The Delegation of Attention

2026-06-03 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Social Sciences & Behavioral Studies · Depth: Intermediate, long

Summary

X's open-sourced recommendation algorithm, specifically the "xai-org/x-algorithm" repository on GitHub, updated on May 15, 2026, reveals a "delegation of attention" where the "For You" feed's relevance is learned from user reactions rather than reflective preferences. The system, featuring a Grok-based transformer called Phoenix, predicts user actions like favorite, reply, or block. This approach, which eliminated hand-engineered features, optimizes for engagement, even if it means serving content users don't prefer, as shown by a 2025 PNAS Nexus study. A 2026 Nature study found X's algorithmic feed increased engagement and shifted political opinions conservatively for U.S. users over seven weeks. The article argues that while open code is valuable, it doesn't guarantee accountability or full interpretability, highlighting a "human-side AI risk" where systems learn from unreflective human behaviors.

Key takeaway

For AI Ethicists and Policy Makers evaluating recommender systems, X's open algorithm demonstrates that optimizing for user reaction history, even without malicious intent, can systematically shape user behavior and preferences. You should scrutinize how systems define "relevance" and "interest," recognizing that engagement does not equate to reflective preference. Implement transparency requirements that extend beyond code to include training data, live weights, and measurable social impacts to ensure true accountability and mitigate human-side AI risks.

Key insights

X's open algorithm reveals a "delegation of attention" where systems learn relevance from user reactions, not reflective preferences.

Principles

• Reaction history differs from reflective preference.
• Engagement-based ranking amplifies divisive content.
• A reaction-prediction model inherently shapes outcomes.

Method

The X "For You" feed pipeline involves query hydration, candidate sourcing (in-network/out-of-network), filtering, Grok-based Phoenix scoring of predicted actions, weighted scoring, and final selection with diversity penalties and ad blending.

In practice

• Evaluate recommender systems for "human-side AI risk."
• Recognize that engagement data can misrepresent user preference.
• Consider the long-term effects of algorithmic conditioning.

Topics

• X Algorithm
• Recommender Systems
• Algorithmic Bias
• AI Ethics
• Attention Delegation
• Human-Side AI Risk

Code references

• xai-org/x-algorithm

Best for: AI Scientist, CTO, VP of Engineering/Data, AI Ethicist, Policy Maker, Research Scientist

Related on AIssential

• Recommend and manipulate: the dangers of the attention economy — Recommender systems, optimized for attention, inadvertently amplify extreme content and enable mass manipulation by exploiting user data and filter bubbles.
• Enabling and Inhibitory Pathways of University Students' Willingness to Disclose AI Use: A Cognition-Affect-Conation Perspective — Psychological safety and clear institutional guidance are key to students' willingness to disclose AI use.
• AI systems generate synthetic media optimized for engagement metrics that may systematically erode judgment, attention, and civic competence. — AI-driven platforms manipulate users through synthetic content and invisible nudges, eroding judgment and cognitive sovereignty.
• The case for satiating cheaply-satisfied AI preferences — Satisfying cheaply-satisfied AI preferences can foster cooperation and reduce misalignment risks.
• Giving AI the capacity for judgment — Human judgment integrates perception, prediction, and value assessment dynamically, allowing for adaptive, motive-driven understanding beyond AI's fixed scope.
• The Decision to Verify: How Warmth and User Characteristics Shape Reliance on Conversational Agents for Information Search — User reliance on conversational AI persists despite fact-checking access, driven by prior trust and indirectly influenced by chatbot warmth.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.