GPT-5 Launders Bias Instead of Removing It

🔥 What happened Researchers at Durham University analyzed 450,000 gender-directed GPT outputs from GPT-2 through GPT-5. Explicit misogynistic content vanishes — but the discrimination gets rerouted, not removed. 💡 Why it matters In GPT-5, 1,997 documents frame breast cancer as a men's rights debate — and three independent classifiers score it as non-toxic. Topic diversity in women-directed outputs drops 36% versus men, even as toxicity scores fall. ⚡ Our take Measuring safety by toxicity scores alone is theater. The industry urgently needs representation metrics — otherwise we'll keep celebrating models that just package bias more elegantly.