Jason JunJason Jun

Detecting toxicity in text with TensorFlow.js

2 March 2023

During my work on a recent side project, Rephraser AI, I discovered that the GPT-3 API had a safety setting for content toxicity. If user input contained toxic content, sometimes the API would ignore the prompt altogether and return warning messages instead of rephrases. To communicate this to the users, I developed this useToxicity React hook for detecting toxic messages.

I will walk through each step of the process, including loading the toxicity model, creating the hook, and using it in a React component.

Set up the project

We will use create-react-app for this demo.

npx create-react-app use-toxicity-demo

Next, navigate to the project directory and install the required dependencies:

cd use-toxicity-demo
npm install @tensorflow/tfjs @tensorflow-models/toxicity

We will use @tensorflow/tfjs for loading the model and @tensorflow-models/toxicity for detecting toxicity.

Create the hook

Now we can create the useToxicity hook. This hook takes a string as an argument and returns an object with two boolean values: loading and isToxic. The loading value indicates whether the toxicity detection process is still running, while the isToxic value indicates whether the message is toxic or not.

import { useEffect, useState } from 'react'
import '@tensorflow/tfjs'
import * as toxicity from '@tensorflow-models/toxicity'
const useToxicity = text => {
  const [loading, setLoading] = useState(true)
  const [isToxic, setIsToxic] = useState(false)
  const threshold = 0.9
  useEffect(() => {
    const checkToxicity = async () => {
      const model = await toxicity.load(threshold, [])
      const predictions = await model.classify(text)
      const toxicPredictions = predictions.filter(p => p.results[0].match)
      setIsToxic(toxicPredictions.length > 0)
  }, [text])
  return { loading, isToxic }
export default useToxicity

To specify which categories of toxicity you want to classify your text against, you can pass a list of labels as the second argument for toxicity.load(). However, if you only need a simple binary classification of whether a message is toxic or not, you can omit the labels and pass an empty array, as done in this example.

Additionally, the model.classify() method returns an array of predictions, each of which is an object with the following structure:

  label: 'identity_attack',
  results: [
      match: true,
      probabilities: [0.97, 0.01, 0.02, 0.01, 0.01, 0.01, 0.01, 0.01],

This information can be used for more specific toxicity detection if needed.

Example usage of the hook

The code below is a simple example of using the useToxicity() hook in a React component.

import useToxicity from './useToxicity'
const TextInput = () => {
  const [inputText, setInputText] = useState('')
  const [textToCheck, setTextToCheck] = useState('')
  const { loading, isToxic } = useToxicity(textToCheck)
  const handleButtonClick = e => {
  return (
      <label htmlFor="text-input">Enter your message:</label>
        onChange={e => setInputText(e.target.value)}
      <button onClick={handleButtonClick}>Check toxicity</button>
      {loading && <div>Loading...</div>}
      {!loading && !isToxic && <div>Your message is clean!</div>}
      {!loading && isToxic && <div>Warning: Your message is toxic!</div>}
export default TextInput

To avoid running toxicity detection on every keystroke, a separate textToCheck state variable is used for toxicity checking. This is because the toxicity detection process can take some time.