Moving Beyond Engagement: Optimizing Facebook's Algorithms for Human Values
Summary
Social media platforms like Facebook, Twitter, and YouTube often optimize for engagement metrics such as comments, likes, and shares, which can inadvertently promote toxic or low-quality content. For instance, Facebook's 2018 shift to "Meaningful Social Interactions" (MSI) led to algorithms boosting content that generated angry comments, ultimately weakening user ties. This issue mirrors problems in search engines, where clicks do not always signify user satisfaction. The article proposes replacing engagement-based metrics with human evaluation, similar to how Google and Bing use human raters for search relevance. By defining a core product principle, such as "Helping users feel closer to their friends and family," platforms can use trained human raters to score content on a 1-5 scale, directly measuring alignment with this principle. This approach allows for data-driven development, A/B testing, and ML model training based on human values rather than problematic engagement proxies.
Key takeaway
For Product Managers designing social media platforms, relying solely on engagement metrics like "Meaningful Social Interactions" can inadvertently promote harmful content and degrade user experience. You should instead define explicit, human-centric product principles and implement a human evaluation system to directly measure content alignment with these values. This shift enables data-driven development that genuinely improves user well-being and strengthens community ties, moving beyond the pitfalls of click-driven algorithms.
Key insights
Optimizing social media for human-rated values rather than engagement metrics can foster healthier online interactions.
Principles
- Engagement metrics often correlate with toxic content.
- Human evaluation can directly measure alignment with product principles.
- Expert raters offer consistency over user surveys.
Method
Define a core product principle, then use trained human raters to score content on a 1-5 scale based on its alignment with that principle. These scores can guide OKRs, A/B tests, and ML model training.
In practice
- Implement human evaluation for content ranking.
- Train ML models to optimize human eval scores.
- Use human ratings for A/B testing algorithm changes.
Topics
- Social Media Algorithms
- Human Evaluation
- Engagement Metrics
- Recommendation Systems
- Content Quality
Best for: Product Manager, AI Product Manager, Data Scientist, Machine Learning Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Surge AI Blog.