I'm wondering whether speculative, non-value-oriented, short-term market volatility might be diminished by AI rendering more accurate estimates of fair value.

I very much doubt it. An LLM might be very good at encapsulating the consensus of current opinions, or reproducing the logic and arithmetic steps of one or more people, but I'm at a loss to imagine how either of those would be an improvement beyond what already happens. The opinions would be no more or less biased, and the valuation methodologies no more or less meaningful.

We've had enormous amounts of automated trading for a few decades now, with algorithms specifically crafted for the task with actual insights on what the result is intended to do, rather than consensus machines, and it has hardly led to flat or rational valuation levels.

Besides, if you had a truly better estimate of the fair value of some asset, what would you do with that information? Most "money" would be uninterested in the result, being interested only in imminent price movements.

Jim