Worried About Germany? You Should Be. What Can We Actually Do?

written by Stefan Christoph

- 12 minutes read
TL;DR: Yesterday the AfD won the Saxony-Anhalt state election with 43.8%. Its state branch has been formally classified as confirmed right-wing extremist since 2023. If that worries you, good. It should. This is an unusual post for this blog, and it doesn’t end with an answer, because I don’t have one. It ends with a question for the people I work with: the machinery that rewards outrage and simple stories is built and run by our industry, so what can IT and media people actually do? I went looking and came back with five research-backed starting points. I want your objections more than your applause.

A quick note before anything else: this is an unusual post for this blog. I normally write about building things, like architecture, AI, and the occasional war story from production. This one is different, and I want to be upfront about something. It doesn’t end with an answer. It ends with a question I need help with.

The result, briefly

On 6 September 2026, the far-right AfD won the Saxony-Anhalt state election with 43.8% of the vote: 39 of 83 seats, three short of an absolute majority [1]. The party took 38 of the state’s 41 direct constituencies, up from a single one in 2021. The governing CDU fell to 17.2%. Turnout hit 77.8%, an all-time record for the state and 17.5 points higher than in 2021 [1].

That last number matters, because it forecloses the comforting explanation. This wasn’t a quiet election that a motivated minority slipped through. Turnout was the highest ever recorded here, so the result can’t be explained away by unusually low participation: 43.8% of valid votes went to the AfD.

Where I stand

I want to be precise here, and I want to own it as my own view rather than dress it up as fact.

Frustration with the established parties is understandable. They have genuinely underdelivered for a long time, in ways that are easy to feel and hard to dismiss. I don’t think anyone who voted out of that frustration is stupid, and I’m not interested in a post that treats them that way.

But from my perspective, voting for an extremist party out of frustration crosses a line. Given German history specifically, voting far-right is not something I can treat as just another protest choice. This is where I stop hedging. The state’s own Office for the Protection of the Constitution has classified the AfD’s Saxony-Anhalt state branch as “gesichert rechtsextremistisch” (confirmed right-wing extremist) since November 2023 [2]. That refers specifically to the state branch, under state law. It is not a slogan, and it is not the separate federal-level question that is still being fought over in court.

So if you’re worried about Germany this morning, you’re right to be. I am too. The question is what we do with that worry, and that’s where this post stops being about one election.

This isn’t a post about party politics

I’m not going to analyse coalitions, prescribe policy, or tell you who to vote for. Party politics is largely outside what people like me can influence, and there are people far better placed to argue it.

What I can talk about is our craft. Most of the people who read this build software, run infrastructure, design products, or work somewhere in the information and media industry. And the uncomfortable truth is that the environment this result grew in, one where the loudest, angriest, simplest version of every story travels furthest, is an environment our industry built. Which means it’s also one we can help repair.

The machine isn’t neutral

Let me be careful with this claim, because the honest version is narrower than the viral version.

Social media did not cause this election result. Anyone who tells you a feed elected a parliament is selling you the same kind of simple story the feed itself rewards. But the machinery underneath is not neutral, and we have decent evidence for how it tilts.

In a pre-registered audit of Twitter’s ranking, researchers found that the engagement-based algorithm amplifies emotionally charged, out-group-hostile content: the kind of content that users themselves said made them feel worse about the other side. The same users, asked directly, did not actually prefer the posts the algorithm had chosen for them [4]. The system optimises for engagement, and in this audit, what won engagement was not what people said they wanted.

And attention in this campaign was strikingly concentrated. Researchers at the University of Potsdam tracked political TikTok accounts and found that roughly 65% of all campaign views in Saxony-Anhalt went to a single account, the AfD’s lead candidate, while the party that posted the most content, the CDU, drew about 9% [3]. One account, two-thirds of the attention. To be exact about what that shows: concentration, not cause. Views aren’t votes, and the numbers don’t prove the Twitter mechanism was at work here [3].

So no proof of causation, and I won’t pretend otherwise. What the evidence describes is the terrain: reach concentrates around whoever masters the medium, and the medium’s ranking machinery has measurable tilts. That terrain is not weather. It’s the product of design decisions, and people like us make them.

Five starting points, not a checklist

If the terrain is partly ours, so are some of the tools to reshape it. What follows is not a program, and it isn’t me telling our industry what to do. These are the five starting points I found when I went looking for evidence, and I’m putting them up for discussion: Are they the right ones? Is something important missing? Would some of them not survive contact with reality? The evidence differs in kind, too: intervention studies for the first three, deployment evidence for provenance, a ranking audit for the last.

1. Media literacy works, and it’s teachable at scale

The reflexive objection is that “you can’t teach people to think.” The research disagrees. A 2025 meta-analysis reviewed 160 media-literacy interventions spanning 40 years and found overall positive effects, with the strongest gains in knowledge and critical evaluation of media; for those outcomes, the time elapsed since the intervention did not reliably weaken the measured effect [5].

The most prominent national example is Finland: media literacy woven into the curriculum from early childhood, and first place on the European Media Literacy Index every year since the index began in 2017 [6]. Coexistence isn’t causation, but it shows a country can decide to build this capability deliberately, at national scale.

This is the slowest lever and probably the deepest. If you teach, mentor, run onboarding, or explain technology to anyone, you’re already holding a piece of it.

2. Prebunking: teach the trick, not the claim

Fact-checking fights yesterday’s falsehood. Prebunking, or inoculation, builds resistance in advance by teaching the manipulation techniques themselves: scapegoating, stripping context, discrediting the messenger. Once you can recognise the move, you don’t need a fact-checker for every individual claim.

This isn’t theoretical. A 2026 study ran 13 surveys across 12 EU nations testing short inoculation videos from a campaign that reached more than 120 million YouTube users before the 2024 European elections. The videos measurably improved people’s ability to spot manipulation, and learning to spot one tactic helped against others too [7]. The effects were small and varied by country (more on that below), but they were real. And this is the kind of intervention that can live inside a product or a feed rather than in a classroom. If you ship anything with a feed, a timeline, or an inbox, that’s yours to try.

3. Evidence can change minds

There’s a fatalistic belief that you can’t reason anyone out of a position they didn’t reason themselves into. The best recent evidence pushes back hard.

In a 2024 study published in Science, people discussed their conspiracy beliefs with an AI system that engaged them with tailored, factual counter-evidence. Belief in the conspiracy dropped by about 20% on average, the effect held two months later, generalised to unrelated conspiracies, and worked even for people whose beliefs were deeply entrenched [8].

At least in this setting, facts moved people when they were delivered patiently, specifically, and in response to the person’s actual reasons rather than fired into a crowd. One study isn’t a law of nature. But if you build with LLMs, it’s an existence proof worth sitting with: the same technology everyone fears as a misinformation firehose has, under controlled conditions, been part of the repair.

4. Provenance: give media a nutrition label

If you work in a newsroom, a studio, or anywhere in the media supply chain, this is the most concrete item on the list. C2PA Content Credentials attach cryptographically signed assertions to a piece of media about its declared origin and edit history, a kind of nutrition label for content. The label can reveal when it has been tampered with; it doesn’t certify that the content itself is truthful. Adoption is past the pilot stage: TikTok joined the standard’s steering committee and says it has labelled over three billion pieces of AI-generated content, and OpenAI has made its image tools conform to the standard [9] [10].

The limitation is baked into the design: metadata can be stripped by a screenshot or a re-upload, which is why serious implementations pair it with durable watermarking and public verification tools [10]. And the evidence here is deployment and capability, not yet a measured reduction in misinformation belief. Provenance is a layer, not a lock. But it’s an adoptable, shipping layer, available today.

5. Ranking is a design decision, not a law of nature

This one lands closest to home for builders. The same audit that showed engagement ranking amplifies hostile content also tested an alternative: ranking by what users said they actually wanted to see. That alternative reduced the angry, partisan, out-group-hostile content in the feed [4]. It came with its own trade-off, a tendency to reinforce content people already agreed with, so it’s not a clean win. But it demonstrates the essential point: the feed is a choice. Every engineer who builds a recommender, a comment system, or a ranking function is deciding what human behaviour to reward.

Here is how those forces sit against each other.

A feedback loop of four stages where engagement ranking can amplify divisive content, with five proposed interventions, from media literacy to preference-based ranking, each pointing at the stage it could interrupt

The diagram sketches possible mechanisms and where each idea could interrupt them. It is not a demonstrated explanation of this or any election.

The honest limits

The effect sizes are modest. The inoculation gains varied a lot between countries, and researchers have flagged that some of them may partly reflect a shifted response criterion, people becoming more willing to label content as manipulative, rather than purely sharper discrimination between manipulative and non-manipulative content [7]. Media-literacy effects are real but strongest for knowledge, and slower to show up in behaviour [5]. And the AI tools cut both ways: the same systems that patiently walked people back from conspiracies can generate the next wave of convincing nonsense, and they’re confidently wrong often enough that I wouldn’t hand any of them the last word.

None of this is a hotfix. It’s a decade of patient work across education, product design, and standards. But “slow and partial” is not the same as “hopeless,” and deciding none of it is worth doing is how the loop keeps winning by default.

This isn’t only a German problem

I’ve written this from Germany, on the morning after a German election, but I don’t think the concern is uniquely German. Research and public debate in several democracies, the United States included, point to similar dynamics: feeds that reward the extreme, simple stories outrunning complicated ones, trust eroding faster than it rebuilds. Their strength and their political consequences differ by country, and I won’t pretend one election proves a world trend. But IT and media are global industries, and in my view that makes the information environment a shared piece of infrastructure whose upkeep is a shared professional responsibility, not a political hobby you take up on election night and drop by Tuesday.

I don’t have the answer

I said at the start that this post ends with a question, and I meant it. Five starting points, collected from the research in one worried day, are not a plan. At best they’re the opening position in a better conversation than the one the feeds keep offering us.

Two things I am sure of. First: doing nothing is also a decision. The machinery doesn’t pause while we look away; its defaults keep running, and its defaults are the loop. Second: the people reading this are unusually well placed. Between us, my network builds feeds and recommenders, trains and deploys the AI systems, produces journalism, runs media pipelines, and teaches the next generation of both. If the five ideas above are even half right, most of the levers are within reach of someone reading this sentence.

So here is the actual ask, and it’s genuine, not rhetorical. Are these the right five starting points? What’s missing? What sounds good in a study but won’t survive contact with reality? And what are you already doing in your corner that the rest of us should copy? Tell me, especially where you disagree.

That would be a conversation worth having, the morning after.

Sources

Stefan Christoph

About the Author

Stefan Christoph is a Principal Solutions Architect at AWS, focused on agentic AI, media & entertainment, and helping builders move from demo to production. He writes about AI architecture, developer productivity, and the future of software.

This is a personal blog. Opinions expressed here are my own and do not represent the views or positions of my employer.

Learn more →

❤️ Created with the support of AI (Kiro)