I'm concerned that the training process, which involves back-propagation to adjust synapse weights, may be an unpleasant experience for the ANN.

This assumption is not based on facts. It’s pretty much like saying that matrix multiplication can have feelings, or that heat stressed silicon is equivalent to pain.

But if this is actually a concern, RNNs have been widespread since the late 90s. Any advanced search engine, translation engine, or weather forecast model, make use of these.

Regardless, it's all a moot point because we have lots of other reasons not to use LLMs.

This may be true, but it’s absolutely outside of the scope of your original point. You dragged the conversation around claiming to be concerned about how models are “treated”, wrapping speculation with philosophical arguments that cannot be applied here, since none of your “what ifs” are remotely based on scientific consensus.