LAION has made a revised version of its widely used AI training dataset available after a safety review. The new release, Re-LAION-5B, is presented as a cleaned successor to LAION-5B and is said to contain no links to child sexual abuse material, or CSAM.
The move directly addresses issues identified in the original LAION-5B by the Stanford Internet Observatory in December 2023. It also gives organizations using LAION-5B derivatives a practical path to remove matching content from their own versions.
What Changed In Re-LAION-5B
Re-LAION-5B is a web-scale dataset of text-image pairs. LAION says it is the first dataset of this kind to be thoroughly cleaned of known links to suspected CSAM.
The updated dataset is available in two versions: Re-LAION-5B Research and Re-LAION-5B Research-Safe. Together, the release covers 5.5 billion text-image pairs.
The central change is the removal of 2,236 links after checks against lists provided by partners. That total includes the 1,008 links identified in the Stanford Internet Observatory report.
LAION notes an important qualification: many links known to child protection organizations are likely no longer active because removal efforts are ongoing across the public internet. For that reason, the 2,236 figure is described as an upper limit for links that may lead to CSAM.
Why The Metadata Matters
The release is not only a replacement dataset. LAION says third parties can use the metadata to clean up existing derivatives of LAION-5B.
That matters because LAION-5B has already been used and adapted by others. A cleaned central release helps, but it does not automatically change copies, derivatives, or downstream datasets that already exist.
The metadata gives those third parties a way to generate diffs and remove matching content from their own versions. In plain terms, it can help teams compare what they have against the cleaned release and identify content that should be removed.
This is an important operational step for AI dataset safety. Cleaning a dataset at web scale is not only about publishing a new file; it is also about making it possible for others to apply the same removals in their own systems.
The Safety Stakes For AI Training Data
The presence of CSAM in AI training datasets is inherently problematic. Datasets used for AI development can shape future systems, and the inclusion of harmful material creates serious ethical and safety concerns.
The issue also reaches beyond the dataset itself. Some trained systems are being used to generate CSAM, according to the source article. That makes dataset cleanup one part of a wider problem involving generative AI and child protection work.
The Internet Watch Foundation reported a sharp increase in AI-generated CSAM in fall 2023. The source article also notes that the volume of AI content can hinder investigations into real child abuse cases.
Another complication comes from AI-generated reports of possible CSAM that are automatically created by social media platforms. These reports can add to the workload facing investigators and child protection organizations.
A New Standard, With Earlier Criticism In View
LAION says Re-LAION-5B sets a new safety standard for cleaning image link datasets at web scale. That claim is tied to both the scope of the dataset and the use of partner-provided lists to remove known links to suspected CSAM.
The release also arrives after earlier criticism of LAION-5B. The dataset had previously faced criticism for containing patient images, according to the source article.
For AI researchers and organizations, the practical takeaway is clear: Re-LAION-5B is intended to replace or improve on LAION-5B in contexts where teams need a cleaned web-scale text-image dataset. Its value is not only in the revised dataset, but also in the cleanup path it offers for existing derivatives.
The broader lesson is that AI datasets are not static technical assets. They require review, correction, and mechanisms that let the wider ecosystem respond when serious problems are found.