Misinformation Mitigation on YouTube

3/13/2025

This post details YouTube's strategy for 'Raising authoritative content' and 'Reducing borderline content and harmful misinformation' as part of its 'Four Rs of Responsibility'. It describes elevating authoritative news sources and experts in search and recommendations, providing breaking news previews and informational panels to contextualize content, and using machine learning models trained on external evaluator input to reduce the spread of borderline and misinforming content. The post highlights a 70% average drop in watch time of this content from non-subscribed recommendations in the U.S. and mentions expansion to international markets.

4/24/2024

This post details YouTube's strategy for tackling misinformation, emphasizing a dual approach of 'increasing the good and decreasing the bad.' It highlights the limitations of solely relying on content removal and instead focuses on elevating trusted sources and reducing the spread of harmful content. The post discusses the reliance on expert consensus for defining misinformation (e.g., for COVID-19) and the challenges of subjective misinformation. It outlines policy evolution, such as removing election-related content after certification, and the importance of balancing free speech with reach. The author also addresses the financial implications, stating that responsible content management is good for business.

2/17/2022

This post details YouTube's evolving strategies for combating misinformation, focusing on three key challenges: catching new viral misinformation by leveraging targeted classifiers, keywords in additional languages, and regional analysts; addressing the cross-platform spread of borderline content through potential share button disabling or interstitials; and ramping up global efforts by investing in regional expertise, partnerships, and localized model updates. It also discusses the challenge of data voids and the careful consideration of surfacing labels that could unintentionally amplify topics.

4/6/2021

Introduces the Violative View Rate (VVR) as a new metric to measure the percentage of views on YouTube that come from violative content. Details the methodology for calculating VVR through content reviewer sampling and explains its importance in understanding the impact of harmful content on viewers. Highlights that VVR has fallen by over 70% since 2017 due to investments in machine learning, with the current rate at 0.16-0.18%. Discusses how VVR will be updated quarterly in the Community Guidelines Enforcement Report and its role in holding teams accountable for progress in protecting users from harmful content.

8/25/2020

This post details the engineering and operational decisions made during COVID-19 to adapt YouTube's policy enforcement. Faced with reduced human review capacity, YouTube prioritized safety by increasing reliance on automated systems to remove potentially harmful content, even at the cost of accuracy and leading to over-enforcement in sensitive categories. The post also outlines technical and policy adjustments to mitigate disruption for creators, such as waiving strikes for non-human-reviewed removals and enhancing appeal processes.

8/4/2020

This post details the 'Remove' principle of YouTube's 'Four Rs of Responsibility' framework. It outlines the process of developing policies for harmful content, emphasizing the balance between free expression and community safety. It describes the use of machine learning and hashing for automated content flagging, noting that over 87% of removed videos in Q2 2019 were first flagged by automated systems. The post also highlights efforts to remove content before it's widely viewed, including the launch of the Intelligence Desk and the significant investment in human resources (over 10,000 people across Google) dedicated to detecting, reviewing, and removing policy-violating content.