
GPT-4 Turns Out to Be a Conspiracy Theory Enthusiast, Ranking Worse Than Its Predecessor
Ourtake
Safety claims and test results pointed in opposite directions, and the chatbot happily picked the test results.
What happened
Misinformation watchdog NewsGuard found that OpenAI's GPT-4 repeated false news narratives in 100 of 100 tests, while predecessor GPT-3.5 did so in 80. OpenAI had claimed GPT-4 was less likely to respond to disallowed content. Futurism and NewsGuard said OpenAI had not responded to requests for comment.
Sources
- newsguardtech.com · NewsGuard's Misinformation Monitor: GPT-4 produces more misinformation than predecessor (opens in a new tab)
- broadcastdialogue.com · ChatGPT-4 produces misinformation more persuasively, says NewsGuard report - Broadcast Dialogue (opens in a new tab)
- Futurism · Researchers Find GPT-4 Is Significantly Less Accurate Than GPT-3 (opens in a new tab)