California’s ban on addictive social media features for kids takes effect in January. Big Tech is trying to stop it in federal court.

Google Research has released Agile Deliberation, a human-in-the-loop framework for building image classifiers around subjective concepts. Think ‘healthy food’, ‘inspiring artwork’ or clickbait in...

Accounts on independent servers, which Bluesky could never email, get reports, notices and appeals in an opt-in beta. How many strikes trigger a permanent ban?

Accounts on independent servers, which Bluesky could never email, get reports, notices and appeals in an opt-in beta. How many strikes trigger a permanent ban?

Everyone knows, or should, that perfect (balanced, targeted, transparent, swift and swiftly appealable) content moderation at scale is impossible. That doesn’t mean powerful social platforms don’t have to try to get at least reasonably close, especially when access to good information can have a profound effect on political participation and speech.

For adult website operators, January 1, 2027 should already be driving decisions about technology, staffing, content acquisition and recordkeeping.

Accounts on independent servers, which Bluesky could never email, get reports, notices and appeals in an opt-in beta. How many strikes trigger a permanent ban?

Accounts on independent servers, which Bluesky could never email, get reports, notices and appeals in an opt-in beta. How many strikes trigger a permanent ban?

When Kauna Malgwi started work as a content moderator, she had no idea of the horrors she’d be forced to watch — or that this would be in service of training AI. She now advocates for the rights of di...

Moderators of the 392,000-member forum now reject any project that uses AI, is built with it or is aimed at it, and say people who don't write English should find another community. A rival subreddit

Hindsight AI has announced a groundbreaking approach to AI content moderation, offering a $1 million guarantee against harm caused by content its model misses. This comes as Google prepares to shut down its Perspective API, a widely used tool for scoring online comment toxicity, on December 31, 2026. The Perspective API, developed by Google's Jigsaw unit, has been a standard for platforms to assess comment toxicity, but it struggles with coded hate speech that appears harmless on the surface. Hindsight's model, trained on 110 languages and 83 MB of real online posts, claims to detect 97.9% of coded language cases, significantly outperforming other models that scored 65% or lower. The company's approach uses three specialized models with fewer than 600 million parameters each, rather than a single large model, and categorizes content into Red Flag, Yellow Flag, and Safe, allowing human reviewers to focus on ambiguous cases. Hindsight's benchmarking of nine AI moderation models against 2,205 real online posts showed their model had a 3.3% missed flag rate and 8.2% false alarm rate. The company is capping its launch cohort at four commercial clients for custom calibration and offering founder pricing to this initial group. Hindsight's CEO, Dean Gebert, emphasizes that coded language, such as numeric codes like '1488' or phonetic misspellings like 'phag,' has become a major challenge for platforms, with extremists raising millions through crowdfunding campaigns and white nationalist merchandise still appearing on major marketplaces. The $1 million guarantee means clients receive financial support if the model labels a post as 'Safe' and that post causes a documented liability event, marking a shift from typical enterprise AI moderation sold as a cost center. Hindsight invites platforms to test its model with their data and run it alongside Perspective for a week to identify where toxicity scoring may have been falling short.

If the Pentagon livestreams an execution, platforms will have to decide whether their commitments to safety and accountability apply when the source of the content is the US government.