“It was a mistake.” A million posts from Bluesky were stolen to create training data for AI

The team of the Bluesky platform, which has been rapidly gaining popularity in recent months, promises not to use user data to train AI. However, no one is stopping someone else from collecting the data.

This week, one million public posts from Bluesky, along with user identification information, were scanned and then uploaded to Hugging Face. The dataset was created by machine learning expert Daniel van Strien and is intended for use in language modeling and natural language processing, as well as general analysis of social media trends, content moderation, and post patterns. It contained decentralized user identifiers (DIDs) and even had a feature to search for content from specific users, 404Media reported.

According to the dataset description, the posts were collected from Bluesky Social’s Firehose API. Bluesky users did not consent to such data use, but the platform does not prohibit such manipulation.

Shortly after this dataset became public, it was removed from Hugging Face.

«I have removed Bluesky data from the repository. While I wanted to support the development of tools for the platform, I recognize that this approach violates the principles of transparency and consent for data collection. I apologize for this mistake,” van Strien wrote in a post on Bluesky.

This could be a wake-up call for users of the platform, which has been rapidly gaining popularity in recent weeks. Although the platform’s owners have promised not to use user data to train AI, they have yet to create tools to force third-party companies to do so without users’ consent.