#139: Die Tasse Kaffee
Germany wants to go back to La La Land
Hello.
You are reading Understanding TikTok. It is me, Marcus. A short federal election campaign in Germany, shaped by open (AlgorithmWatch) and covered (Correctiv) foreign influence operations (New York Times) presumably jointly responsible for the strongest results for a far-right party ever since the Nazis and a general right-shift (The New Statesman), came to an end. Just like the entire postwar order (The Atlantic). Let’s talk about:
🗳️ The German 2025 Elections
🤖 Using ChatGPT for analysing TikTok data
🍏 What else
🗳️ The German 2025 Elections
The defining sound meme on German TikTok right before the federal election has been a nostalgia-infused “we meet again years after a break-up and have a coffee” pop song performed by Florian Silbereisen and Helene Fischer who dated in the noughties. A perfect metaphor for a general “take me back to when everything was somewhat okay” vibe even though it never was – just in time of geopolitical upheaval. Nearly every party has used the song in their campaigns with Olaf Scholz and team using an Anti-Afd parody version of the song one day before the election on his official account.
First take-aways on election eve:
TikTok for political campaigning has gone full mainstream in 2024/25
TikTok’s cultural impact is undefeated with an older target audience growing
Political campaigning in Germany has never been so meme-infused
Political messaging from posters to platforms is heavily fragmented
Politicians and parties finally and fully embrace the platform era
The role of sound memes needs further investigation. Hook me up!
Despite TikTok’s assertions to “protect election integrity” (stop me if you think that you’ve heard this one before) problems already pointed out after Germany’s last federal election in 2021 (Mozilla) still exist. Several studies point out:
"TikTok displayed the greatest right-wing skew, per its findings — showing right-leaning content 74% of the time." (Global Witness Study, Techcrunch, Nieman Lab)
“TikTok classifies only 59% of the content of party or candidate accounts and only 47% of the content of fan pages as political.” (ISD Germany)
I am currently scraping data for a more in-depth report on The German TikTok Elections 2025 that will be published by Friedrich Ebert Foundation in the coming months. Wish me luck on my deep dive where i will investigate stuff like this or this this or this.
🤖 Using ChatGPT for analysing TikTok data
Analysing the complexities of multimodal, layered and dense video content poses challenge to researchers. Large language models like ChatGPT might be of help here. But how? I asked Carlo De Gaetano (IT) – designer and researcher with the Visual Methodologies Collective at the Amsterdam University of Applied Sciences.
Carlo presented his experiments in a workshop at a winter school on AI Methods at Siegen University (slide deck) in February 2025 that i had the pleasure to attend. This is the condensed version of a longer Q&A with Carlo.
Hi Carlo, how are you using ChatGPT to analyze TikTok data?
One of my recent experiments focused on maps in sea level rise videos—a recurring visual trope on TikTok. I wanted to test whether ChatGPT could recognize place names, countries, and regions just by looking at maps. After processing around 30 thumbnails, I asked ChatGPT to compile the identified geographical entities into a structured dataset. I was pretty excited by the results. Read more in the full interview here.
How can ChatGPT be most helpful here?
For me, ChatGPT is most useful as a thinking partner—it helps structure and refine my research process, making it easier to go from a messy collection of videos to a more organized way of analyzing them. One of the simplest but most practical ways ChatGPT helps is with search queries. Finding the right videos on TikTok isn’t always straightforward, and hashtags or captions don’t always reflect the full content of a video. ChatGPT helps me brainstorm in the early stages of research to find precise keywords and hashtag combinations that improve my chances of pulling in the most relevant material.
Once I have a dataset, ChatGPT assists in developing workflows for analyzing the collection. As I have already mentioned, since I can’t upload videos directly (for now), I have to rely on extracted frames, video thumbnails, or timeline snapshots to represent the content visually. ChatGPT can help sketch out different ways to process these images systematically: for example, the suggestion to track color transitions over time to detect sea level rise maps came from ChatGPT, and was then refined collaboratively through iterations.
What surprised you the most when analyzing TikTok data with ChatGPT?
One of the biggest surprises was how well ChatGPT remembers past conversations and can build on them. When working on a research project that spans multiple sessions, it’s easy to lose track of earlier ideas or minor refinements that were made along the way. But ChatGPT often recalls previous discussions, references them, and even reminds me of things I had considered before but didn’t follow up on. This makes it feel more like an actual research collaborator rather than just a tool for answering isolated questions. Another unexpected aspect is that ChatGPT doesn’t just give answers—it helps refine questions. When researching something as visually complex as audiovisual collections, it’s not always clear what the best approach is. Instead of just providing pre-formed insights, ChatGPT acts as a sounding board, helping me clarify what I’m looking for and suggesting different ways to structure an analysis.
Also, I was surprised by how much creativity is required when working with AI in this way. Since ChatGPT can’t directly analyze video, a lot of the process involves figuring out workarounds—how to extract meaningful data from videos in a way that AI can process. As an information designer, this has made me think more deeply about how research is structured and has pushed me to develop new methods I wouldn’t have considered otherwise.
What are common mistakes when using ChatGPT for TikTok analysis?
I can share some of my own mistakes while using ChatGPT as a visual assistant. The first one was trying to ask for everything in a single prompt. Another mistake is not refining prompts enough. A vague question leads to a vague answer. For example, when I first started asking ChatGPT to describe TikTok thumbnails, I wasn’t specific enough about the kind of details I wanted. This meant I got overly generic descriptions that didn’t really help my research. But after a few rounds of tweaking, I realized that giving explicit instructions on what to focus on—colors, text overlays, recognizable landmarks—made a huge difference.
I learned the hard way that uploading too many images at once makes things go sideways. At one point, I tried uploading a zip file containing 200 TikTok thumbnails, expecting ChatGPT to neatly categorize them. Instead, I found myself in a strange situation where ChatGPT confidently made up descriptions that didn’t match the images at all—a perfect example of how LLMs, as Harrison (2023) puts it, are "mansplaining machines: often wrong, yet always certain." It was frustrating but also a good reminder that AI models don’t “see” images in the way we do—they process metadata, pixel data, and patterns, but they can still hallucinate descriptions when overloaded. So, the lesson? Take it step by step, test small batches first, and don’t assume the AI is always right—because sometimes, it’s just confidently wrong.
Would you mind sharing 1 or 2 tricks or tips when analyzing TikTok data with ChatGPT?
There is a misconception that because ChatGPT is presented to us as a machine, it must be great at math—which is absolutely not the case. I quickly learned that asking it to count how many times a certain object appeared in a timeline led to nonsense results. It’s much better at what it’s actually designed for—describing images, summarizing text, and categorizing things using a few examples. A useful strategy when analyzing videos is to build a structured tagging system: when working with a collection of TikTok videos, it helps to design a clear system for categorizing images. Instead of expecting ChatGPT to "figure out" a pattern on its own, I use the few-shot learning method, where I manually tag a small set of images with labels first and then have ChatGPT apply those categories to the rest of the dataset.
Another key strategy is to combine AI with other tools to scale up the research. As a visual research assistant, ChatGPT works best when it’s used in combination with other software rather than as a standalone solution.In this way, ChatGPT does not do the heavy lifting itself but helps outline the logic and structure of the solution, while Python hands the execution. AI doesn’t replace manual work, but it can massively speed up and refine the process when used strategically.
Thanks so much!
🍏 What else
There is a new TikTok-Scraper out there (Github)
What we know about TikTok content creators (Pew Research)
Politicisation of influencers workshop coming up (Utrecht University)
Why TikTok’s loss might be Instagram’s gain (DigiDay)
Wie beeinflusst TikTok die Wahlentscheidung? (BR)
That’s it for now. I am in Innsbruck next week for the Annual conference (Jahrestagung) 2025 of the specialist groups “Communication and Politics” (DGPuK), “AK Politics and Communication” (DVPW) and “Political Communication” (SGKM) to present on TikTok & participatory propaganda in the context of elections campaigns. Are you nearby? Come on over. Speak soon. Ciao
P.S. Anyone else’s FYP escalating between Snailrave and Party 4U?