Data Skeptic

AutoLike

35 min

This episode of Data Skeptic explores the AutoLike framework, a tool that uses reinforcement learning to audit opaque social media recommendation systems like TikTok's 'For You' page. Hugh Lee explains how AutoLike can automatically drive these black-box systems to serve specific content topics, including potentially harmful ones, thereby providing regulators and platform designers with a scalable method to measure and characterize algorithmic content curation. The discussion highlights the challenges of auditing closed platforms and the need for better data access and transparency.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

AutoLike formulates the audit of recommendation systems as a reinforcement learning problem, where an agent learns through trial and error which actions (like, skip, share) most efficiently drive the system toward a desired content topic.
2

Using AutoLike, researchers were able to drive TikTok to serve significantly more content on a chosen topic (e.g., pets, sports) within just 100 interactions, compared to a control strategy of simply skipping all videos.
3

The tool can also manipulate sentiment dimensions; for example, selecting 'sad cats' resulted in the overall content stream becoming sadder, not just cat-related, suggesting cross-topic emotional contagion in recommendations.
4

A key challenge is scaling real-time audits, as each experiment requires thousands of video interactions and classifications; future work involves creating simulated TikTok environments to train agents faster.
5

AutoLike is designed for auditors like regulators (e.g., FTC) to gather evidence on how easily platforms serve problematic content, and it is agnostic to platform or topic, only requiring the app's interface.

💬 Notable Quotes

"Social media is a black box. It's not open source, it's definitely not open data. So how can you study it from the outside? Well, AutoLike is one great example of how to do that." — Hugh Lee, Senior Technologist at FTC"
"The agent will keep learning that over and over until it will eventually learn the optimal action to take per state to get to what you want." — Hugh Lee"
"If the platforms are able to provide special environments or APIs that researchers can hook into, that would be very helpful." — Hugh Lee"