Google's Jigsaw unveils Perspective, an AI-powered tool for publishers to weed out abuse from online comments, available as an API for free
A research team tied to Google unveiled a new tool on Thursday that could have a profound effect on how we talk to each other online.
Context & Ripple Effects
This launch is the delivery of what Jigsaw previewed months earlier as Conversation AI, its machine-learning project against online trolls — now shipped as Perspective, a free API any publisher can call instead of building moderation in-house. The bet is that abuse, not lack of interest, is what keeps comment sections closed.
The corpus shows the bet paying out along two paths: The New York Times adopted the tool within months and used it to open comments on more articles (NYT adoption), and by 2023 Jigsaw engineer Lucy Vasserman described how OpenAI, Anthropic, and other LLM makers use the same API to flag their models' toxic output — a customer base that did not exist when Perspective launched.
First-order effects
- Publishers get toxicity scoring at zero marginal cost, removing the staffing barrier that kept comment sections off most articles — the NYT's expanded commenting after adopting Jigsaw is the direct proof case.
Second-order effects
- Moderation spreads beyond publisher sites: Jigsaw's Tune extension applies the same classifier to Reddit, Twitter, Facebook, YouTube, and Disqus, putting one company's definition of 'toxic' in front of readers across platforms.
- Free distribution makes Perspective a default reference point for researchers and moderators, pressuring rival platforms and vendors to match its attributes or explain why they can't.
Third-order effects
- A tool built for human comments becomes safety infrastructure for machines: once OpenAI and Anthropic rely on Perspective to police LLM output, Jigsaw's classifiers sit inside the generative-AI supply chain, and its thresholds effectively become industry standards.
- If the pattern holds, moderation consolidates around a few shared ML APIs rather than per-site rules — concentrating norm-setting power in whoever operates the classifier, with little public accountability over what gets flagged.
The trend: Online moderation is consolidating around shared machine-learning APIs that begin as free publisher utilities and end up as de facto safety standards for both human speech and generative AI.