When recommendation algorithms become systemic risk under EU law

What this risk is

AI recommendation systems that optimize for engagement metrics systematically amplify content that provokes emotional responses — outrage, fear, anxiety, tribal identity — at the expense of accuracy, balance, and epistemic health. The resulting filter bubbles, radicalization pathways, and polarization are not incidental byproducts; they are the predictable output of optimizing for engagement in a world where emotional content drives engagement.

How it occurs · Mechanisms

Recommendation systems face a misaligned objective:

• What they optimize: Engagement (clicks, watch time, shares, reactions)
• What correlates with engagement: Emotional arousal, in-group/out-group content, novelty, outrage
• What doesn’t correlate with engagement: Accuracy, nuance, balanced perspective, epistemic quality

The system is not “trying” to polarize users. It is doing exactly what it was designed to do. The problem is the design.

Mitigations · Governance

Objective Redesign

• Multi-objective optimization — Incorporate measures of information quality, source credibility, and user wellbeing into recommendation objectives
• Engagement quality vs. engagement quantity — Distinguish between content users find valuable and content they find compelling

Architectural Changes

• Recommendation diversity requirements — Ensure recommendations expose users to diverse perspectives
• Friction for extreme content — Reduce amplification of content flagged as extreme; require user action rather than passive recommendation
• Explanation transparency — Show users why content was recommended; enable meaningful control

Governance

• Third-party algorithmic audits — Independent audit of recommendation system impacts on information diversity and political polarization
• Transparency reports — Public reporting on recommendation system behavior and impacts