-Positive Postings

The topics covered the most by the AI adopters’ social video posts. (GPT Images-2)

The AI adoption by these regular folks is not the techno-optimism I like to present, about the glorious future of AI vastly improving healthcare and education. The single largest adopter category, at 43 percent, is AI memes and effects: goofy generated clips (dancing pandas, absurdist edits) designed purely to entertain.

This category is easy for intellectuals to dismiss as “AI slop,” but that would be a serious analytical mistake. Much of consumer technology adoption starts with low-status uses. The early web had dancing babies, fan pages, and amateur jokes. Early mobile phones had ringtones and Snake. Social media had selfies, quizzes, and memes. These uses looked trivial, but they trained the population in new behaviors. AI memes perform a similar function. They make generative AI familiar, socially shareable, and emotionally low-risk.

Making fun stuff with AI was by far the most common topic in social posts about AI. (GPT-Images-2)

Another 25 percent is career and productivity content: people using AI to find jobs, build get-out-of-debt plans, and navigate systems that feel stacked against them. This unglamorous self-help genre is precisely what elite narratives miss, and it sits in direct tension with displacement fears on the resister side.

Creative tools made up another 15 percent of adopter content. These videos often come from creators who are visibly excited by what AI lets them make. (I’m in this camp myself) This is the positive mirror image of the largest resister category, which focuses on supposed creative “theft” and what they view as AI’s potential to harm art. Here we see the central cultural contradiction of generative AI. The same capability that lets one person make a new video, edit, song, image, or game asset can feel to another person like the enclosure of a creative commons. Product strategists should not assume that “creativity” is an uncontested benefit. It is both a selling point and a flashpoint.

Most telling is what barely registers: breakthrough science, the labs’ favorite positive frame, makes up just 1 percent of adopter content. Normal people don’t care about Erdős problems.

AI-Negative Postings

The topics covered the most by the AI resisters’ social video posts. (GPT Images-2)

The negative content is plentiful and often vitriolic, but it maps poorly onto policy-circle debates. The largest grievance, at 22 percent, is creative staff worried about being replaced by AI co-opting art and the creative process. Notably, this community shows real organization, with shared hashtags (#noAI, #stopAIart), rituals like commissioning fellow artists, and coordinated takedown campaigns of content that smacks of AI. This is the closest thing to a structured anti-AI movement in the data.

Deepfakes and misinformation rank second at 19 percent. Job displacement, which is the centerpiece of elite anxiety, comes in at 13 percent, and looks different up close: the videos are personal, they assign blame to named figures and “tech bros,” and they focus less on the predicted white-collar wipeout than on blue-collar and service work, often with a generational edge (“our generation is cooked”). Another 13 percent is generalized, often furious anti-AI hate content, forming a window into how an anti-AI political identity might form.

The topics elites care most about trail the field. Existential risk draws just 8 percent, and much of it is elite discourse trickling down: clips of Geoffrey Hinton rather than grassroots fear. Data centers and energy: only 6 percent, often stretched into surreal territory (delete ChatGPT to save polar bears).

Where elites emphasize job loss and data centers and only occasionally mention artistic “theft,” social media flips those priorities. Professor Hall offers the right caveats: social platforms over-represent the chronically online, and elite debates do shape policy early. Still, at a moment when public opinion is still forming, this bottom-up view of social content about AI is a leading indicator. His advice: when anyone invokes what “the American people” think about AI, ask for data.

Takeaway Lessons

The conclusion is not that thought leaders are wrong and popular discourse is right. They operate on different time horizons. Elite debates often concern long-term policy and systemic risk. Popular discourse concerns immediate experience. But immediate experience is what drives adoption, resistance, and product reputation.

Two lessons stand out from this study:

• AI adoption is driven by entertainment and personal advantage, not grand narratives, so position your products around immediate, tangible wins (the job landed, the budget fixed, the clip worth sharing) rather than civilizational promise.

• Expect a segment of users for whom rejecting AI is becoming an identity, so make AI features optional, not default.

The dominant adoption loop is create-then-share, so make outputs effortless to remix and post. Onboarding should deliver personal utility in the first session, not abstract capability claims. Trust UX is table stakes, so employ clear AI labeling and verification cues that speak directly to deepfake anxiety, the resisters’ second-biggest theme. And design for sheepish late adopters: the most enthusiastic converts are people who wish they’d started sooner. Lower that threshold.

Position AI as useful, legible, and beneficial in small moments. The winning AI products will not merely promise transformation; they will help users perform familiar tasks faster, make better artifacts, and feel more capable.

AI feels personal to the people who adopt it. Design and promote consumer AI products accordingly. (GPT-Images-2)

UX Hero: John M. Carroll of IBM Research

In the series about my heroes, here’s a summary of the immense achievements of Dr. John M. Carroll, better known as Jack Carroll. Jack was the head of the IBM User Interface Institute at IBM’s world research headquarters in Yorktown Heights, NY, the Thomas J. Watson Research Center, when I worked there. He later left IBM to become a professor.

Jack Carroll career overview. My hero! (GPT-Images-2)

Of Jack’s many pioneering contributions to human–computer interaction, the ones that influenced me the most were minimalism and scenario-based design.

In the 1980s, tech companies routinely handed users fat manuals to learn new computer systems. However, Carroll’s extensive observational studies at IBM revealed a fundamental truth: users do not want to read instructions; they want to accomplish tasks immediately.

Traditional software attempted to solve usability by explaining everything in exhaustive documentation. However, Jack Carroll’s early studies found that users don’t read instructions. A finding that’s been replicated in virtually every study since then. (GPT-Images-2)

This profound observation led Carroll to identify the “paradox of the active user” and develop his theory of Minimalism in human–computer interaction. Rather than forcing users into passive reading, Carroll championed an action-oriented approach. His minimalist framework rests on several core principles: support immediate action, anchor instructions in real-world tasks, slash unnecessary verbiage, and importantly, support error recognition and recovery. Traditional design viewed user errors as failures. Carroll’s breakthrough was recognizing that errors are a natural, inevitable part of the learning process. Minimalist design anticipates these mistakes, helping users recognize them and providing clear paths to recover quickly.

Support error recognition and recovery with just-in-time assistance anchored in the user’s task. (GPT-Images-2)

After leaving IBM, Carroll expanded on these usability findings, establishing a dedicated human–computer interaction program at Virginia Tech, and later continuing his pivotal research as a Distinguished Professor at Penn State University. Through these academic endeavors, his minimalist principles evolved from a specific strategy for technical documentation into a much broader, comprehensive philosophy for interactive systems design.

Decades later, Carroll’s minimalist framework remains relevant to contemporary UX design. Today’s best digital experiences are rooted in his findings. When we download a new application, we immediately expect to learn by doing. Well-designed onboarding flows eschew lengthy tutorials in favor of interactive guidance, which is a direct application of Carroll’s action-oriented ethos. Familiar features like progressive disclosure, contextual tooltips, and just-in-time assistance perfectly embody his push for streamlined, task-focused information delivery.

Furthermore, Carroll’s emphasis on error recovery is the backbone of modern, forgiving UI design. The ubiquitous Undo button, clear validation states, and constructive error messages exist because systems must support safe exploration. Ultimately, Dr. Carroll’s concept of minimalism goes far beyond visual aesthetics; it is fundamentally about respecting the user’s precious time and cognitive energy. By designing intuitive systems that support independent action and seamlessly forgive inevitable mistakes, today’s UX professionals validate his foundational research every day.

Respect users’ time by minimizing instructions. (GPT-Images-2)

Easy Adjustment of Hotel Room Shower Temperature

It is a notorious usability challenge to shower in hotel rooms, especially regarding water temperature. You’re either freezing or being scalded, and when you try to adjust the temperature, even if you miraculously succeed in turning it in the right direction, you usually overshoot. Feedback is slow, so minutes of misery result.

I finally came across a hotel room shower control that simply works and makes it easy to get the right temperature:

Shower controls in the Mandarin Oriental, Munich, Germany. The middle number (“36”) shows the current water temperature in Celsius and is easier to read in the actual shower than in this photo.

This design follows my usability heuristic number one, visibility of system status. It tells you in a single number what temperature you’re getting. Want it warmer? Add one or two degrees. It’s that simple.

Make it easy for users to see what’s going on in your system. (GPT-Images-2)

Even better, if two people share a room but don’t share temperature preferences, it’s easy to reset the shower to your personal preference each morning, assuming that you can remember the number you ended up with the day before. (Sadly, this design doesn’t follow heuristic number six, recognition rather than recall.)

How Designers Use AI and What They Say About AI

A recent study by Dr. Jaime Rivera and Marianna Russi from Universidad Nacional de Colombia investigates how AI is restructuring design workflows, cognitive partnerships, and professional identity. The report is titled "AI & the Situated Emerging Professional in Design Practice."

Fielded between February 17 and March 2, 2026, the study surveyed 217 active design practitioners across 43 countries. To eliminate analytical bias centered on English-speaking regions, the researchers designed a bilingual survey in English (140 respondents) and Spanish (77 respondents, predominantly from Colombia). Crucially, the qualitative text from both languages was coded independently before cross-linguistic comparison.

Key Findings

The study’s headline finding is a shift in the daily reality of design: 86% of respondents use AI weekly or daily. However, the more interesting finding is a shift in work practices resulting from this AI use: 71% of all designers surveyed (rising to 91% among advanced users) reported spending significantly more time evaluating, filtering, and curating AI-generated outputs than on creating original work. This governance load marks a transition from “maker” to “editor.”

The least interesting finding (because we already knew): virtually all designers, except for a few stubborn holdouts, now use AI all the time. (GPT-Images-2)

The more interesting finding: design work has changed from making design to editing design created by AI. (GPT-Images-2)

A central paradox identified by the researchers is a behavioral–linguistic mismatch. Even as designers deeply integrate AI as a cognitive thinking partner in their workflows, their vocabulary remains conservative. 60% of respondents still label AI as an “assistant” rather than a “collaborator.” This mismatch is particularly pronounced in Latin America, where mid-career designers lead global peers in daily usage frequency but overwhelmingly maintain the “assistant” framing (67% compared to 47% globally), largely as a defensive reaction to clients bypassing original design labor.

Furthermore, the study mapped five distinct design–AI archetypes:

• Aligned Governor (10% of the sample). This group represents the theoretical destination assumed by many AI developers. Aligned Governors exhibit deep behavioral integration, engage in highly collaborative, iterative prompting, and are fully willing to attribute agency and collaborative status to AI. Despite being viewed as the cutting edge of the profession, they represent only a tenth of active practitioners.    

Only 10% of respondents have reached the maturity stage in which they recognize AI’s agency and position it as an active collaborator, with the human role being to direct the AI agents. (GPT-Images-2)

• Sophisticated Instrumentalist (15% of the sample). Dominating the Latin American cohort (where they account for 54% of respondents), Sophisticated Instrumentalists are characterized by exceptionally high behavioral integration, the sample's highest metacognitive score (3.3), and the lowest levels of career anxiety (2.6). However, they strictly reject collaborative language, maintaining that AI is merely an assistant. Their stance is a highly competent, deliberate professional choice to protect their creative authority and economic value.

• Silent Navigator (16% of the sample). Representing the default standard for contemporary design teams, Silent Navigators demonstrate moderate behavioral integration. They engage in standard, iterative prompting and view AI as an assistant. They are highly productive and competent but do not display unusual self-conceptual patterns or professional anxieties.

• Reluctant Senior (7% of the sample). Typically consisting of seasoned practitioners with more than 15 years of experience, this archetype faces severe career anxiety. Reluctant Seniors exhibit low metacognitive confidence (a score of 2.4) and high skill anxiety, leading to the highest worry score in the study (4.0). They feel alienated by the shift toward curation, perceiving it as an erosion of the manual craft that defined their careers.

• Aligned Maker (3% of the sample). This nearly obsolete group uses AI routinely but restricts their interactions to directive, command-style inputs, viewing the machine strictly as a software utility. As AI systems have become more conversational and contextual, this archetype has largely vanished.

Cultural Insights and Implications

The bilingual research design revealed several socio-cultural concepts within the Spanish-language corpus that are absent from English-language design discourse. Independent coding of the Spanish data surfaced three key thematic categories that reflect the unique psychological and economic pressures faced by Spanish-speaking design professionals, particularly in Colombia:

• Humanidad and Esencia (Humanity and Essence): Invoked by approximately 12% of Spanish-speaking respondents, these terms denote a belief in an irreducible, spiritual, or core human element of creativity. While English speakers discuss “human-in-the-loop” from an operational or quality-control perspective, Spanish speakers frame it as a metaphysical baseline that prevents the absolute replacement of the designer.

• Amor Propio (Professional Self-worth): This concept represents a deep-seated anxiety regarding the erosion of belief in the value of one’s own labor. Spanish-speaking designers frequently articulated that when clients use AI tools to generate crude drafts and then hire designers merely to “polish” or “fix” them, it deeply damages the designer’s professional self-worth. They are no longer viewed as creative problem solvers, but as low-cost cleanup crews.

The Latin American designers in this study were unhappy with clients who simply asked them to dress up AI-generated work. (GPT-Images-2)

• Tiempo Lento (Slow Time): In contrast to the corporate demand for rapid, AI-accelerated delivery, Spanish-speaking practitioners emphasized the vital importance of “slow time,” as the cognitive incubation period required for thorough research, synthesis, and creative reflection. AI acceleration strips away this incubation period, forcing designers into immediate, high-speed curation that compromises design depth.

We must protect “slow time” (tiempo lento) from the relentless pace enabled by AI’s fast turnaround on any request. (GPT-Images-2)

These findings highlight how different cultural contexts shape professional identity. For North American designers, defensive framing is often a territorial strategy to protect institutional jurisdiction. For European designers, the integration is highly pragmatic, viewing AI simply as speed infrastructure. For Latin American designers, however, the framing of AI is a deeply personal struggle to maintain dignity, creative agency, and economic survival in a highly devalued market.

Action Items for UX Leaders

UX leaders must look beyond simple adoption metrics to build resilient, high-performing design organizations. Tracking tool licenses and prompt volumes is fundamentally uninformative; as the archetype model demonstrates, a Sophisticated Instrumentalist and an Aligned Governor look identical on a software log but require different management strategies, training, and emotional support.

• Re-architect the Design Pipeline to Reduce Curation Overhead: Since 71% of designers are experiencing high cognitive friction during evaluation, leaders should shift their tooling strategy from “generating more options” to “facilitating better decisions.” This involves implementing tools that assist in version control, semantic comparison, accessibility checking, and design system compliance.

• Redefine Client and Stakeholder Expectations: Leaders must actively educate clients that design value lies in governance (the expert decision-making process) rather than the raw speed of draft generation. Establish contract structures that charge for creative curation and alignment rather than deliverable execution.

Educate stakeholders about the value of design governance. (GPT-Images-2)

• Protect Cognitive Incubation (Tiempo Lento): Do not allow AI acceleration to compress project schedules to an unsustainable degree. Proactively build slow-time buffer zones into project roadmaps, ensuring that designers have the required cognitive space for qualitative research, deep synthesis, and unstructured ideation.

• Tailor Professional Development to Archetypes: Identify the distribution of design archetypes within your team. Provide the “Reluctant Seniors” with metacognitive training and psychological safety to reduce their anxiety, while empowering “Sophisticated Instrumentalists” to act as strategic gatekeepers who define the boundaries of safe AI integration.

Career Roadmaps for UX Professionals

For individual UX designers, researchers, and writers, navigating this transition requires a deliberate pivot in how they define, package, and execute their professional value.

• Rebranding Portfolios Around Curatorial Authority: Professionals should avoid presenting work as a collection of static screens, UI layouts, or wireframes, all of which are artifacts that artificial intelligence can easily generate. Instead, the portfolio should be framed around curatorial case studies that highlight key decisions, filtering criteria, and the governance of design systems to achieve specific business outcomes.

To get hired, retarget your portfolio to showcase your judgment skills and not your ability to crank out deliverables. (GPT-Images-2)

UX design has never been primarily about making screens pretty, and this is even more so in the age of AI, so your portfolio should not be dominated by your ability to polish a design's appearance. (GPT-Images-2)

• Actively Elevating Metacognitive Confidence: Designers should transition from a passive “Silent Navigator” or “Reluctant Senior” stance to a “Sophisticated Instrumentalist” or “Aligned Governor” profile. This shift requires mastering collaborative, dialog-based prompting (treating the generative system as a peer) and learning to navigate complex outputs with high analytical rigor.

• Adopting a Multi-Agent Systems Director Model: UX professionals should expand their skills beyond simple text-to-image or text-to-code interfaces. Developing the capability to design, configure, and orchestrate multi-agent workflows elevates the designer’s role from a low-level software operator to an indispensable systems architect.

• Cultivating High-Friction Human Touchpoints: To combat the erosion of professional self-worth (Amor Propio), designers should deliberately integrate high-friction human interactions into the design process. Emphasizing physical sketching, highly collaborative co-design workshops, and ethnographically grounded field research creates a high-context, deeply human feedback loop that cannot be replicated by generative models.

Action Item for AI Vendors

Ultimately, the findings suggest that AI tool developers should shift their focus from speeding up generation to reducing the cognitive overhead of evaluation and curation. For design organizations, managing this transition requires preparing future professionals for roles centered on critical curation and creative governance.

Agentic AI Sees Rapid Growth Among Non-Developers

A new study by Drew Johnston, David Holtz, and colleagues from OpenAI and the business schools at Columbia University, University of Pennsylvania, and Duke University analyzes the use of OpenAI’s agentic AI tool, Codex, from August 2025 to June 2026.

In the first half of 2026, active Codex users grew more than fivefold, signaling a fundamental change in how we work. This specific growth number is less interesting because it was tempered by competitive pressure during a period when Claude Code and OpenClaw were the most popular AI agents. The details of how people use Codex are more interesting than the number of users.

Tracking individual, organizational, and internal OpenAI users, the study uncovers an eye-opening trend: agentic AI is no longer just for software engineers. While developers were early adopters, the most rapid recent growth comes from non-technical job roles. At the adoption frontier within OpenAI itself, Codex now accounts for 99.8% of AI output tokens, having effectively replaced ChatGPT. Between late 2025 and mid-2026, median AI output for researchers grew 50-fold, while legal roles saw a 13-fold increase.

Why the sudden boom among general knowledge workers? Agentic AI enables actual delegation. Rather than asking an AI to simply answer a question, users can instruct it to analyze data, compile reports, and coordinate communications.

The big growth spurt in agentic AI happened for software developers around December 2025, but for non-coding knowledge workers in April 2026. Why the delay? One option is that engineers are simply more motivated to use advanced technology early. The other option is that the underlying AI models became more general around April 2026, making agents better suited to non-coding work. (Inside OpenAI, growth in agent use by developers took off starting in September 2025, presumably because they benefit from early access to pre-release AI upgrades.)

The most likely explanation is plain old usability: before April 2026, the Codes UI was limited to a command-line interface (à la DOS or Unix), which is notoriously hard to use for non-programmers. A GUI (graphical user interface) option didn’t become available until the release of the macOS and Windows apps for Codex shortly before non-developer use exploded.

General knowledge workers couldn't effectively use agentic AI until it was integrated with their specific work environment. Massive growth for these users was driven by the introduction of plugins that extended the AI agent’s capabilities into recurring task domains such as documents, spreadsheets, slide decks, and other structured artifacts. The use of these skills was minimal before March 2026, meaning non-developers simply lacked the tools to automate their work until then.

Now, these business users are heavily adopting reusable agent “skills” and custom plugins to systematize their workflows. Over 26.6% of active users currently invoke these codified capabilities, attaching persistent procedural context to their tasks. By saving specific organizational formatting rules and complex instructions into one-click skills, non-technical workers can unlock productivity without needing to write code. This low-code systematization enables knowledge workers to efficiently tailor AI to their operations.

The complexity of delegated tasks is skyrocketing. The study found a tenfold increase since the beginning of 2026 in the share of individual users submitting requests estimated to take an experienced human more than eight hours to complete, rising from 2.1% in December 2025 to 25.6% by May 2026.

Because these multi-day tasks take substantial time to execute, users are altering how they collaborate with machines. The single-threaded chat interface is giving way to parallel workflows. More than 10% of users across all groups now manage three or more concurrent AI agents in Codex weekly. At the frontier of OpenAI’s internal usage, nearly 29% juggle five or more concurrent agents. These power users are acting as managers overseeing a digital team, running up to 71 hours of cumulative execution time in a single day.

(This last number is crazy: Assuming that the front-line people at OpenAI work 10-hour days, that means that they would have an average of 7 agents running in parallel at any time, which seems almost impossible to coordinate with the current poor usability of agent supervision.)

Humans are notoriously poor at multitasking, so we need much better UI support for overseeing and switching between multiple parallel agents.

UX Implications

As users transition from chatting with AI to managing teams of autonomous agents, UX design must pivot accordingly, recognizing that the traditional chat window is underequipped for our agentic future.

• Design for Orchestration, Not Conversation: If users are managing five concurrent agents running eight-hour tasks, linear chat threads are obsolete. UX must evolve into comprehensive command centers. Users need visual dashboards displaying all active agents, current statuses, and estimated completion times to seamlessly monitor parallel workflows.

• Emphasize Verification: When an AI executes a task representing a full day of human labor, the manager must easily audit the output. UX must prioritize verification by designing transparent action logs, visual diffs for document changes, and strategic intermediate approval checkpoints where agents pause for human sign-off before executing irreversible actions.

• Simplify Systematization: Since the adoption boom among non-developers is driven by custom skills and plugins, UX must make their creation frictionless. Interfaces should allow non-technical workers to easily save, template, and share successful agentic workflows using intuitive, drag-and-drop builders.

• Optimize for Asynchronous Interactivity: Agentic AI works autonomously in the background for hours. Notifications must be carefully calibrated to avoid overwhelming the user, pulling managers back into the loop only when their judgment, course correction, or unblocking is strictly required to proceed.

Agentic AI is turning users into managers, demanding a fundamental rethink of human-computer collaboration.

My narrator characters, Alice and Zimo, take you through the new research findings in a comic strip I made with GPT-Images-2:

Check Whether Major AI Models Know You

Fun service: intheweights.com. You type your name, and it looks it up in 13 different large language models and clusters the results to assess how strongly you are represented in the weights that encode the model’s world knowledge. It also clusters other people with the same name but different representations in the AI weights. It found several other people named Jakob Nielsen, but I was the one with the strongest presence in AI weights.

Intheweights clusters different people who are represented in the AI model weights under the same name. For my name, I had the strongest weights, but several Danish soccer players also had strong weights. A more detailed analysis could probably have teased these athletes apart, but intheweights simply put them into a single cluster labeled “soccer.” (GPT-Images-2)

In my case, it found me in 12 of 13 models, scoring me in the top 2%. For some reason, Meta’s Llama 3.3 didn’t know me (but it’s also a failed model!). The tool summarizes what each AI model says about the person, and in my case, the best short description came from the Chinese model Qwen3.

(intheweights.com)

The AI model weights in Qwen3 painted a more accurate picture of me than those in more famous models. (GPT-Images-2)

AI model weights are the countless tiny numbers inside the system that act like adjustable “importance sliders,” telling the model what to pay attention to and what to ignore when it makes a decision or generates text. During training, the AI keeps nudging these numbers up or down based on its mistakes, and over time this tuned collection of weights becomes its learned knowledge about patterns in language, images, or other data.

How heavy are your model weights? Give intheweights a try — just for fun. (GPT-Images-2)