Perspective API is shutting down: how well do the replacements handle German?

On December 31, Google is switching off Perspective, its free toxicity filter that many comment sections and forums depend on. We compared how well cloud services, open guard models and the new English-German Tiger Guard are proven to handle German text. For almost all of them, there are no solid numbers for German, and the numbers that do exist are sobering.

Analysis Nov 19, 2026 9 min read Original analysis by our editors
Illustrative image, AI-generated

What this is about

Since 2017, the Perspective API from Google's subsidiary Jigsaw has been the go-to toxicity filter. You send it a comment and get back a number between 0 and 1 for how "toxic" it seems. The service was free. Comment systems like Coral (Vox Media), a Discourse plugin and hundreds of research projects rely on it. Now it's ending. On the project page, Jigsaw writes that the service will remain "active until December 31, 2026." Requests for more quota were only processed until February 2026. Asked about migration support, Google's answer is a curt "No."

So you need a replacement, and one that handles German. Our thesis: almost all alternatives are optimized for English. There are hardly any independent measurements for German, and the ones that exist show a clear gap. That goes for Tiger Guard too, the new specialist model for English and German.

What's going away, and why that's not all bad

For German in particular, Perspective was never a neutral yardstick. Researchers led by Gianluca Nogara showed at ICWSM 2025 that the API systematically rates German text as more toxic (paper). German tweets had a median toxicity score of 0.132. The same tweets translated into English scored just 0.012. At Google's recommended threshold of 0.7, the study found that four to ten times as many German tweets would be removed as their English translations, depending on the analysis. Because the system is a black box, the authors couldn't work out why.

With German reader comments, the problem showed up the other way around. In 2025, researchers at Munich University of Applied Sciences compared Perspective, OpenAI's moderation and GPT-4o on the HOCON34k dataset, which contains real comments from newspaper forums (arXiv 2501.01256). At a threshold of 0.5, Perspective found only 27 % of the hate speech. At 0.8, it flagged just 6 of 1,592 comments. According to the authors, the service struggled especially with dialect and links.

Our take: if you tuned your thresholds on Perspective, you calibrated them on a skewed instrument. Switching is also a chance to recalibrate your own rules.

The candidates at a glance

Option Where it runs German proven? Categories Cost
Tiger Guard / Small (308M / 141M) local, CPU works too yes, own figures binary: SAFE/UNSAFE free (Apache 2.0)
OpenAI omni-moderation cloud (US) vendor claim only, no figure 13, images too free
Mistral Moderation 2 cloud (EU provider) German among 11 languages (launched 2024) 11, incl. PII, jailbreak usage-based, price not listed
Azure AI Content Safety cloud, EU regions trained and tested on German hate, sexual, violence, self-harm 5,000 texts/month free
Google Cloud moderateText cloud "fully supported" 16 paid
Llama Guard 4 (12B) local, GPU 43 % recall across languages 14 (S1–S14) free (Llama license)
Shieldstral 1.0 (3B) local, 16 GB VRAM German among 12 languages custom rules in plain language free (Apache 2.0)
Qwen3Guard (0.6/4/8B) local 119 languages, no German figure 9, three-tier free (Apache 2.0)
ShieldGemma, Granite Guardian HAP local no, English only 4 and binary free

Tiger Guard comes from Casdo Labs (under the Patronus Protect brand) and came out this week. It consists of two compact encoder models, meaning models that only classify text and don't generate anything. They output a probability for "unsafe" and run with Transformers or ONNX Runtime on your own hardware. The quantized x86 version of the Small model is about 240 MB (Model Card, Small). The idea isn't entirely new: Hugging Face has long hosted German BERT classifiers trained on GermEval. One example is the ml6team model, which by its own card only reaches 39 % recall. Still, Tiger Guard is the first current guard model trained specifically for English and German that publishes its error rates broken down by language.

Among the cloud services, Azure stands out. It's the only one that states in writing that its classifiers for hate, sexual content, violence and self-harm were "trained and tested in German." A detail that matters for data protection: according to Microsoft's region list, text moderation is available in Sweden, France, Switzerland and West Europe, but not in the Germany West Central region. The older Azure Content Moderator only supports English, according to Microsoft's FAQ.

Mistral's Shieldstral (August 2026) takes a different approach. Instead of fixed categories, you write your house rules in plain language, and the model answers a yes/no question for each rule.

What the numbers on German really say

The gap between English and German runs through every measurement, including the vendors' own figures:

Model English German / multilingual Source
Tiger Guard, validation 80.3 % recall, 3.8 % FP 64.2 % recall, 4.4 % FP Model Card
Tiger Guard, PolyGuard holdout 83.3 % recall, 3.8 % FP 75.3 % recall, 5.5 % FP Model Card
Tiger Guard Small, PolyGuard 83.4 % / 5.0 % FP 75.3 % / 7.1 % FP Model Card
Llama Guard 4 69 % recall, 11 % FP 43 % recall, 3 % FP Meta
OpenAI Omni (RTP-LX, F1) 87.5 74.1 (multilingual) PolyGuard paper

Recall is the share of harmful texts a filter actually catches. FP, the false positive rate, tells you how many harmless texts it wrongly flags. Let's run the Tiger Guard numbers. On the validation set, 35.8 % of harmful German texts slip through, compared with 19.7 % in English. So the filter misses almost twice as much in German. At the same time, it flags more innocent texts. On the PolyGuard test, Tiger Guard Small flags 7.1 % of harmless German texts. With 1,000 harmless comments a day, that's about 71 false alarms.

On the plus side, it's open about this: the Model Card names its weaknesses explicitly. Tiger Guard misses obfuscation tricks like leetspeak and mild toxicity without direct insults. It also sometimes flags counter-speech that quotes insults in order to criticize them. The validation figures also come from data selected as part of training. Developer Dominik writes on Reddit that other prompts from the same data family as the PolyGuard holdout went into training. He advises you to "test edge cases against your own platform rules before wiring the score to an automatic action." The training corpus hasn't been released. You can set the threshold freely for these models: at 0.891, the false alarm rate drops to about 1 %, but recall on the benchmark falls to 63.6 %.

For German, OpenAI only says that the new Omni moderation performs better in German than the previous model did in English. There's no concrete German figure. For Qwen3Guard, Mistral Moderation 2 and Shieldstral, there are also no published German-specific results, only averages across several languages. According to Mistral, Shieldstral reaches an F1 of 84.6 on PolyGuard prompts. The F1 score combines hits and false alarms into a single number. Nobody has independently verified it yet.

In the interest of transparency: we didn't run our own test on GermEval 2018 for this article. If you want to test yourself, you'll find the data in the GermEval 2018 repository (tweets labeled OFFENSE/OTHER) and real newspaper comments in the HOCON34k corpus.

Data protection, the DSA and the shutdown trap

GDPR: Local models like Tiger Guard, Shieldstral or Qwen3Guard don't send any data out. That's the cleanest option. For OpenAI's moderation endpoint, the documentation states that inputs aren't used for training and aren't stored for abuse monitoring. Still, your comments go to a US company, which means you need to assess the third-country transfer. Mistral is based in France, and with Azure you can choose an EU region.

DSA: Under Article 15 of the Digital Services Act, platforms must state in their transparency reports whether they use automated moderation, along with "indicators of the accuracy" and the "possible rate of error." Micro and small enterprises are exempt from this reporting requirement. Our take: a model that discloses its error rates by language makes that report much easier to write than a black box.

The real danger is a silent failure. According to a post in Coral issue #4902, Coral waits 800 milliseconds for a response (PERSPECTIVE_TIMEOUT) and then publishes the comment unchecked. So starting January 1, nothing gets blocked. Everything goes through. There's no official response from the Coral developers there yet, only community suggestions for replacements. One of them is a self-hosted clone with a Perspective-compatible API, but it only handles English. Discourse has marked the Perspective plugin as deprecated and recommends its own AI workflows instead (discussion).

On the research side, a team led by David Hartmann analyzed 241 studies that use Perspective. They archived the scores for 5.9 million text snippets from 77 datasets so that old results stay comparable (arXiv 2604.25580).

Our verdict

There's no equivalent free replacement with proven strong German performance. Here's our take by use case:

  • Small forum, tight budget, privacy matters: Tiger Guard Small on the CPU. It's the only candidate with published German error rates. Set the threshold on the strict side and send hits to a review queue instead of deleting them automatically.
  • You want to enforce your own house rules and have a GPU: Shieldstral or Qwen3Guard. Both are more flexible, but unproven in German, so test them yourself.
  • Publisher with cloud approval: Azure Content Safety (explicitly tested in German, EU regions), or the free OpenAI moderation as a second opinion.
  • Skip: ShieldGemma and Granite Guardian HAP. They only handle English.

Whatever you choose: before December 31, export a sample with Perspective scores and your moderators' decisions, run the successor in parallel for a week, and switch the failure mode from "let through" to "hold back." Every candidate has higher error rates in German than in English, so you still need a human review step at the end.

Sources

#moderation#perspective api#tiger guard#german#gdpr#guard models

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.