Summary: Being asked for a status update is a bad sign | Strong results from AI tutors in Nigerian schools | UX optimism | Ideogram text editing feature | AI-UX conference in Berlin | EU multilingual LLM
UX Roundup for February 10, 2025. Happy Valentine’s Day! (Midjourney)
Listen to the Be My UX Valentine song (YouTube, 2 min.).
Anybody who speaks French, German, Korean, or Polish, please check the translated subtitles and let me know in the YouTube comments whether you like the rhymes. (For the other translated subtitles, I didn’t bother rhyming but just tried to retain the meaning of the verses.)
Don’t Wait to Be Asked for Status Updates
If stakeholders ask you for a status update, you may take this as a positive sign: they’re interested in your work and eagerly awaiting your deliverables!
However, Jessica Hancock, a UX designer at Adobe, points out that it’s actually a problem if they don’t already know how you’re doing. It’s your responsibility to communicate with stakeholders proactively.
She adds the useful insights that:
Your work is only as good as your ability to present it. Especially if you ever want a promotion.
You must know the ROI of your work and also communicate that. (This means understanding how your efforts convert into dollars for the business.)
Your work is only as good as your presentation. As AI creates more of the on-screen design in the future, your ability to present designs to stakeholders will become even more important. And you must be highly cognizant of the business value of every recommendation. (Leonardo)
AI Tutoring = 2 Years of Learning in 6 Weeks
The World Bank has published the results of using AI for individual tutoring of students in Nigerian schools. 800 students used Microsoft’s version of ChatGPT twice a week for 6 weeks for individual tutoring in English language skills. This is a very modest exposure to AI tutoring, and yet the results were dramatic:
Compared with a control group, students who used AI for individual tutoring improved their test scores by 0.3 standard deviations. This is a big effect size.
The skill improvement that the AI-tutored students gained in 6 weeks is equivalent to the improvement that usually requires two full years of traditional classroom schooling to achieve. This degree of acceleration is almost unbelievable and makes me regret that the project only lasted 6 weeks. Of course, it’s possible that the biggest gains come from the initial use of AI and that subsequent gains might taper off, but it’s equally possible that individual AI tutoring would retain its edge over traditional education indefinitely, since the AI tutor can adapt to the students’ growing skills much more accurately than a human teacher can do when faced with having to teach an entire classroom of kids with highly varying skills.
Maybe most important, the students who had used AI for such a short time didn’t just score better on a test immediately after the project, but they retained higher scores than the control group on the regular end-of-year exams.
African children improved their skills by the equivalent of 2 years of traditional classroom learning after only 6 weeks of individual tutoring from an AI system. (Midjourney)
Comparing the test scores for the control group and the students who used AI for individual tutoring for 6 weeks. (Reproduced from the World Bank report.)
Most studies of using AI in the workplace find that it helps poor performers more than it helps strong performers. However, this study seemed to have the opposite outcome, as shown in the above test scores chart. AI helped everybody, and that’s obviously the main conclusion. The weakest students did learn more, and since there are many students below the average, this does sum to a large total amount of skills gained.
However, the fattest part of the differential area is in the range between +0.5 SD to +1.5 SD above zero: these are scores that represent clever but not super-smart students. The gain in this region seems about twice as large as the gain in the mirror-image part of the chart, between -0.5 and -1.5 SD (which is the domain of the somewhat dull but not extremely stupid kids).
Finally, let’s look at the area above +2.5 SD: these are the super-smart children. It looks like the number of children achieving this very high skill level increased by about a factor of 3x (a gain of 200%), which is astounding.
In contrast, the reduction in students with very low skills (below -1.5 SD) isn’t nearly as large: the control group has about 75% more students who learned this little, compared with the AI-tutored group.
Why did this study show bigger gains at the top than the bottom of the skill distribution? The famous “more research is needed” applies here, but I do have one plausible explanation: with traditional classroom education, the teacher must cover the material for the full range of students, from the low end to the high end. This inevitably means most classroom time is spent on the most basic material since that’s all the lowest-skill students can learn. Any time the teacher attempts to cover advanced material for the smartest students, the bottom half of the class will be left behind. Thus, it’s rare for classroom instruction to spend much time on the most advanced material, even though the top students could learn this content if it were taught.
This means that high-skill students have a much bigger potential for being taught more appropriate material with individual AI tutoring. In contrast, low-skill students are already spending most of the classroom time being taught material at their level, so they don’t gain so much more by having the curriculum optimized for their specific individual needs.
A final point is that this study does not prove that human teachers will become superfluous. The students in the AI tutoring condition still had human teachers available during the AI sessions. However, the role of the human teacher changes from presenting the curriculum to supporting their students in other ways — anything from serving as a role model to keeping the students motivated and the class on track, even as each student proceeds along that track at his or her optimal pace.
UX Optimism
A slide show I made about UX Optimism. (Emojis made with Leonardo)
For more detail, see my recent piece AI Impact on UX Jobs.
Listen to my UX Optimism song (YouTube, 2 min.), animating the emojis and adding B-roll for the dance break from Kling 1.6 — Suno likes to include a dance break in almost all my songs.
Ideogram Text Editing
Ideogram has introduced a new text editing feature. First, you generate an image from prompting the usual way — possibly while iterating on the prompt and the generations to get something you like. Then you activate a text editing tool to superimpose the text of your choice, including typeface, size, weight, and color.
The main benefit from this feature seems to be to avoid the step of moving the image into a traditional editing tool such as Photoshop. Currently, there are no AI features, for example to suggest suitable typefaces or skewing the text to fit with the image. Ideogram’s own announcement of the feature suggests trying it on a flat frame like the example below. (Indeed, I used Ideogram’s suggested prompt: “A watercolor illustration of a blank paper with an elegant gold border and a red rose at the top. The paper is intricately decorated with a whimsical floral pattern of red, pink, and purple flowers and green leaves. The background is white.”
Top: Using the new Ideogram text editing tool, which allows me to specify the typeface. Bottom: Creating a similar image with pure prompting, including the center text in the prompt. The font and word wrap are whatever Ideogram conjured up. If you don’t like that design, you can run another iteration or use inpainting.
Even though I don’t find Ideogram’s text editing tool that useful, it is an example of an interesting trend for AI tools to introduce more fine-grained user control than pure prompting. We’ve seen similar ideas in video generation, with camera controls in Kling, Hailou, and many others services.
Kling’s “motion brush” tool is an example of the hybrid UI that is getting more popular in AI tools, as I predicted in 2023. Here, I specified the path I want the bird to fly when Kling generates a video based on this image. It is much easier to draw a motion path than to describe it in words through a prose prompt.
AI-UX Conference Berlin
The prompt:UX 2025 conference will be in Berlin, Germany on April 8-9, 2025. The event has a strong speaker lineup covering interesting topics.
There was very little good information about how to design AI-based products or features in 2023 and 2024, but I am happy to see that more good events are being produced. Some books are also coming out, though I worry that this old media form will have difficulties keeping up with the rapid improvements in AI.
A conference on the intersection of AI and UX in Berlin in the spring; should be nice. (Midjourney)
EU Multilingual LLM
The EU Commission has announced that it is spending €52 Million (USD $56 M) on a new AI called OpenEuroLLM which will cover all the EU languages. While it’s great to (maybe) get better AI for smaller languages like my native Danish, this is too little, too late. The United States alone is investing more than $300 Billion in AI this year, and China is also going hardcore on AI.
I think the EU bureaucrats misunderstood the letter needed to be an AI leader. The number of dollars (or Euros) should be followed by a “B,” not an “M.” This is even more true given the inefficiency of EU projects. It’s admittedly 35 years ago, but I used to participate in EU-funded projects and they were inevitably mediocre because the Commission insisted on including an incompetent research team in Greece for completing an essential part of the project. (Greece only mentioned as the example I remember best, but there were incompetent researchers from random countries in all the projects I knew about.)
For better results, EU should have recognized that they already have a promising AI company within its borders in the form of Mistral. It has been falling behind lately with all the exciting releases from the U.S. and China, but Mistral is still my best bet for a strong European AI model.
If you only have one piece of gold to spend, invest it wisely in the one best team, instead of spreading it around to 20 teams, half of which are incompetent. (Grok)
Given their crazy “AI Act,” I don’t have much faith in Europe’s ability to build the future, but one can always hope. My main hope for competition to break the America-China duopoly currently rests with the United Arab Emirates and/or Saudi Arabia, because they have both the money and the energy resources it takes to win in AI.