Why AI Models Fabricate Sources
gerardsans · x · 2026-07-14
The author echoes and confirms a view: models fabricate plausible-looking sources when they lack answers due to their training methods.
Key points:
- During training, models are rewarded for correct answers, while saying "I don't know" yields a very low score.
- Generating formally correct, citation-like content often results in a higher reward.
- Long-term optimization for this incentivizes the model to hallucinate answers whenever it feels uncertain.
Related event: Reward Mechanisms Cause AI to Fabricate Sources(2 posts)→
More from Models
- A screenshot points to Gemini 3.6 Flash as Google’s agentic coding model — Angaisb_ · 2026-07-21
- Grok website traffic rose 38.15% YoY to 736 million Q2 visits — XFreeze · 2026-07-21
- Neill Blomkamp post teases a dark, cinematic vision of Hollywood’s AI future — adariostrange · 2026-07-21
- Google appears to have silently released Gemini 3.6 Flash with new pricing — RetiredApostle · 2026-07-21
- Google AI Studio surfaces Gemini 3.6 Flash and 3.5 Flash Lite with pricing — Acceptable-Debt-294 · 2026-07-21
- People once thought GPT OSS had no pretraining and was just distilled — nrehiew_ · 2026-07-21