Phase 2b: LLM-Specific Responsible AI · 45 min · Python · openai · transformers

Toxicity & Harmful Content Filtering

Safety filtering is not censorship. It's the difference between a tool that helps and a tool that harms.

Hiring signal: Candidates who understand the false positive/false negative tradeoff in safety filtering — and can articulate why over-filtering harms marginalized communities while under-filtering exposes users to harm — demonstrate the nuanced thinking RAI roles demand.

What you will learn

Define a harm taxonomy: hate speech, harassment, self-harm, sexual content, violence
Use the OpenAI Moderation API and Google Perspective API for toxicity detection
Build a custom safety classifier by fine-tuning a small language model
Analyze false positive/negative tradeoffs and their disparate impact on marginalized communities
Address multi-language toxicity detection challenges

Introduction

Toxicity & Harmful Content Filtering

Why This Lesson Matters

When you deploy an LLM, you're deploying a system that can generate harmful content. Toxicity filtering is the safety layer between your model and your users. But it's not as simple as "block bad words" — the tradeoffs between safety and access are real, and they disproportionately affect marginalized communities.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Harm Taxonomy, Detection Tools, The False Positive Problem, Multi-Language Challenges, Key Takeaways — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.

Key terms in this lesson

Fine-Tuning