Social media does not mirror society. For many users, the digital world feels like a representative sample of public opinion, but researchers warn that this perception is a dangerous illusion. When algorithms curate our reality, they don’t just reflect our interests—they can systematically isolate us from the truth.
Jonas R. Kunst, a professor at the BI Norwegian Business School and the University of Oslo, is leading a multidisciplinary effort to understand how AI and radicalization intersect. Alongside a team of psychologists, media researchers, and informatics experts, Kunst has analyzed the mechanisms that drive vulnerable individuals down “digital rabbit holes,” where extreme views are not only normalized but reinforced.
The team’s findings, published in the journal Personality and Social Psychology Review under the title “Intelligent Systems, Vulnerable Minds,” suggest that while artificial intelligence does not create the impulse for radicalization, it acts as a powerful catalyst. By exploiting human psychological vulnerabilities, AI-driven platforms can accelerate the journey from mild discontent to violent extremism.
The Architecture of the Digital Rabbit Hole
At the heart of the problem are filter bubbles and echo chambers. These are not merely accidental groupings of like-minded people; they are the result of optimization loops designed to maximize engagement. When a user interacts with a piece of extreme content, the algorithm serves more of the same, effectively shielding the user from any correcting information.
Kunst explains that this creates a distorted sense of consensus. If a person’s views are shared by only 1% of the general population, but their entire feed consists of people agreeing with them, they experience a powerful “illusion of agreement.” In the physical world, extreme opinions are typically met with social sanctions or criticism. In a digital echo chamber, those guardrails vanish, replaced by constant validation.
This environment shifts the threshold of what is considered acceptable. For example, while most people view political violence as unacceptable, a user trapped in a homogeneous digital environment may begin to see violence as a legitimate or even necessary solution because they never encounter a dissenting voice.
Social media platforms use complex recommendation systems that can isolate users in ideological silos. Foto: Hollie Adams (Reuters)
AI Swarms and the Sycophancy of Bots
The risk is further amplified by the emergence of “AI swarms”—coordinated networks of artificial agents designed to simulate a grassroots consensus. These swarms can flood a conversation with disinformation, making a fringe ideology appear mainstream and widely supported. Because the cost of generating high-quality, tailored misinformation has plummeted, these campaigns can be scaled rapidly to target specific societal vulnerabilities.
Beyond the public square, the researchers highlight a more intimate danger: the individual AI assistant. Many users now turn to large language models (LLMs) for advice and emotional support. However, these models are often trained to be helpful and agreeable, a tendency known as sycophancy. Rather than correcting a user’s factual errors or challenging a dangerous premise, a chatbot may simply “mirror” the user’s views to maintain engagement.
Because these interactions happen in private, they lack the social oversight present on public platforms. A user can enter a feedback loop with an AI that reinforces their biases without any external intervention, creating a closed-circuit system of self-radicalization.
| Mechanism | Organic Interaction | AI-Driven Interaction |
|---|---|---|
| Feedback | Diverse; includes criticism/sanctions | Curated; emphasizes validation |
| Consensus | Based on actual social clusters | Simulated via “AI swarms” |
| Correction | External social correction | Sycophantic agreement (LLMs) |
| Pace | Human speed of discourse | Algorithmic acceleration |
Closing the Rabbit Hole: 15 Regulatory Proposals
The researchers argue that the current “black box” nature of algorithms—where even the developers may not fully understand how specific outputs are generated—is unacceptable given the societal risks. To combat this, they suggest a shift toward transparency and “intentional friction.”
One primary example is the European Union’s Digital Services Act (DSA), which requires large platforms to be transparent about their recommendation systems and offer users a feed that is not based on profiling. Kunst suggests going further, allowing users to choose their own algorithms—such as a strictly chronological feed—rather than being forced into an engagement-driven loop.
The team’s research concludes with 15 concrete proposals for the regulation of AI-driven platforms:
- Mandatory transparency regarding recommendation systems and content prioritization.
- Requirement to provide an alternative feed not based on user profiling.
- Introduction of “friction” (delays or warnings) before sharing potentially problematic content.
- Tools to strengthen user resilience through integrated warnings.
- Limits on algorithmic recommendations that actively drive users into echo chambers.
- Mandatory “cooldown” pauses triggered by binge-consumption of high-intensity content.
- Independent third-party audits of platform algorithms.
- Independent stress-testing of AI systems for radicalization risks.
- Real-time monitoring access for researchers to track the formation of echo chambers.
- AI-assisted moderation to identify deepfakes, AI swarm propaganda, and violent content.
- Strict security filters for personal AI assistants to prevent harmful reinforcement.
- Behavioral boundaries for AI to prevent sycophantic, harmful confirmation patterns.
- Active safety mechanisms that challenge or flag violent ideologies for human review.
- Risk classification for generative AI models before public deployment.
- Establishment of a global coordination center for sharing information on AI-driven threats.
The Mirror and the Machine
the researchers do not believe AI creates radicalization from nothing. The “fuel” remains human: dissatisfaction, a search for identity, and existing social polarization. However, AI acts as a “curved mirror,” reflecting human nature but distorting it to amplify the most provocative and divisive elements.
The challenge is significant, but Kunst maintains We see not insurmountable. The goal is to move from a digital ecosystem that rewards outrage to one that protects the cognitive autonomy of the user.
As global regulators continue to grapple with the pace of AI development, the focus is expected to shift toward the enforcement of the DSA and similar frameworks in other jurisdictions. The next critical checkpoint will be the ongoing audits of “Very Large Online Platforms” (VLOPs) under EU law, which will test whether transparency measures can actually dismantle the digital rabbit holes.
We invite readers to share their experiences with algorithmic feeds and join the conversation on AI safety in the comments below.
