Google's Jigsaw expands Perspective API, a set of ML tools that help moderators and researchers identify toxicity, with seven attributes, including nuance
Billy Perrigo / TIME : X: @mer__edith , @prestonjbyrne , @jigsaw , @billyperrigo , @jigsaw , and @jigsaw . LinkedIn: Reza Ghazinouri See also Mediagazer X: Meredith Whittaker / @mer__edith : TLDR: to ‘solve’ trash content on trash surveillance platforms, BigTech AI will flag whether content is ‘good’ or not. Can we please stop falling for this? This isn't ‘solving’ the problem. It's simply consolidating control over platforms w/o attending to who will wield it. Preston Byrne / @prestonjbyrne : “The last set of AI moderation tools just suppressed viewpoints I disagree with. The next set of AI moderation tools will promote views I agree with too” Saved y'all the click @jigsaw : What if AI could be leveraged so that online discourse brings people together? Seven years ago Jigsaw launched Perspective API, a suite of machine learning tools that help moderators and researchers identify toxicity in online speech. Billy Perrigo / @billyperrigo : We often hear about how LLMs are going to degrade the internet. But what if they could heal its divides? My exclusive story about a set of new AI tools from Google Jigsaw, released today: https://time.com/... @jigsaw : These new attributes are intended to be used in conjunction with the original set of Perspective tools. Moderators could use these attributes to rank top comments according to nuance and respect. @jigsaw : We're excited to announce the addition of seven new experimental bridging attributes, empowering researchers and moderators to identify and rank comments based on attributes such as nuance, reasoning, personal stories, compassion, and more. LinkedIn: Reza Ghazinouri : [Jigsaw] revealed a new set of AI tools, or classifiers, that can score posts based on the likelihood that they contain good content: Is a post nuanced? … See also Mediagazer
Context & Ripple Effects
Perspective originated as a free API for publishers to flag abusive comments, then broadened into a toolset used beyond publisher moderation. Its use to flag toxic LLM output was documented in Jigsaw's work with major AI labs, making this an expansion of the system's role in both platform and model-safety workflows.
The new attributes move the product from identifying abuse toward assessing constructive qualities such as reasoning, personal experience, and compassion. That broadens the practical and governance significance of automated speech classification.
First-order effects
- Moderators and researchers using Perspective gain experimental signals for ranking or reviewing comments on qualities beyond toxicity; the initial rollout is indicated as experimental in India.
- Jigsaw must validate whether these more subjective attributes perform reliably enough for moderation and research decisions, rather than simply detecting clearly abusive language.
Second-order effects
- Platforms and AI developers already using Perspective for toxic-language screening can test more granular intervention rules, such as elevating constructive contributions rather than only filtering harmful ones.
- The added categories make configuration choices more consequential: adopters must decide how strongly automated assessments of nuance or compassion should affect visibility, review queues, or enforcement.
Third-order effects
- If adopted broadly, AI moderation could shift from a narrow safety function toward automated quality ranking of public speech—raising the importance of transparent criteria, appeal paths, and human oversight.
- The pattern extends Jigsaw's evolution from publisher comment-abuse detection to infrastructure that can shape how platforms and AI systems classify acceptable discourse; whether it improves deliberation or concentrates editorial power will depend on deployment controls.
The trend: Content-moderation AI is evolving from detecting prohibited speech toward scoring the perceived quality and social value of speech.