UX Roundup: AI Empathy | Transparency About Compute Limits | UI Modes | Seedream 4.5 | AI In/Out of Silicon Valley | Kling Speaks

Summary: AI exhibits better empathy than human clinicians | Compute limits should be made transparent in the user experience | User interface modes usually lower usability | Seedream 4.5 image model now with improved text rendering | How feasible is it to build an AI company outside Silicon Valley | The Kling video model now generates speaking characters

UX Roundup for December 8, 2025. (Nano Banana Pro)

Creation by Discovery

New music video about how creators explore the latent space of possible design solutions and use their judgment to navigate in new directions, guided by what they discover. (YouTube, 5 min.)

The video is an example of the very point it discusses: my concept was a K-pop stage performance, and I started with a nice image of the group on stage. Then Nano Banana Pro gave me about 30 alternate poses and angles on the singers, which made for more variety when designing the dance breaks.

You may have noticed that in my previous music videos, the dancer strikes the same pose in every dance break, because they were animated using the start-frame feature of modern video models. My new video is no different in terms of animation: the video models still don’t understand music or choreography, so they animate dance moves guided by that start frame. The difference is that still-image creation now allows exploration of the latent space of possible poses.

(Nano Banana Pro, based on a lead singer, made with Grok, and the remaining K-pop idols outpainted with Seedream: three different image models combined to create one image.)

AI Exhibits Better Empathy Than Human Clinicians

I’ve covered several studies showing that patients rate AI higher than human clinicians for empathy. How well people feel that the AI treats them is obviously a separate question from whether the AI is providing the correct clinical diagnosis and treatment advice, but other studies show that AI is getting better than human doctors in this aspect of medicine as well.

Patients rate AI as exhibiting more empathy than human doctors. (Nano Banana Pro)

A new meta-analysis, published in the December 2025 issue of the British Medical Bulletin, now provides a more systematic review of AI empathy, going beyond the individual studies. Alastair Howcroft and colleagues analyzed 15 studies, published between 2023 and 2024, where the “AI” condition was ChatGPT 3.5 or 4.

Clearly, this research is already outdated, since we have moved on to models like GPT 5.1 Pro and Gemini 3 Pro, with even better models expected in 2026. However, academic publishing moves slowly, and meta-analyses move even more slowly, since they cannot be conducted until after the source papers with the underlying research studies have been published. All we can say is that the results of using AI in 2026 are likely much better than AI in 2023 and 2024, but we don’t have the exact numbers.

The speed of academic publishing is completely misaligned with the pace of AI progress, so for research to be useful and to impact future product development and field deployment of AI, something must be done to expedite data analysis and result dissemination. At least AI can help scientists analyze their data more rapidly. (Nano Banana Pro)

Returning to the data we have on AI in 2023 and 2024: Of the 15 studies the authors analyzed, 13 reported statistically significantly higher empathy ratings for AI than for human clinicians, with only two dermatology studies favoring humans. (Whether human dermatology doctors are particularly nice, compared with human doctors specializing in other fields, is not known. There could be many other explanations for these two outlying results.)

Sometimes, bad news or difficult conversations are easier when you know you are dealing with a computer. It would be valuable to have deeper research into why people prefer dealing with an AI, but pragmatically, the reasons don’t matter for now (even though they matter immensely for designing better AI in the future). All that matters is that patients do prefer AI. (Nano Banana Pro)

13 of the 15 underlying papers provided data suitable for meta-analysis. Overall, AI demonstrated significantly (p<0.00001) higher empathy than human practitioners, with a standardized mean difference [SMD] of 0.87. Given the variability in the scores, 0.87 standard deviations roughly correspond to 2 points on a 10-point rating scale. For example, if human doctors were rated a 6, then AI would be rated an 8.

As I pointed out, having only data about GPT 3.5 and 4 is an inevitable weakness of this meta-analysis. But at least the data does encompass two AI generations. The two AI models scored as follows, relative to the human clinicians:

• GPT 3.5: 0.51 SMD better than humans (across 4 studies)

• GPT 4: 1.03 SMD better than humans (across 9 studies)

Because of the small number of studies, this difference between the two AI models was not statistically significant (p = 0.16), but the data indeed suggest that AI at least didn’t get worse and, more likely, improved from 3.5 to 4.

The measures of empathy are not particularly strong. The most common measurement (8 of 15 studies) was single-item 1–5 Likert scales, where responses were rated from ‘not empathetic’ to ‘very empathetic’ or similar descriptors. Of course, empathy is a subjective quality that lies in the eye of the beholder. When interacting with somebody else (whether human or an AI), do you feel that this “other” being is appropriately empathetic? So we probably mostly have to rely on subjective instruments, but if AI empathy is an important quality to develop further, a more detailed (and validated) questionnaire would be better.

While the simplistic 1–5 empathy rating scale was acceptable for early studies, we need more insightful survey instruments for future research into AI empathy. (Nano Banana Pro)

New meta-analysis: AI exhibits much higher empathy than human clinicians, as rated by patients. (Nano Banana Pro)

Early research on clinical AI has been limited to text communication between the patient and the provider (whether human or AI). Newer AI has a voice mode for more natural communication, and real-time avatars will soon provide visual images for patients to relate to. Robots that can provide hands-on care are still in the future, though I would not be surprised if patients start preferring robots to humans for medical treatment. We already know from revealed preferences that passengers far prefer being driven by Waymo’s autonomous cars over riding in human-operated Uber cars. (Nano Banana Pro)

The UX of Compute Limits

OpenAI and Google both have too little AI inference compute to serve the user demand for AI services. As a result, they have both recently introduced severe limits on the number of requests free users are allowed, especially for compute-intensive workloads such as video rendering and Nano Banana Pro images.

For example, Sora 2 dropped the allocation for free accounts from 30 to 6 videos, and Google recently cut image generation with its new high-end Nano Banana Pro image model to only two images per day for free users.

I cry dry tears for unpaid users, because there’s a limit to how much service you can expect to get for free. (Remember that “free” means that somebody else is paying. In this case, people with paid subscriptions.) Much worse that users on high-end paid subscriptions often see their generations fail due to overloaded AI data centers.

The problem of insufficient AI inference compute will likely persist for the next decade, despite the extensive build-out of giga-watt-capacity data centers. Superintelligent AI will likely require at least 1,000x the compute per query, and possibly even 10,000x what we’re spending now.

Similarly, for media generation, by 2035, I expect that video generation will produce millions of feature-film-length videos every day that are better than anything ever to come out of Hollywood, and that music generation will create millions of songs every day that beat the entire Beatles catalog by far. (When I say “better,” I mean in the judgment of the person generating the video or song exactly to his or her taste.) Such generative media will require millions of times more AI tokens than what I spend now on a simple 3-minute music video with a decent song.

Maybe by 2040, we’ll have enough compute to meet humanity’s full demand for AI, but for many years to come, severe rationing will continue. This again means that the design of AI products must make these limits and their consequences transparent in the user experience, as we see in this comic strip:

(Comic strip made with Nano Banana Pro)

For designers and builders, this shift carries significant implications. AI-first tools can no longer be treated as always-on, unlimited resources. Experiences built on that assumption risk breaking when quotas are exhausted. The solution lies in what might be called "compute transparency," making resource costs visible within the interface itself.

Practical patterns include quota meters showing remaining allowances, graceful degradation to faster but lower-quality models when limits hit, and scheduling options for off-peak processing. Progress indicators should communicate realistic wait times, while quality-cost toggles let users choose between draft and premium outputs. Pre-flight estimates and post-run receipts build trust by eliminating surprise costs. Ultimately, giving AI generation clear visibility and thoughtful fallbacks will define resilient user experiences in this new compute-constrained landscape.

Here are some sample UX design patterns to consider:

The Honest Banner

When system constraints require pausing functionality, display a transparent banner that explains the situation without alarming users. The copy should be direct and reassuring; something like "Uploads are paused right now while we scale storage. Your files are safe." Provide actionable controls such as "Notify me," "Try again," and "Download existing" so users retain agency. Include a detail link labeled "What caused this?" that leads to a short, skimmable explanation rather than a dense technical document.

Soft-Lock Action Buttons

Design primary action buttons to display predictive wait labels that set expectations upfront. For example, "Generate (≈3 min)." Offer secondary options like "Schedule for later" and tertiary alternatives such as "See alternatives" to give users flexibility when immediate processing isn't ideal.

The “Tiny Win” Fallback

When full functionality isn't available, offer a degraded-but-useful path that still delivers value—such as a low-resolution preview, partial search results, or local export options. Label this mode honestly with copy like "Basic mode (faster, fewer features)" so users understand the trade-off they're accepting.

The Queue Card

Replace ambiguous spinners with a queue card that shows the user’s place in line alongside a dynamically updating estimated wait time. Effective microcopy might read: "You're #42. ~6–8 min. We'll email if it's your turn while you're away." Always include an escape hatch like "Save and exit" so users can leave without losing their place or progress.

The Staged Uploader

Build upload flows that accept files and pre-validate them locally before attempting any server-side operations. When capacity is constrained, defer the server step with a clear promise: "We'll start as soon as capacity frees up. You'll get a timestamped receipt." This approach keeps users informed while preventing frustration from failed uploads.

The Status Capsule

Maintain a compact, ever-present status chip visible throughout the interface that communicates current system conditions. For instance, "System load: High • Est. slowdowns ~5–7 min." Tapping this chip should open a short status drawer within the application rather than redirecting users to an external status page.

The Friendly Deferral

When processing will take significant time, offer users the option to receive results asynchronously with specific details about what they'll get. Frame this as "Send me the result" with the exact artifact named: "Email me the PDF (estimated in 18–25 min)." This respects users' time while maintaining their connection to the outcome.

What to Avoid

Steer clear of humor that trivializes the situation, such as jokes about hamsters powering servers or systems being “on fire.” (I can run such jokes in my newsletter, but jokes don’t belong in a UI where they add insult to injury when you can’t deliver what users want.) Avoid vague error messages like “something went wrong” that leave users without actionable information. Never blame users for system limitations, and resist the temptation to display fake-precise completion times like “done in 2m 03s” when such accuracy isn’t actually possible.

Compute limits must be reflected in user interface design, rather than blowing up the user experience. Even if you don’t face compute limits in your service right now, plan to introduce some of these design patterns in the future. If you design them now and run through user testing, you will be able to react fast when the problem inevitably hits home. (Nano Banana Pro)

Just for fun: the previous infographic converted into a Renaissance painting. (Nano Banana Pro)

User Interface Modes Lower Usability

Google has reportedly started work on simplifying the user experience for using AI in its search engine, aiming to integrate better the three separate UI concepts of “AI Mode,” “AI Overviews,” and plain old Dumb Search. I applaud this initiative and look forward to better Google usability.

However, the problem was having three different concepts in the first place and introducing modes in the user interface.

Modes are usually detrimental to usability, as famously championed by Larry Tesler, who even had a custom license plate made for his car reading “NO MODES.” Tesler is one of the heroes of usability, and in particular, a hero of the graphical user interface, where he made major contributions first at Xerox PARC (where most of the modern GUI was invented) and then as UI lead for the Apple Lisa, where the GUI was made practical. (Lisa was the precursor to the Macintosh.)

(Nano Banana Pro)

After his work on the Lisa GUI, Larry Tesler eventually become Apple’s Chief Scientist and later held roles as VP for shopping experience at Amazon.com and VP of UX at Yahoo! Based on his many contributions, Tesler received the SIGCHI Lifetime Achievement Award for Human–Computer Interaction Practice in 2011. When I received the same award two years later, Tesler attended the award ceremony, and we had a good discussion about many of these early events. But even the publicly known stories will have to wait for another comic strip.

Seedream 4.5 Launched

The Chinese image model Seedream is now at version 4.5, with improved text rendering. I used it to draw its version of the last page of my comic strip about Larry Tesler (see above). I used the exact same prompt from the storybook as I had used with Nano Banana Pro, and also the same style reference for the cartooning style and the same character reference sheet for how to draw the “Larry” character.

Seedream 4.5 pretty much nailed the storybook description of this page, except that frame 2 (upper right) was supposed to have a caption box, not a speech bubble. It’s unclear who’s saying “Computing became accessible to everyone” in this frame. I also think it rendered Tesler’s car as too modern, given that he first got his famous “NO MODES” license plate in the 1970s and used it first on a Dodge Valiant and later transferred it to a Subaru station wagon. (Nano Banana Pro also got the car brand wrong, probably thinking that a Silicon Valley VP would own a fancier car than what Tesler actually drove. But at least it drew a period-appropriate design for the car.)

Seedream didn’t match the style of pages 1–5 of the comic strip as well as Nano Banana Pro did, but in fairness, the Banana had drawn all of these pages itself before being asked to draw page 6 in the same style. If I had made the entire strip with Seedream, style consistency would probably have been better.

In total, I would score Seedream 4.5 as almost as good as Nano Banana Pro for this comic strip assignment. For sure, it got all the text right, which was very rare for image models until recently, and most image models still can’t make a coherent comic strip.

Silicon Valley or Bust?

Alex Danco has an interesting, if rather long [as if I am in any position to complain about wordy articles 😬], discussion of the benefits of building AI-Native companies in various locations. He starts with the observation that nearly 100% of the interesting AI companies outside China are within a 20-mile (30 km) radius of Silicon Valley. But why?

During the previous decade (the 2010s), many important technology companies were started outside the Silicon Valley region, which had dominated the initial dot-com bubble boom in Internet companies. Internet technology itself had become a commodity, so that one could build useful applications of it anywhere in the world. In fact, it was more important to draw upon local talent and local knowledge of important business processes than being located in the area where Internet technology was originally invented and where its initial applications were built. Furthermore, the basic structuring principles for managing such companies and development projects were widely known. (Within my own area of expertise, it’s certainly true that the legacy UX design process that I used to recommend before the AI era was exceptionally well known: there were many highly qualified UX experts located in any significant part of the world. Including Europe. When I visited home in 1995, almost nobody knew how to build Internet businesses or how to design websites. These days, thousands of Danes are fully up to speed.)

If it weren’t for the EU’s oppressive AI Act, Europe would be a great place for many small AI-Native companies, building in the application space, as opposed to foundation models. (Nano Banana Pro)

Now, everything is up in the air again. We don’t know what AI will be next year, let alone after we reach superintelligence around the year 2030. We don’t know the proper methodology for designing and building with AI or for AI. This means that the tight loop of everybody being in the same location, going to the same parties, and hopping back and forth between the same companies suddenly has value again. (In Europe, the EU’s “AI Act” furthermore makes it extremely burdensome to run advanced AI projects, which of course means that fewer even get attempted, let alone become successful.)

For now, paperwork is holding Europe back in AI. (Nano Banana Pro)

Ten years ago, a widely held saying held that “only an idiot starts a company in California,” because of the many expenses and impediments to execution imposed by the state. These expenses are even worse today (I pity anybody trying to rent an apartment in San Francisco), but have temporarily become irrelevant: when an AI company expects to invest several hundred billion dollars in building training and inference supercomputer centers (outside California, of course), it becomes a rounding error to waste a hundred million dollars on California overhead. Getting to superintelligence a few months faster makes that expense worth paying.

Right now, a Silicon Valley location helps a company get to superintelligence faster because of the high bandwidth knowledge sharing. (Nano Banana Pro)

While it’s true that the biggest and most important AI companies are in Silicon Valley, Danco points out the paradox that it’s simultaneously true that many smaller AI operations are starting locally, in almost any place you can think of. As he says, “local A-player builders now have a really viable alternative to working for a local tech company, which is working for themselves.” With AI, it’s easy to run a small company with a superefficient team. You hire a handful of the kind of people who used to be known as “10x developers,” and they become “100x” builders because of AI acceleration and the lack of that bureaucracy that would slow them down in a bigger company.

Some of the most aggressive builders elect to stay in their own city and start a focused AI company there. (Nano Banana Pro)

The A-players have always been able to develop code 10x faster than average programmers. But now they can become 100x by joining a small AI-Native super-accelerated team. (Nano Banana Pro)

As a result, companies outside Silicon Valley are likely doomed if they are either more than 10 years old or have more than 20 employees. (A small company could pivot even if it’s old, if it goes hardcore “founder mode.” And a big, new company may still not be so set in its ways that it couldn’t go AI-First and deploy transformative AI to reinvent all its workflows over the next two years.) By being “doomed,” I don’t necessarily mean they will go out of business before 2030, but they may become irrelevant to the future of the world, which again means they will lose their A players to companies that matter.

Smaller AI companies may be better off avoiding the extreme Silicon Valley overhead expenses and staying in the founder’s local area, scooping up the local A-players. (Nano Banana Pro)

Kling Video Model Now with Native Audio

Chinese video model Kling is now in version 2.6, with the main new feature being an ability to natively generate audio to match the video. Version 2.5 already had Foley effects, but v. 2.6 adds human speech in English and Chinese. It does fairly good lip synch, but I don’t think it is quite at the level of the leading American video models Veo 3.1 (Google) and Sora 2 (OpenAI).

(Nano Banana Pro)

I made a short demo reel with the new model (Instagram, 1 min.), featuring three different kinds of video: photorealistic image-to-video and text-to-video, as well as a 3D-animated character of a cute tiger. The photorealistic clips depict the Viking attack on Paris in 886, with a Viking anachronistically recommending that viewers subscribe to my newsletter. For the sake of comparison, my demo reel includes a clip from Sora 2 made with the same text-to-video prompt. Unfortunately, I could not use Sora 2 for image-to-video, because it refuses to make videos based on an uploaded photo of “real people” for so-called “safety” reasons. (Leave aside the fact that my start frame image was AI-generated and depicts a made-up horde of barbarian Vikings in the year 886.)

Kling also released a much-improved avatar model that can animate videos up to 5 minutes, setting a new record. HeyGen’s Avatar IV only does 3 minutes, which often means that I have to stitch together two generations to make a video. (For example, my explainer video AI Helps Old Knowledge Workers Stay Creative ran 6 min. 36 secs., so it would even have been too long for Kling.) I made a short comparison video showing Kling and HeyGen animating the same avatar singing a minute of the same song.

The best way to compare AI avatar models is to have them animate the same avatar with the same soundtrack. I chose a segment from my latest song.  (Nano Banana Pro)

While I am happy to see competition in the avatar space and certainly congratulate Kling on a major step up from its previous avatar model, my conclusion is that HeyGen remains King of Avatars. Higher image quality and more dynamic gestures that are less repetitive than Kling's gestures. What do you think?

I continue to like Kling for B-roll clips of dance breaks in my music videos, but I won’t be using it for talking videos until the next release hopefully improves the quality of the acting.