What a toxicity classifier is actually trained to catch
Most toxicity models on the market, including the ones bolted onto generic comment widgets, were trained on social media: replies, posts, forum threads. Their training data skews toward slurs, harassment, threats and personal attacks between strangers. That is the pattern they learn to flag, and on that pattern they perform well.
A news comment or debate section is a different kind of text. Readers are not chatting with strangers. They are reacting to an article, often an opinion piece, with strong views on a policy, a public figure or a decision the newsroom itself made editorial choices about. The vocabulary is combative by design: reckless, dishonest, indefensible. None of that is harassment. All of it can trip a model trained to associate strong language with abuse.
Where opinion journalism breaks the model
A generic classifier scores text on surface signals: aggressive words, capitalization, exclamation marks, sentiment polarity. An opinion-heavy comment thread is full of exactly those signals, attached to legitimate argument rather than abuse. Toxicity detection built without press content in its training set has no way to tell the two apart, because it was never shown the difference.
The result is a moderation system tuned for the wrong distribution. It does the thing generic filters are known for on newsrooms: it over-flags sharp, well-argued disagreement, and it lets through the calm, politely worded false claim that reads nothing like "toxic" text but does the most damage to a debate.
The two failure modes, and why both are expensive
A false positive on a newsroom's comment section removes a reader's legitimate argument, the exact contribution a debate space exists to host. Enough of those and the loudest, safest opinions are the only ones left, which is the opposite of what a healthy structured debate is for.
A false negative is worse. A well-written, calmly stated falsehood or a coordinated talking point can sit in a thread for hours, read as reasonable, and shape what other readers believe, while the classifier that would have caught an insult stays silent. Neither failure is rare with a generic model on press content. Both come from the same root cause: the model was trained on the wrong kind of text.
What changes with press-tuned moderation
The fix is not a stricter threshold on the same model. It is training on the kind of text a newsroom actually receives: argumentative, opinionated, sourced in a comment thread rather than a social feed. Logora's AI moderation is trained specifically on press debate, and filters around 85 percent of toxic content before it reaches a human moderator, on a model built from roughly 45,000 human-labeled examples across more than 50 million reader contributions moderated since 2019.
The measurable effect is not fewer comments, it is better ones. At Milenio, the share of contributions approved for publication rose from about 60 percent before Logora to 80 to 85 percent after: a press-tuned model lets more legitimate argument through while still catching what a generic one would have missed.
If you are still comparing generic providers
Not every team is ready to commit to a press-tuned model on day one. Some are still testing generic providers, OpenAI, Azure, Google's Perspective API, Hive, against their own comment volume before deciding anything. For that stage, Logora also built and open-sourced OpenModeration, a unified API that plugs into more than ten moderation providers through one integration and one dashboard, hosted in the EU or self-deployed. It will not close the press-content blind spot described above on its own, that is a training-data problem, not an integration problem, but it makes running that comparison across providers a lot faster.
Why the DSA makes this a compliance question, not just a quality one
Under the Digital Services Act, every moderation removal needs a statement of reasons, and a model that over-flags legitimate opinion generates far more of those decisions to justify. A press-tuned model that gets the call right the first time is not only better for the reader, it is fewer disputed removals for the moderation team to document and defend. A public moderation charter that states the standard clearly makes the same distinction explicit for readers before they post.
Conclusion: match the model to the text
A toxicity classifier is only as good as what it was trained on. One trained on social media harassment will misread a newsroom's opinion pages, in both directions, no matter how the threshold is tuned. The fix is moderation trained on press debate from the start. Start with the AI moderation module or the broader comment system comparison to see how that changes the trade-off.