UX Roundup: Designers in Denial | 3 Levels of Corporate AI Use | AI Improves Medical Diagnosis | AI Math | Ideogram Character Consistency | Bad Study Methodology

Summary: Many UX designers are still in denial about AI | The 3 different types of AI use in companies, and their implications for your career | AI improves medical diagnosis in real-world clinical use | AI does mathematics differently than humans | Ideogram improves character consistency | Leading questions in usability studies

UX Roundup for August 4, 2025. (GPT Image 1)

All Things Design Podcast

I was on the All Things Design podcast (YouTube, 51 minutes)

Here are some of the points I made in the podcast:

Designers in Denial: A Dereliction of Duty

I made a direct accusation that the UX profession has collectively abandoned its responsibility to the world regarding AI. I see a widespread “AI denial” among designers who prefer their old, comfortable methods. This is not merely a missed opportunity; it is a failure to uphold our duty to the billions of users who will have to navigate this new world. I urged designers to “stop sitting in a corner and sulk,” and to start leading the critical transition from command-based interfaces where the user gives explicit instructions, to designing for the new paradigm of intent-based interactions, where the user specifies a goal and the AI figures out the steps. Sounds elegant, but most folks are terrible at articulating their intent, and even we intellectuals overestimate our skills there. We have a responsibility to shape this future, and right now, we are failing. Drawing from past revolutions, such as the web or GUIs, resistance merely leads to obsolescence. We can’t deny it; we have to play with the technology as it evolves.

This might be controversial because I was not mincing words. I issued a stark and uncomfortable challenge to the professional identity of my audience. For a field that defines itself by user advocacy, this is a harsh critique. But it is an interesting point because it frames AI not as a threat or a tool, but as our next great moral imperative.

The “Pancake” Organization and the 10x Productivity Leap

I predicted the rise of the “pancake” organization. I stated that AI will make product teams 10 times more productive within about three years. A team of 10 will do the work of 100. The inevitable result of this hyper-efficiency is a flattening of corporate hierarchies. The multiple layers of management we have today become wasteful overhead. Therefore, the traditional management ladder — until now, the primary career path for senior practitioners — will disappear. I said that aspiring to be the Director or VP of UX is a flawed ambition because most of those jobs simply won’t exist.

I know this is controversial. It directly attacks the established model for career advancement in our field. I am telling senior designers and managers that their path to success is becoming a dead end. This is unsettling, but it is the logical consequence of technology evolution. It is interesting because it forces a necessary re-evaluation of what a successful design career looks like: it will be defined by deep, skill-based expertise, not by the number of people who report to you.

Big enterprises will cut design teams in half, but mid-sized and small companies can finally afford UX now, expanding the market hugely. Drawing from history, like how the dot-com era favored interactive design over glossy ads, this efficiency will let us do way more good design across industries.

AI Supremacy: Your Hard Skills Are Expiring. Agency Is the New Core Human Skill.

I made the point that the “hard skills” designers have spent their careers perfecting are on a fast track to obsolescence. Craft-based tasks like creating mockups, perfecting typography, and even running usability studies will be done better and faster by AI within 5 to 10 years. The truly valuable human skill, and the one I emphasized repeatedly, is agency. This is the proactive ability to identify opportunities, define the correct user intent, and strategically direct AI to achieve a goal. This is what will differentiate valuable contributors from simple AI operators.

This is a profoundly difficult message for many to hear. I am telling highly skilled practitioners that the very foundation of their craft (the tangible work they produce) is becoming a commodity. This is controversial because it threatens their professional identity. It is interesting because it signals a pivot to a more abstract, strategic role. The future is not about your ability to use Figma; it is about your judgment and your ability to make things happen.

History rhymes: mainframes scoffed at GUIs, early web was dismissed as geeky, but they became essential. Same with AI, which has its problems now, but is getting better at an even faster pace than the web once did.

Buy Your Own AI or Quit Your Job

Fourth, I offered some stark, practical advice for designers trapped in slow-moving, AI-resistant enterprises. I told them not to wait for their company. They should buy their own AI subscriptions now, which is cheap “unemployment insurance.” I suggested they use these tools on their own time, or even hide their use from management, to gain the necessary experience. And I gave even more aggressive advice: for those serious about their future, they should consider quitting their stable job to join an “AI native” or “AI first” company, even if it means a pay cut, because the experience gained there is invaluable.

This is a controversial stance because I advocate for employees to be proactive and take significant career risks. I place the full responsibility for adaptation on the individual, because my experience shows that legacy organizations will simply not move fast enough to save their employees’ careers. I offered a direct, actionable, and urgent prescription that cuts through corporate inertia and highlights the personal stakes of this technological shift.

Software Ate the World. Now, Design Will Eat Software.

Finally, I offered an optimistic, if disruptive, vision for our field's future. I took Marc Andreessen's famous line (“software is eating the world”) and asked the next logical question: “Who's eating software?” My answer is design. As AI makes implementation trivial and cheap, the bottleneck for creating value is no longer engineering. It is a matter of figuring out the right thing to build. Defining user needs and specifying the correct intent—the core of design—becomes the single most critical activity. This will lead to an explosion of design work in countless niche industries that could never afford it before.

This is an empowering and interesting message because it elevates the strategic importance of design above all else. It is also controversial, however, because it implies a fundamental restructuring of the job market. The future for most designers may not lie in big, stable tech companies, but in highly specialized roles within a vast number of smaller companies across every industry imaginable. This requires a new career mindset focused on domain specialization.

5 themes I discussed in the All Things Design podcast. (GPT Image-1)

Three Levels of AI Use in Companies

The coming age of artificial intelligence is not a gentle evolution. It is a tidal wave, and most companies are either building sandcastles or, worse, turning their backs to the sea. For knowledge workers, the single most important career decision you will make is not your role, but the type of company you join. Your future employability depends entirely on the rate at which you learn AI, and this rate is dictated by your company's fundamental integration of AI.

We can classify companies into three distinct tiers based on this integration. Do not be fooled by marketing claims or C-suite pronouncements; observe the operational reality.

Three types of companies, classified by their use of AI. (GPT Image-1)

⭐⭐⭐ AI Native: These companies are born of AI. They were founded from scratch with the core assumption that AI would permeate every product, feature, and internal process. There is no legacy business to overcome. AI is not a tool they use; it is the medium in which they exist. For employees, every task is an opportunity to co-evolve with advanced AI systems. The learning is constant, deep, and structural because the very workflow is an AI-human hybrid.

Building is everything in an AI-Native company because they recognize that we don’t know how to use AI in business optimally. We have to invent our own future, and these companies are doing it. Surfing the wave of change, instead of resisting, is the way to improve fast, since every iteration of a new build gains more learnings. (GPT Image-1)

⭐⭐ AI First: These are legacy companies making a valiant, but compromised, pivot. Management has declared AI the top priority and allocates massive resources to AI initiatives. However, they are shackled by decades of inertia. For every new AI process, there is a legacy system it must integrate with. For every executive pushing for change, there are two middle managers who quietly resist because their old skills are being devalued. Working here means you will learn about AI, but you will spend at least half your time fighting organizational friction instead of building the future. It’s a step up from the past, but it remains the slow lane.

Working in an AI-First company that has supposedly transitioned from its old-school ways means that you still must devote substantial efforts to overcome inertia from legacy processes and (worse) legacy middle management. You can do it, but it’s a pain, and correcting the past diverts attention from your professional growth toward the AI future. (GPT Image-1)

⭐ AI Forward: This is the most deceptive category and a dangerous illusion of progress. It often manifests at the team level, not company-wide. Enlightened management provides budget for AI tools and looks favorably on experiments. This feels progressive, but because AI is not a strategic imperative, the core business remains unchanged. You might get a ChatGPT Pro license, but your main job is still to update the same old reports. This is a career holding pattern, giving you just enough exposure to AI to make you feel current while you fall behind your peers in AI-Native and AI-First jobs.

AI-Forward companies support the use of AI tools to expedite individual tasks, but their legacy workflows usually remain unchanged. (GPT Image-1)

The conclusion for your career is simple and stark. To remain relevant in an era that will be defined by superintelligence, you must maximize your learning velocity. This is only possible in an environment where AI is not an initiative, but the foundation. Roughly speaking, one year in an AI-Native company is worth two years in an AI-First company and five years in an AI-Forward company. Remember, five years is all you have to pivot your career before superintelligence takes over in 2030. (Of course, if you stay in an AI-Denying company, you will be unemployed come 2030 and never catch up with those of your colleagues who spent the transition period learning.)

Your best option is to join an AI-Native company. Here, your skills will grow at the pace of the technology itself. Your only viable second choice is an AI-First company, where you will at least be part of the main battle, even if it’s an inefficient, uphill one.

To willingly stay in an AI-Forward company is to accept a managed decline. And as for the laggard companies still in denial about AI? Flee. Staying is career suicide. The cost is not just your current job, but your entire future.

Sadly, it will likely soon be impossible for AI-Native companies to start up in my native Europe, unless they abandon their misguided AI Directive, which puts immense bureaucratic barriers in the way of AI innovation. Europe is on the straight path toward irrelevance for the AI-fueled world economy of the 2030s. However, even if you’re European, your individual career doesn’t have to be doomed, because AI-First companies should still be available in the EU.

The cozy sandcastles of the legacy world won’t survive the AI wave. To mix metaphors, you must be the third little pig and build your house from the bricks found in an AI-Native, or at least AI-First, company. (GPT Image-1)

My advice for your career. (GPT Image-1)

AI Improves Medical Diagnosis in Real Clinics

We have endless studies showing that AI performs better than human physicians at diagnosing patients in a research setting. But so far, most studies have been artificial, in order to create a controlled research experiment: clinical information was collected before the research and then given to either humans or AI (or human doctors working with AI), to see who would deliver the most accurate diagnosis, when given the same data.

However, for AI to help real patients, it must improve treatment under real-world conditions: a new patient walks into the clinic and complains about something, and AI must enhance that patient’s diagnosis and treatment. It’s not enough for AI to do better when presented with a data set of previously-collected information. Human patients, not data sets, are what medicine is supposed to help.

A new study does exactly that. Penda Health⁠, a primary care provider in Nairobi, Kenya, analyzed 39,849 patient visits across 15 of its clinics to measure the outcome of providing clinicians with real-time AI assistance during patient visits. Clinicians with AI assistance delivered a 16% relative reduction in diagnostic errors and a 13% reduction in treatment errors compared to those without.

The “AI Consult” product operates as a safety net that runs in the background of a patient visit to identify potential errors. It was iteratively designed with clinicians, providing outputs with green/yellow/red severity and issuing alerts only when needed to prevent harm.

Interestingly, clinicians learned during the study to avoid “red” alerts from AI Consult before the tool even provided feedback, suggesting clinicians were using it to improve their own practice: one called it a “learning tool,” and another said it “broadens my knowledge.”

We’re starting to see real-world clinical use of AI to improve healthcare for real patients. (GPT Image-1)

The study findings of a 16% improvement in diagnosis are less impressive than the tests of AI versus human doctors in theoretical settings outside the clinic. When AI and physicians are both given case study write-ups, AI usually outperforms humans by much greater margins. To me, this larger theoretical advantage for medical AI indicates that clinical use still has ways to improve, as we get more real-world experience with how to integrate AI with clinical practice. (And, of course, superintelligent AI in a few years will do even better. However, even if all progress in AI foundation models stopped today, progress in the practical use of AI will advance, since we’ve only scratched the surface so far.)

One reason for the success of the “AI Consult” product is the use of the user-centered design process, at least to some extent. (I don’t know the full details of how it was designed, other than the twin points that they used iterative design and relied on clinicians during the iterations.) Many AI products are driven by engineers, rather than UX, which means they will have suboptimal use when delivered to real users.

AI Does Mathematics Differently Than Humans

Epoch AI analyzed how the new Grok 4 Heavy model does mathematics. The project confirmed that Grok 4 excels in math: Grok 4 scored 88% on a set of medium-hard math competition problems for advanced high school students, representing a slight improvement over the previous winners, Gemini 2.5 Pro and OpenAI o4 mini (high), both of which scored 85%. (I would rather have it be great at creative writing and judgment of business priorities, but it’s certainly good to have AI improve at mathematics, which most humans are incapable of doing well, meaning that AI assistance is sorely needed.)

These standard benchmarks are close to being saturated, and I find the qualitative analysis of how Grok solved the problems to be more interesting than the raw scores.

Grok performed better when solutions could rely on a broad knowledge of existing mathematics, such as familiarity with a large number of theorems. AI has read everything, and contrary to human mathematicians, AI can remember all the theorems and apply them when needed.

However, when it comes to solving new problems, AI currently takes a different approach than the best humans, who “find clever shortcuts which allow them to solve problems without having to spend as much time chugging through equations,” as the report says. AI (whether Grok or competing models) relies more on a “grind it out” approach, where it quickly tries out a large number of steps. Of course, AI doesn’t get tired, it does know all the theorems and formulas so that it can try out a wide range, and it can think for as long as we have compute budget to support solving the problem. Many problems in the real world are worth thousands or millions of dollars to solve correctly (imagine a structural engineering assessment of whether a $100M bridge will collapse in a hurricane), and so can easily justify hours of GPU time.

(Let’s say we’re willing to pay $1,000 to solve a math problem. That money wouldn’t go far in hiring truly skilled consultants, but it would buy us 264 hours of B200 (Blackwell) compute from an AI cloud provider. I doubt any reasoning model reasons for that long in current production.)

AI excels in mathematics, albeit by approaching math problems differently than the most talented humans do. (GPT Image-1)

Epoch also asked two mathematics professors to evaluate Grok 4’s abilities in academic literature searches. They were impressed and felt it would be helpful. Professor Bartosz Naskręcki said, “In literature searches, Grok 4 shines by quickly identifying relevant papers and connections across mathematical fields, helping me discover ideas and references I might have otherwise missed.”

Does it matter that AI solves math problems differently than human mathematicians do? Probably not: it’s a different intelligent species, and we should not expect it to work the same as us. As long as it gets the results, the methods may not matter much. At the same time, the fact that high-IQ humans can solve math problems in more creative ways demonstrates the potential for further improving AI. It’s very plausible that, given that this capability for more efficient intelligence has evolved once, it could be constructed again.

Ideogram Character Consistency

Ideogram released a feature named “Character” which allows you to upload or create an image of a character and then have that character used with decent fidelity when it generates new images. As an example, I gave it a simple selfie snapshot of myself and asked it to generate 3 images: Myself as a pirate, transformed into Studio Ghibli style, and as a paleolithic mammoth hunter.

Images created with Ideogram’s “Character” mode.

Character consistency has long been a problem for AI image generators. Even with fairly detailed descriptions of each character (like I did for my series of manga with UX career pivot recommendations), characters don’t look truly the same from one image to the next. This impedes professional comics projects, and even serves as a block for more fun-driven projects like my own (see, for example, my video about Sovereign AI, where the two characters — Jensen Huang of Nvidia and Arthur Mensch of Mistral AI — look very different between the two parts of the video).

In my experimentation, the two photoreal images (pirate and hunter) replicate my likeness rather faithfully, whereas the animate image doesn’t look how a great artist would have drawn me. Besides character consistency, we also want images to represent the author’s vision, and Ideogram’s stone age hunter fails in this regard. The prompt said that I was hunting a wooly mammoth, but the picture looks more like I’m posing in front of it. (And the mammoth is a rather scrawny specimen.) Compare with the image ChatGPT’s native image mode made for the same prompt: much more engaging hunting scene (which I used as the base image for an even better B-roll clip made with Kling 2.1 Master in the explainer video I made from that article).

One more interesting point about Ideogram’s new feature: my pirate image was made with what they call “templates,” which are basically pre-made ideas for character transformation that are presented in a gallery. This is a good use of prompt augmentation to get users started with the new feature by giving them fun results from a single click and by showcasing the range of creations the new feature supports.

Spot User-Testing Methodology Flaw

My recent video “Usability Testing as Theater” (YouTube, 2 min.) contains a few examples of poor study facilitation methodology. What do you expect? It’s theater! (And more to the point, what you get from the current state of the art in video prompting.)

One mistake is so blatant that I felt I had to call it out in the video captions, to avoid unsuspecting souls from copying my example: the facilitator instructs the user to “try clicking on the big red shiny button.” This is obviously blatantly leading and will invalidate the test.

Another instance of flawed methodology is slightly less blatant; I have observed it many times with less-skilled test facilitators. Can you spot it? I’ll give the answer next week.

Instructing a test participant on where to click will clearly invalidate a usability study. My “Usability Testing as Theater” video includes another slightly less obvious methodology mistake. Try to find it. (GPT Image-1)