After redefining and adjusting the idea, it was time to give shape to all these thoughts. Before thinking of colours, typefaces or any kind of identity, the first step was to create a base layout for the following designing phase. The wireframes have been made in Figma and kept as basic as possible, with placeholder texts, simple shapes and an overall idea of its web architecture. In this way, the attention remains on what the tool actually does and how the user moves through it, rather than how it looks.
The tool/website is still based on the two main sections mentioned in the previous post: a survey part to gather quantitative data and a page with open access to them. The open data hub was imagined as dashboard where people can browse the results, which visualization will be particularly curated to ensure clarity.
The result I reached was a fine basis for its further development, even if it could have been more detailed in some sections (e.g. downloading process, better about page, more questions typologies) and texts. Indeed, trying to keep the actual copy in the areas I designated in the wireframes was an actual challenge. Maybe in the future would be better to already have texts, or at least the most similar version to the official one.
he heuristic evaluation in the last post named five UX patterns and grounded each one in a mechanism and a source. But naming a pattern in a research document and getting someone to actually recognize it while scrolling are two different problems. To close that gap, I built something more concrete than a catalogue: a wireframed prototype that simulates the experience of being algorithmically escalated, and then tests whether people can spot what just happened to them.
The prototype is a fictional social media feed, structured in two phases.
Phase one is a straight scroll through five posts(in a more refined version probably more), with no annotation or warning. The content itself escalates in a deliberate sequence: it opens with something completely mundane (a weekend BBQ post), moves through generic economic frustration (a rising energy bill, “does this ever stop?”), into nostalgic in-group framing (“our grandparents built this country… share if you see it too”), into outrage-framed conspiracy (“the media won’t cover this, ask yourself why”), and ends on an explicit call to join a named movement. Two things escalate in parallel with the content itself: the engagement numbers (from 235 likes and 15 comments on the first post to over 74,000 likes and 8,400 comments on the last) and the comment threads underneath each post, which shift from neutral to increasingly validating (“finally someone saying this,” “I knew it,” “count me in”).
After the scroll, users hit a short reflection screen: “What did you just notice?”, with three answer options. It’s there to test whether the escalation registers consciously at all, or whether it just works on people the way it’s designed to.
Phase two replays the identical feed, but now the posts carry annotation boxes naming the specific mechanism, for example, labeling the energy-bill post as an “Emotional Trigger” (frustration without a named target, which the algorithm treats as highly relevant because it drives comment activity), the grandparents post as “Social Validation + Frictionless Sharing” (the like count signaling consensus, the oversized share button removing any pause before spreading it further), and the media-conspiracy post as “False Credibility + Outrage Architecture” (borrowing the visual language of legitimate news to make a vague accusation feel credible). This second pass is the direct, hands-on version of the heuristic evaluation from the previous post, the same three mechanisms (amplification, validation, frictionlessness) made visible on the exact content that used them.
At this moment, this prototype has only been walked through informally with three fellow design students, not yet with any of the young people the tool is actually meant for.
Still I gained valuable feedback from my peers:
The engagement-number escalation is probably the easiest thing to miss. Peers who read UX interfaces daily are likely to focus on the text of the posts rather than the like/share/comment counts sitting quietly to the side, meaning even a design-literate tester might land on “the content got more extreme” while missing “the posts had more and more likes,” even though both are true and both are the point.
The “Join the Group” call-to-action is probably the moment that lands hardest. Everything before it is observational; that button asks for an action. It’s a reasonable guess that this is where testers would say the discomfort became concrete.
The annotated second pass likely produces a genuine “oh” reaction, precisely because pass one deliberately withholds any framing. The value of the two-pass structure is exactly this contrast. But it also means the specific wording of the annotations (how technical vs. plain the labels are) probably deserves the most iteration.
Peers may find the labels (“Emotional Trigger,” “Outrage Architecture”) a bit academic for a tool aimed at teenagers. As they are now, they are very technical and too complex for my actual target group.
Outlook: Testing With Actual Young People
In my next steps I would refine my current prototype and start testings with my actual target group. This could look like the following:
Small workshop groups (4–6 participants, ages roughly 14–17) rather than one-on-one sessions, since group discussion after the reflection question is likely to surface disagreement about what counted as “extreme” or “unusual”.
Co-facilitation with a media-literacy or extremism-prevention professional, given that even a fictional simulation of this content needs someone in the room who can handle a participant reacting more strongly than expected, and a clear opt-out built into the session, not just offered at the start and forgotten.
Interviews with people that are part of the extremist movement. This is one part of which I am very unsure of, but I believe I could gain valuable insight from people that are on the political right. Understanding how they got there could help me create a tool for prevention. This part requires the most preparation, as I would need to find a way for people to talk to me in the first place.
The context review in the last post mapped out where and when algorithmic radicalization has been documented across YouTube, TikTok, and X. This post does the actual evaluation: taking Gray et al.’s (2018) dark pattern taxonomy, nagging, obstruction, sneaking, interface interference, and forced action, and adapting it into a working heuristic grid, then applying that grid to five specific interface elements drawn from the platforms reviewed above.
Rather than asking the standard usability question (“does this help the user complete their task efficiently?”), each element below is evaluated against three project-specific criteria derived from LaCroix and Berkovits’ (2026) pipeline model (see Post 2):
Amplification: Does the element systematically increase exposure to escalating content?
Validation: Does the element convert exposure into a social signal of identity or belonging?
Frictionlessness: Does the element remove the pause a user would need to reconsider before acting?
Autoplay Chaining (YouTube)
Autoplay queues the next video automatically, based on the recommendation model, without requiring any user action. Against the grid: high amplification, low validation (it’s not a social action), high frictionlessness (zero taps required to keep watching). In Gray et al.’s taxonomy, this most closely resembles forced action combined with a passive form of nagging, the system continues the experience on the user’s behalf unless they actively intervene.
The “For You” Feed Reset (TikTok)
Unlike a subscription feed, TikTok’s FYP has no stable “unsubscribe”. Every swipe is a fresh optimization pass. Baumann et al. (2025) found strong content amplification setting in within roughly 200 videos. Against the grid: very high amplification, low validation, very high frictionlessness (a swipe requires less effort than a click). This is closer to interface interference, the feed’s ranking logic is invisible to the user, so there’s no way to see, let alone interrupt, the drift while it’s happening.
Engagement Counters and Like Badges (cross-platform)
Visible like/share/comment counts turn viewing into a public scoreboard. This is the element LaCroix and Berkovits identify as converting resentment into performative identity. Against the grid: low direct amplification, very high validation (the entire function of the element is social signaling), moderate frictionlessness (liking is one tap, but the effect persists after the tap). This overlaps with social proof exploitation, a pattern documented across e-commerce dark pattern research (Mathur et al., 2019) that maps cleanly onto political content: a high counter tells a new viewer “many people already validated this,” which lowers their own threshold for engaging further.
One-Tap Resharing
The mechanical simplicity of turning a private reaction into a public post. Moderate amplification (it multiplies existing content rather than generating new exposure), high validation (sharing is itself an identity act), maximal frictionlessness, this is the exact mechanism LaCroix and Berkovits single out as removing “cognitive friction” between private conviction and public propagation. In dark pattern terms, this is the near-total absence of what Gray et al. would call a deliberate obstruction, most dark pattern research studies obstruction added to slow users down (e.g., making cancellation hard); here the notable design choice is the deliberate removal of a step that would ordinarily exist.
“Trending” and “Recommended for You” Badges
A badge asserting that content is popular or personally relevant, regardless of how that judgment was computed. Against the grid: moderate-to-high amplification (badges are themselves a ranking signal shown back to the user), high validation (trending implies social consensus), high frictionlessness (the badge pre-empts the user’s own evaluation of the content). The actual basis for “trending” (engagement velocity, not necessarily quality or accuracy) is concealed behind a label that implies broad, neutral endorsement.
It was tempting, once I’ve decided to study radicalizing UX patterns, to jump straight into cataloguing dark patterns. Screenshot a manipulative feed, label it, move on to the next one. But before any pattern can be named responsibly, I needed to understand the context, and that’s where a context review comes in.
Before naming any concrete UX pattern, this project ran a context review across three platforms, YouTube, TikTok, and X (formerly Twitter), to understand when, how, and under what conditions these dynamics were actually documented. Patterns don’t mean the same thing in every era; a review has to establish what changed, and when.
YouTube
The earliest and still most cited empirical anchor here is Ribeiro et al.’s 2020 audit of YouTube’s recommendation system. The researchers tracked user migration across three overlapping communities, the Intellectual Dark Web (I.D.W.), the Alt-lite, and the Alt-right, using more than 72 million comments and over 2 million video and channel recommendations collected between May and July 2019. The context matters: this was the period when “algorithmic rabbit hole” had already become a popular metaphor (helped along by public testimony from former YouTube viewer Caleb Cain about his own radicalization), but there was no large-scale empirical confirmation of how the pathway actually worked mechanically. Ribeiro et al. supplied that: users measurably migrated from milder I.D.W. content toward more extreme Alt-right content over time, and channel recommendations, more than video recommendations, were the specific mechanism connecting Alt-lite audiences to Alt-right channels.
Later audits (e.g. Hosseinmardi et al., 2021; Ledwich & Zaitsev, 2020) complicated this picture, finding smaller or more contested effects depending on methodology. For a context review, that disagreement is itself useful information: it tells us the “pipeline” isn’t a fixed, permanent feature of YouTube’s algorithm, but something that varies with the specific recommendation model in place at a given time, and with how the audit itself is designed.
TikTok
TikTok’s context is different in almost every respect: newer platform, shorter content, faster iteration cycles. Media Matters’ 2021 investigation seeded a fresh account with exclusively transphobic content and, within a few hundred recommended videos, found the “For You” page surfacing extremist and far-right material, including content referencing violent figures. That single-account case study was followed by more rigorous quantitative work: Baumann et al. (2025) ran a sock-puppet audit and found that TikTok’s algorithm produces strong content amplification along a bot’s expressed interests within the first 200 videos watched, with content diversity declining sharply after that point. Shin and Jitkajornwanich (2024) reverse-engineered TikTok’s recommendation logic specifically for far-right content and found multiple distinct pathways feeding it, largely traceable to platform recommendations rather than active user searching.
The context here, a platform built entirely around a single, extremely responsive short-video feed, optimized almost exclusively for watch time, helps explain why the numbers look faster and starker than YouTube’s: 200 videos on TikTok is an afternoon, not months of viewing history.
X
X’s context is the most politically visible of the three, partly because ownership and moderation policy changed dramatically after the platform’s 2022 acquisition. Huszár et al.’s 2022 study (discussed in Post 1) predates that change and already found algorithmic amplification of right-leaning content in six of seven countries studied. More recent investigative work, including a 2026 cross-platform audit by People vs. Big Tech comparing X and TikTok in the French political context, found continued asymmetric visibility for far-right and radical-left accounts relative to moderate parties, independent of follower counts. The context review here has to hold two things at once: the underlying amplification dynamic Huszár et al. identified pre-dates the platform’s ownership change, and post-change investigations suggest it has, if anything, become more visible rather than less.
Running this review rules out treating “algorithmic radicalization” as one uniform phenomenon with one uniform UX signature. YouTube’s pattern is slow and comment-thread-mediated; TikTok’s is fast and almost entirely feed-driven; X’s is amplification-plus-visibility in a much more overtly contested political environment. Any pattern catalogue this project produces needs to specify which platform context it was observed in, and roughly when, otherwise a pattern documented on 2019 YouTube gets silently treated as if it still describes 2026 TikTok, which the context review shows is simply not a safe assumption.
In the first post of this series, we established that UX design is never politically neutral. But amplification is only half the story. The more unsettling question is what happens after someone starts seeing more of this content: how do they go from scrolling to sharing, from watching to believing, from believing to spreading?
LaCroix and Berkovits (2026) offer a useful concept for this: the Digital Radical, an individual who has made the transition from passive consumer of extremist content to active propagator of it. Their argument is that this transition isn’t purely psychological or ideological. It is actively encouraged by the architecture of social platforms.
They break this process down into a kind of pipeline, each stage powered by a specific design feature:
Recommendation engines move users from generic, everyday frustration toward ideologically charged resentment. The frustration was already there, the algorithm gives it a direction and a villain.
Like and share systems convert that resentment into performative identity. Engaging with extreme content becomes a public signal of who you are.
One-click sharing removes the last barrier. What used to be a private conviction becomes public propagation, instantly, without any cognitive friction to slow the person down and make them reconsider.
The conclusion LaCroix and Berkovits draw is blunt: the interface itself is the medium of radicalization. Not the ideology alone. Not the individual alone. The interface.
This is arguably the most design-relevant insight in the whole framework. Radicalization researchers have long focused on content and community. But LaCroix and Berkovits point to something UX designers understand intuitively: friction, or its absence, determines behavior. Every “are you sure?” prompt, every extra tap, every moment of pause is a design decision. Removing friction isn’t inherently sinister; it’s usually done to boost engagement metrics. But when the content being frictionlessly shared is extremist, the design choice that was meant to drive growth ends up driving radicalization instead.
It would be a mistake, and an oversimplification this project wants to explicitly avoid, to conclude that platform design causes right-wing extremism. It doesn’t, at least not on its own. Radicalization emerges from an interplay of individual vulnerabilities, social context, and political crises (Youngblood, 2020; Banywana, 2026). Digital infrastructure is an increasingly important factor in that interplay, not a standalone explanation for it.
But “not the sole cause” doesn’t mean “not worth studying.” Digital platforms are, for a growing share of the population, the primary place where political information is absorbed, social identity is formed, and civic participation happens. Whoever doesn’t understand the design logic behind that space is simply at its mercy. Whoever does understand it can start to question it as a user, as a designer, or as a policymaker.
This is the communicative and educational ambition behind the project: building awareness of how platform design encourages radicalization, as a precondition for pushing back against it, individually, and at the level of policy and platform design itself. Concretely, this means identifying and classifying radicalization patterns in interface design, and eventually building a tool that strengthens users’ digital media literacy, aimed especially at young people, whose critical awareness of digital media still needs active support.
When we talk about right-wing radicalization online, the conversation usually centers on content: which videos, which posts, which influencers. What gets less attention is interface itself. This blog series starts from a simple but often overlooked claim: UX design is not politically neutral.
Every digital platform is built on thousands of small design decisions. How a feed is ordered, what gets a notification, how easy it is to share something with one tap, none of these choices are accidents, and none of them are ideologically empty. They are the product of algorithmic personalization, social validation mechanisms, and platform architecture. Together, these design layers quietly shape what we see, how we react to it, and who we become as users.
The most extensive study on algorithmic amplification of political content comes from Huszár et al. (2022), who analyzed how Twitter’s recommendation algorithm treated political content across seven countries. Their findings showed, that in six of the seven countries studied, right-leaning and conservative accounts were amplified more strongly by the algorithm than their left-leaning counterparts. The differences in reach were statistically significant and consistent across contexts.
What’s important here is why this matters for design research specifically. The study doesn’t tell us definitively why this asymmetry exists, the underlying mechanism is still debated. But it tells us something arguably more important: the design of the algorithm produces this inequality, regardless of what its engineers intended. Nobody needs to sit down and decide “let’s amplify right-wing content” for a system to end up doing exactly that. The outcome emerges from the architecture itself.
This is the core problem with treating UX design as a purely technical, apolitical discipline. A recommendation engine has to rank content somehow. A feed has to be ordered somehow. There is no version of these systems that makes zero choices, “neutrality” isn’t actually real nor feasible in this context. Every design decision has political consequences, whether the designers intended them or not, and whether they’re even aware of them or not.
For a master’s project sitting at the intersection of design and political radicalization, this reframing matters. It shifts the question away from “did designers mean to help radicalize people?” (usually no) and toward “what does this specific interface structure do, mechanically, to the people using it?” That’s a question we can actually study, pattern by pattern.
If design decisions produce measurable political effects, regardless of intent, then understanding which decisions produce which effects becomes a genuinely urgent task. That’s the gap this project is trying to fill: not just showing that platforms behave asymmetrically, but building a vocabulary that names exactly which design choices drive which radicalization dynamics.
I went to the library of the University Ulm yesterday to get my hands on the book “Überwachen und Strafen” by Michel Foucault, but I wasn’t able to find it, When asking for it at the front desk I was told that I have to order it first, because it’s not accessible in the public shelves there. Luckily, my sister studies at the university of Ulm, so she will order it for me and also borrow it so I can read it at home. But it will take a while until I get it, so I won’t be able to read it before the deadline for these blogposts. I will have enough time to read through it though during the summer break.
The Trainship Book
Unfortunately, I was sick during both “Bring Your Prototype” sessions this semester, and in the days leading up to them as well, so I didn’t sit down to develop my first 30-minute prototype until Friday. For this, I went back to the long-distance train ride setting, since I’d already had an idea for it some time ago.
The idea is to create a kind of “friendship book” that passengers can fill out during the train ride. It features questions that everyone might secretly ask themselves while observing their fellow passengers on the trip. Train passengers constantly construct stories about the people around them. They infer destinations, relationships or emotions based on clothing, luggage or behaviour. The friendship book shifts this process from imagination to voluntary self-disclosure. Instead of guessing, travellers can decide which parts of their story they want to reveal. It serves several purposes at once: On the one hand, the book is an interesting way to pass the time on long trips; on the other hand, it satisfies people’s natural curiosity without requiring direct interaction. The book can still serve as a conversation starter without forcing social interaction.
I’ve come up with the following questions for the book:
The Layout
There will be a short introduction text on the first page, explaining the concept of the book and how it’s meant to be used. Then there will be double pages for the persons to fill out with questions related to their trip. The personal section on top will be very short, but I still decided to include it, so the readers would get a rough image about who that person is. After that it starts with some basic questions and goes more into detail when it comes to the reason why the person is travelling to this destination and how what is going to await them makes them feel. I think if the travellers feel comfortable to be honest here, it could create a sense of authenticity and connection, if other people find themselves in similar situations or struggeling with the same anxious thoughts, for example. Then there will also be some pages where people can tell stories about the best or funniest thing that happened to them on the train. This could as well be very entertaining to read. I also thought about including a spread where people can recommend songs to each other which they then can listen to while on the train. It might also be nice to sort the songs into different categories depending on the current mood, e.g. “Listen to this song when you are sad.” or “Listen to this song when it’s raining outside”.
The Prototype
I did a quick 30-minute prototype of the three spreads I described earlier. I tested it on the way back from Ulm, but I forgot the filled out version at my sister’s place, so I will have to ask her to send me a picture of it.
The Concept
There is different options when it comes to where the book is going to be placed / who owns it.
OPTION 1: The book belongs to an individual person and they can bring it on train rides whenever they feel like it and then just ask people to write in it. That way they “collect” friends on their journays over time. This might be a good option for more extrovert and outgoing people, but also for people who wish to engage more in conversations with strangers, as it causes active social interaction. This is also the option that I tested on my way to Ulm on the weekend.
OPTION 2: There is several books on the train that stay on fixed spots, so each person sitting in that specific seat would fill out the book. That way people would find out things about the people that were in the exact same place in the past, which could be also very interesting, since it causes asynchronous interaction.
OPTION 3: the books would be placed at the entrances for people to take and give to other people. These books would be more like booklets, only for this one trainride to keep as a memory afterwards. What is interesting about this option is that it connects more than two people on the same train ride.
My favourite would be option 2, since it’s the most appealing to me personally. I would enjoy to read about the people who were there before me. Apart from that I think one of the advantages is that after a while there is many filled out pages to read through, so it might be more interesting. Additionally, it serves as a temporary pastime for many people and is therefore more sustainable than individual books, which may eventually be forgotten or end up in the trash because they’re too much of a burden to carry while traveling.
The Testing
I found it easier to talk to someone on the train this time, because I had a “valid” reason for it. I realised that the prototype did not only test the book itself, but also gave me permission to approach a stranger as a way of “social excuse”, lowering the barrier that usually prevents conversations between passengers.
The journey from Augsburg to Ulm only takes around 45 minutes, which made the situation feel less intimidating, as I knew I would be leaving the train soon anyway if the interactionbecame awkward.
There was a girl around my age sitting next to me, so after briefly explaining my project, I asked if she would be willing to fill out the prototype pages I had prepared the evening before. Although I was a little nervous at first, but I soon realized that we were getting along quite well. While she was filling out the pages, we started talking about our experiences of travelling by train and realised that neither of us would usually start a conversation with a stranger unless we were approached first.
She told me that she found the questions interesting and enjoyed answering them. Nevertheless, I plan to conduct a small survey to evaluate which questions people find most interesting and whether some should be revised or replaced.
Overall, the prototype confirmed that the friendship book can work as an effective conversation starter. It provides a clear reason to approach someone while still making it easy for the other person to decline without feeling uncomfortable.
I wondered whether the book functions more as a social mediator than as the actual activity itself. In this particular situation, I think the physical object was probably more important than the questions it contained. Having a tangible artefact provided a natural reason to approach someone and made starting a conversation feel much less awkward. The questions mainly served as a gateway to further discussion about train journeys and interactions with strangers. However, I believe this would change if the books became a permanent feature on trains, as proposed in option two. In that case, the content of the questions would likely become much more important, since passengers would spend more time reading previous entries and contributing their own thoughts.
Would the same conversation have happened without the physical book? Probably not, or at least it would have been much more difficult to reach that point. Without the book, I would most likely have had to start with conventional small talk before changing the topic to how someone feels about talking to strangers.
Limitations
Of course, this was only a very small test, so I cannot draw any general conclusions from it yet. I only asked one person to participate, who also happened to be around my age. She was open to talking to me, so the situation was probably easier than it could have been with someone else. I also tested the prototype during the day on a relatively short train ride. I think the results could be different on a longer journey, in a crowded train, or with people who are in a hurry or simply do not want to interact. If I want to stick with the idea of the trainship book I will definitely have to test the concept with more people before I can properly evaluate it.
Continuing my research on the topic of “Slowness”from last semester, I will be doing further work on this topic as part of the Design & Research II course with Birgit Bachler PhD.
For the final assignment, I will need to record a 2-minute video of the final prototype — in my case, I plan to use prototyping in Figma, which each of you will be able to test. The research will again focus more on contextual analysis, as well as user testing, bibliography, and heuristic evaluation.
During the development of my prototypes, I realized that presenting separate activities was not enough. Each activity explored a different aspect of children’s experiences, but they did not naturally connect with one another.
Instead of treating them as individual prototypes, I started combining them into one participatory toolkit. The goal changed from asking isolated questions to guiding children through a simple design journey.
The toolkit now follows three connected steps;
First, children choose a material or playground element that they would like to include. They then select a feeling that best represents how this element should make them feel. Finally, they choose a type of play they associate with it. Rather than giving fixed answers, the activity encourages children to create their own combinations and explain their choices.
After completing these steps, children receive a blank sheet where they can use the selected cards as inspiration to draw or design their own playground.
Although the prototype is still low-fidelity, I think structuring the activities into one continuous process makes participation much more natural and easier for children.
