What's Causing Bard's Frequent Factual Hallucinations?

[Catalina]

Despite having internet access, Bard increasingly makes up information when asked for specific data - like debates, trades, career timelines, etc. Nearly half of Bard's recent "facts" in my tests are complete fabrications.

This excessive hallucination renders many responses useless misinformation. Has anyone else experienced Bard's imagination wildly distorting factual answers? I've had to stop using it and switch back to ChatGPT and Bing for reliable information. Hopefully Bard's fact hallucination issues can be curbed in future iterations.

 

[bomb]

LLM creativity and hallucinations often stem from high "temperature" prediction settings. Bard is known to be a cold model, likely to maintain factual reliability.

However, if Anthropic recently increased Bard's temperature to boost creativity before fully filtering out false information, it could explain the rise in hallucinations. Balancing creativity with truthfulness via temperature tuning seems tricky for Bard.

Have others noticed more imagination but less factual accuracy lately, possibly due to adjustment of its under-the-hood temperature/uncertainty levels? This temperature tweak hypothesis may account for Bard's new tendency toward creative factual errors.