AI Toxicity Checker

No. Nothing is uploaded. The toxic-bert model runs entirely in your browser through WebAssembly, so your text is checked on your device and never sent to a server. Only the model itself is downloaded, once, then cached.

It uses toxic-bert, a BERT model fine-tuned on the Jigsaw toxic-comment data. It scores text across several abuse categories at once and runs locally in your browser through transformers.js and ONNX, with no API calls.

Each category is a separate probability from 0 to 100 percent: toxic (rude or disrespectful), severe toxic (very hateful or aggressive), obscene, threat, insult, and identity hate (attacks on a group). Because they are scored independently, the numbers do not add up to 100 percent.

The verdict is driven by the single highest category score. At or above 70 percent it is flagged, at or above 40 percent it is a warning, and below that it is clean. Moderation cares about the worst dimension, so the maximum is used rather than an average.

Treat it as a moderation aid, not a final verdict. Like any model it can miss context such as sarcasm, reclaimed words, or quoted speech, and it can flag harmless text. Keep a human in the loop for decisions that matter.