If anything, iterative prompting introduces more bias.
Correction only matters only in the hypothetical where the LLM has already produced an output that mischaracterizes the original content. What gives you confidence that correction n+1 is different? Given that susceptibility to error is assumed, without reference to the original, the best assurance you can have is that subsequent outputs no longer contradict. But this gives no certainty that it has converged on an accurate characterization.
Correction only matters only in the hypothetical where the LLM has already produced an output that mischaracterizes the original content. What gives you confidence that correction n+1 is different? Given that susceptibility to error is assumed, without reference to the original, the best assurance you can have is that subsequent outputs no longer contradict. But this gives no certainty that it has converged on an accurate characterization.