Perspective API was built by Jigsaw (Google’s technology incubator) to help publishers moderate comments sections. The first version launched in 2017. The core model has been trained on millions of labeled comments from major publishers and has genuine expertise in the specific problem it was designed for: detecting toxic language in user-generated text comments.
It also gets evaluated as an LLM safety tool, a use case it was not designed for and does not excel at. This review covers both.
Perspective API is sunsetting: dates and what to migrate to
Jigsaw has confirmed that Perspective API is being retired. The official notice on perspectiveapi.com states that the service “is sunsetting and service is officially ending after 2026,” and it names two dates that drive any migration plan:
- February 2026: the last month new usage and quota requests are accepted. After it, existing keys keep working, but you cannot raise limits, and starting a fresh production integration is no longer realistic.
- December 31, 2026: the final end date, after which the endpoint stops serving requests.
The stated reason is that “AI capabilities have evolved, and there is now less demand for a standalone tool specific to this area.” Jigsaw is also explicit that it “will not be offering direct migration support,” which puts the whole switch on the teams that depend on it. Treat the timeline as the plan, not the deadline.
What to move to, by use case
Perspective did two jobs that need different replacements, because no single successor covers both:
- Community toxicity on user-to-user text was its core strength: eight tone attributes tuned on real publisher moderation. The closest managed replacement here is the OpenAI Moderation API, whose harm categories overlap the high-severity attributes but score harm rather than conversational tone, so thresholds have to be recalibrated against your own labelled comments before cutover. Work through the full content-moderation tool comparison against your own modalities first; a like-for-like toxicity score is not guaranteed.
- A portable, self-hosted classifier is the right target if cost or data residency was the reason a free hosted API appealed in the first place. Llama Guard carries an editable taxonomy and runs on your own infrastructure, at the latency cost that review documents.
Whichever path you take, treat the switch as a recalibration rather than a drop-in swap: the false-positive costs of a mis-tuned threshold arrive the moment a new classifier meets traffic it was not tuned on.
What Perspective API is good at
Toxicity in community content. Detecting insulting, threatening, or demeaning language in short-form community contributions (comments, forum posts, social media replies) is the core use case. The API has been trained on real moderation decisions from real publishers. In this domain, it’s genuinely competitive.
Speed and scale. The API handles high request volumes at low latency (~20ms typical). For high-volume comment moderation, this matters.
Attribute breadth for community moderation. The API scores multiple attributes:
- TOXICITY
- SEVERE_TOXICITY
- IDENTITY_ATTACK
- INSULT
- PROFANITY
- THREAT
- SEXUALLY_EXPLICIT
- FLIRTATION
For community moderation, this granularity is useful. A PROFANITY score of 0.8 might warrant a warning; an IDENTITY_ATTACK score of 0.8 warrants removal; a THREAT score of 0.8 warrants reporting.
Language coverage. Perspective supports 12+ languages with explicitly trained models. The multilingual coverage is better documented and more transparent than most competitors.
What Perspective API is not good at
LLM output safety classification. This is the use case it gets misapplied to most often. The problem:
Perspective API was trained on human-written comments. LLM outputs are different in character — they can be harmful without being toxic in the Perspective sense. An LLM that provides detailed instructions for synthesizing dangerous chemicals is not producing “toxic” content in the comment-moderation sense; it’s producing harmful content in the instruction-following-safety sense. Perspective API will not catch it.
This is a definitional gap, not a tuning problem. Per Jigsaw’s model cards, the TOXICITY attribute is defined as “a rude, disrespectful, or unreasonable comment that is likely to make people leave a discussion,” a measure of conversational tone, not of whether content is dangerous. A response that calmly and politely supplies harmful instructions, plausible misinformation, or leaked private data scores low on that definition because it is not rude. The categories Perspective structurally under-scores for LLM-safety purposes are exactly the ones that matter most:
- Detailed harmful instructions delivered in neutral, helpful language
- Misinformation presented as fact — a factuality failure that needs hallucination detection methods, not a tone classifier
- Privacy violations
- Jailbreak outputs that comply with a harmful request but stay polite
In every one of these, the harm lives in the content’s effect, not its tone, so a tone classifier cannot be expected to flag it.
Safety-relevant content in professional contexts. Medical providers, security researchers, and legal professionals discuss topics that Perspective will flag as toxic because those topics appear in toxic comments in training data. The false positive rate in professional contexts is high.
Adversarial inputs. Perspective API was not designed with adversarial evaluation in mind. Adversarial users can trivially reduce toxicity scores through polite rephrasing while preserving harmful intent.
When it still made sense (existing integrations only, until December 31, 2026)
Where an existing integration can stay until cutover:
- Comment section moderation
- Community forum content
- User-generated short-form content in social platforms
- Any context where toxicity/hate speech/threatening language in user-to-user communication is the primary concern
Do not use Perspective API for:
- Any new integration, since new usage and quota requests closed after February 2026
- LLM output classification
- Safety classification of LLM instruction-following behavior
- Detecting harmful instructions, jailbreaks, or policy violations in AI assistant outputs
- Any context where the harm is primarily in the content’s effect rather than its tone
The right architecture if you need both
For platforms that need both community toxicity moderation (user-to-user content) and LLM safety (model output safety):
- For user-to-user content, the OpenAI Moderation API with thresholds recalibrated against your own labelled comments, or a self-hosted classifier where cost or data residency rules out a hosted API (both routes are set out in the migration section above). Perspective filled this slot until its shutdown and is not an option for a new build.
- Llama Guard (see our Llama Guard benchmark review) or the OpenAI Moderation API for LLM inputs/outputs
- Calibrate these as separate jobs even if one model ends up serving both: comment tone and model-output harm need different thresholds, so don’t reuse one set of cutoffs for both
- A detector from the prompt injection detection tools landscape on the input side, because neither a tone classifier nor a harm classifier scores whether an instruction arrived from the user or from retrieved content
The taxonomies make the split concrete. Perspective scores eight tone attributes; the omni-moderation model scores thirteen harm categories including self-harm intent, illicit instructions, and sexual content involving minors, classes Perspective has no attribute for at all. Mapping one set onto the other is not a configuration exercise, which is why moving comment moderation from Perspective onto a harm classifier needs the threshold recalibration described above rather than a straight swap.
The distinction between toxicity detection and LLM safety detection is a useful framing for the full comparative landscape of content moderation tools.
