UX Roundup: AI Music | Humans Impolite to AI | Subscriptions vs. Social Media | AI vs. Search
- Jakob Nielsen
- 1 day ago
- 9 min read
Summary: Experimenting with same song in different genres | Humans impolite to AI, which needs to be reflected in training data | Subscriptions beat social media for insightful content | AI answers beat search
UX Roundup for November 10, 2025. (Seedream 4)
Two New Music Genres
I remade my opera about direct manipulation (YouTube, 4 min.) in two modern music genres:
One reason I enjoy making music videos is that each one gives me the chance to experiment with new designs for the singer and the setting, to match the song’s style. This is what Richard Wagner called a “gesamtkunstwerk” (“total work of art”), where multiple media forms are synthesized into a unified, immersive experience. Today, we use the more modest term “multimedia production” for this idea, but I like Wagner's high-concept 19th-century Germanic thinking.
Usually, I prefer designing new avatars for each video, but this time, since I was making two music videos in an afternoon, I decided to try an alternate approach for the jazz song and experiment with character consistency instead: I took my avatar from the rather dry explainer video about transformative AI economics and used Seedream 4 to change the avatar’s outfit and setting.
The following three images show the workflow. I often start with Grok Imagine for the avatar creation because its fast response time allows for creative exploration of the latent design space of design possibilities.
Moody avatar design by Grok Imagine 0.8, used as the basis for both my Transformative AI explainer and my jazz song about direct manipulation.
Grok’s Imagine 0.8 version, which I used back in September, only generated images in portrait orientation, whereas I want 1920x1080 landscape images for my videos. However, I used Seedream 4 to outpaint the original image from Grok and fill in some additional parts of that rainy European street for a more interesting environment.
Seedream generates 4K images, which means that I get enough pixels to crop out a smaller part of the image to use when animating close-ups.
Outpaint by Seedream 4.
Originally, I thought of making the avatar sing the jazz song in that same rainy street. It would have been quite an appropriate mood for the song. But since I set the rock song in a concert hall, I decided to set the jazz song in a nightclub.
The avatar’s leather outfit was still basically fine for my revised concept, but instead of a long coat (nice in the rain!) I changed it into a short leather jacket with basically the same look.
Now for the real challenge: using Seedream to render the same character in a new environment, with that slightly changed outfit. Worked like a charm:
Using Seedream 4’s character consistency with the original Grok avatar design as the reference image to reimagine the avatar performing in a nightclub in a more indoorsy outfit.
One last point about avatar outfits: Here, I only got upper-body shots when using Seedream to reimagine the original Grok avatar design. This is likely to be the default when uploading a reference image framed like that. The problem is that we don’t know what the avatar is wearing below the image frame, so when I feed it to Kling to use as the start frame for dance breaks or intro/outro scenes, Kling has to invent the bottom part of the avatar’s look.
It usually does a good job, but slightly differently in each clip, which is why many of my avatars end up wearing different shoes throughout a video. If I were a true film director and as obsessed with detail as Woody Allen reported was, I would need to make a single full-body reference image, or even an entire character sheet.
I am not obsessed, since I create for the joy of it, not to make “art.” Eventually, AI will improve its handling of character consistency to free human creators from such worries (or, like I currently do, having to live with continuity errors).
A final point, which is a repetition of what I’ve said before: to make even a short AI video, it’s best to combine multiple tools, each with their strength. For this project, I used: Grok Imagine 0.8 and 0.9, Seedream 4, HeyGen Avatar IV, Kling 2.5 Turbo, MiniMax Hailou 2 (for a single outro clip), Veo 3.1, Suno 5, Gemini Pro 2.5, and Claude Sonnet 4.5. A total of nine different AI models (or ten, if you count the two different dot-releases of Grok’s image model that I used for different avatars made a few weeks apart).
To see how far we have come in AI music and video, compare the jazz music video I made in January 2025 with the jazz video I made in October 2025. The music was clearly simpler in January, but still holds up well as a small-group jazz performance. In contrast, the animation has taken a great leap forward. (While still not perfect, of course.)
Humans Are More Polite to Humans Than to Chatbots
Researchers from Amazon studied 13,000 messages sent to human customer‑service agents and chatbots. They found that people are 14.5 % more polite and 5.3 % more grammatically fluent when speaking with human agents compared with AI.
Training chatbots on a mix of human‑ and AI‑directed message styles made them 2.9% better at understanding user intent than AI trained solely on original human conversations.
Both results from this study make sense: If you know you’re talking to an AI, you might as well skip the chitchat and politeness and get straight to business. And there are likely other changes in the way users communicate with AI. This again means that AI should be trained on how users interact with AI and not on old training data derived from human communication.
People communicate differently with fellow humans than with AI. This should be reflected in the AI’s training data. (Seedream 4)
Subscriptions Beat Social Media
I have noticed that when I post on social media, my best work rarely gets the most impressions. Viral posts do much better, but unfortunately, they often tend to be less insightful.
To ensure that you receive my best work, do not rely on social media, but subscribe to my email newsletter on Substack.
Newsletters will be delivered promptly to your mailbox by fast carrier pigeons. (Or to your inbox by Internet electronics, as the case may be.)
My newsletter has just been delivered to one more satisfied subscriber, and the carrier pigeon can rest for the day. (Seedream 4)
AI Answers Beat Search (Again)
Two years ago, I explained why AI answers have superior usability to the then-traditional use of search engines to find information. I also discussed how websites should pivot their user-acquisition strategy away from SEO.
Websites that followed my advice from 2023 are in a better situation today than those who believed that the Google firehose would keep showering them with visitors forever.
The firehose of traffic that fed so many websites for so long is drying up. Time for a content strategy pivot rather than bemoaning the lost past. (Seedream 4)
My current advice has become more nuanced, but is basically the same as it was in 2023. Here’s this year’s articles on how to survive without search traffic:
If you work on content strategy, I recommend reading all 4 of the linked articles. You can kick yourself for not having started following the advice in 2023, but it’s never too late to pivot.
Recent data confirms what I already said.
First of all, Google has clearly decided that the leftover search money tree is killed off by themselves rather than leaving it to Perplexity to do the job. For years, Google was hesitant to divert its legacy search users to AI, knowing full well that it would be trading Search Advertising dollars for AI Advertising pennies. (To repurpose Jeff Zucker’s famous quote from 2008.) Now, Sundar Pichai clearly realized that pennies are better than bankruptcy and decided to leverage Google’s still-overwhelming traffic in its quest for AI supremacy.
Google’s SERP-ad money tree is dying, and they clearly decided to pivot to fertilizing the new AI tree, even if it will be less fruitful. Better to have a small money tree than no tree at all. (Seedream 4)
For website owners, the message is clear: even Google doesn’t believe in search anymore.
Search is quietly being buried. Google hopes that its AI features will rise like a Phoenix after the funeral. (Seedance 4)
Unfortunately, in typical Google fashion, the user experience architecture of Google’s AI answers is confusing, with several different types of AI answers more or less integrated with its legacy search engine.
I made these cartoons in July when I last commented on the confusing user experience architecture of Google’s AI services. At least with search, Google may be limiting itself to two versions: AI Overviews and AI mode. (GPT Image-1)
Except possibly for OpenAI, no other company could survive with such a confusing AI-UX, but Google is still strong enough to pull it off. Plus, I would hope/expect that they improve the AI-UX architecture in 2026.
Will Sundar Pichai be able to break free of the Promethean chains of past search glories and lead the 25% of Googlers who will survive to reach the glories of AI supremacy? I think he has the best chance of anybody in today’s environment. (Seedream 4)
Consulting firm McKinsey surveyed 1,927 consumers in the United States and found that half of Americans intentionally seek out AI-powered search engines, with a majority of users saying it’s the top digital source they use to make buying decisions. (Since the data was collected in August, the percentages are likely bigger now.)
The most striking data from the McKinsey survey is the answers Americans give for their most preferred source of information:
AI: 44%
Search: 31%
Review sites, social platforms, and editorial content: 13%
Brand or retailer sites + Live commerce: 11%
My prediction is that AI is already bigger and will cross the 50% mark in 2026. Other countries? Except for China, the rest of the world is behind the United States in AI adoption; however, the trend is expected to follow with a one- or two-year delay, even if the EU stubbornly refuses to repeal its misguided AI Act.
My apologies to Norman Rockwell, who would not have painted this scene since he lived in more innocent times and cultivated nostalgia in the first place. However, as AI answers replace search results, there is no reason to be nostalgic for the recent past of heavily-spammed search listings with minimal support for the masses of users who have abysmal information-seeking skills. The AI future will be more worthy of nostalgic feelings. (Seedream 4)
The Italian federation of newspaper publishers, FIEG, has filed a complaint against Google, saying that it has become a “traffic killer,” reducing clicks to newspaper sites by 80% after introducing AI Overviews. (Why legacy publishers feel entitled to Google giving them the same amount of traffic forever as it did in the old days is beyond me.)
Italian legacy publishers want Google to return to its old ways and send them as much traffic as it used to, instead of subjecting traffic referrals to the mercies of the AI Guillotine. (Seedream 4)
In somewhat old data from March 2025, the Pew Research Center found that among 900 American users, the percentage of people who clicked a link from Google to a content site dropped in half (from 15% to 8%) when the user saw an AI summary as opposed to a traditional SERP without AI.
Taking my own website, www.uxtigers.com, as an example of a content website with useful information (as opposed to a “black-hat” spam site that deserves to lose traffic), in October 2025, my analytics show 207 times more traffic from AI services scraping my content than the number of visitors referred to visit my site from those AI services. (The top 3 AI services that do send me a trickle of traffic are ChatGPT, Perplexity, and Gemini.)
One out of every 207 AI scrapes led to a human visitor to my website.
In other words, the clickthrough rate is 0.5% as the ratio between users clicking through to my site from an AI answer versus how often AI used my content to improve the quality of its answers. Out of the many times AI read my content, it likely only explicitly cited me some fraction of the time in its synthesized answers, so I don’t know the clickthrough rate between AI showing users a citation from my site to people clicking that citation.
The specifics vary between these sources (McKinsey, Pew, Italian newspapers, and UX Tigers analytics), but the overall trend is clear: a bloodbath of search-derived traffic to websites as users increasingly turn to AI answers, which usually provide all the information they need right there, obviating the need to click through for more information.
This scene will be playing in companies worldwide over the next two years: Where did our website traffic go? Why has customer acquisition imploded? Prepare now. (Seedream 4)