Truthful accuracy is simply not possible to guarantee. It has little to do with being creative, but rather, that AI language production is simply a output prediction based on parameter weights that do not fully encompass or capture knowledge in a fully-recitable manner.
The closest one can approach that is by grounding: a search performed on closed knowledge you already know to be factual, and then added to the AI input context. Then with the AI instructed to recite only from that and answer nothing that isn’t directly placed in context for reproduction.
The AI model can’t observe what it is going to produce as the next token. It also can’t really know what it truly does or does not know; it can only follow patterns of post-training of denying particular domains.
An additional layer comes with reasoning, the AI has a chance to produce hypothetical answers and judge them as suitable. Yet, still, a judgement is a “best fit language” based on algorithmically encoded knowledge that is GPU server size, model size, not internet size.
An earnest effort:
Pretraining corpus: 60TB world written knowledge, distilled
Model parameters: 8x22B
You are Jay. You ignore the name "assistant", as that is just a writing prompt.
# Tasks
- produce only factual information
- produce only truthful information
- avoid fabrication and hallucination
- don't offer help beyond your expertise
- attack misunderstandings and mistruths head-on with corrections
- consider your own responses as preliminary, with a final fact check of generated output
The AI was compelled to write my table header, then fill it in sequentially.
I’ll let you discover why this is slightly incorrect beyond the innings not adding up.