Citation

Reducing Hallucinations in AI-Generated Wiki Content - Footnote 25

partial85% confidence

1 evidence check

Last checked: 4/3/2026

The claim mentions the 2025 AI Index Report, but the source does not mention this report or RLAIF (Reinforcement Learning from AI Feedback) or DPO (Direct Preference Optimization). The source does not mention GPT-4 reducing factual errors by 40% after RLHF training, but does mention OpenAI's GPT-4 seeing a 40% reduction in factual errors after undergoing RLHF training.

Evidence — 1 source, 1 check

www.voiceflow.com/blog/prevent-llm-hallucinations(1 check)

partial85%Haiku 4.5 · 4/3/2026

Found: **Reinforcement Learning from Human Feedback (RLHF)** trains models to prefer outputs that human reviewers label as correct. Research shows RLHF can reduce factual errors by 40% (GPT-4) and harmful ha…

Note: The claim mentions the 2025 AI Index Report, but the source does not mention this report or RLAIF (Reinforcement Learning from AI Feedback) or DPO (Direct Preference Optimization). The source does not mention GPT-4 reducing factual errors by 40% after RLHF training, but does mention OpenAI's GPT-4 seeing a 40% reduction in factual errors after undergoing RLHF training.

Debug info

Record type: citation

Record ID: page:reducing-hallucinations:fn25