If you benchmark is to be better than a human familiar with the task at hand, then you will be disappointed.

If you are not expecting them to be as good as people then their failures seem wholly unremarkable. The specific nature of tbe failures can be quite interesting in fact. They can expose deficiencies in the architecrure, training data, or presumptiona of those controlling the models.

In my experience very few people are claiming the capabulities that would be required for me to expect them to produce better than they currenrly are.

I'm not bothered by a fish's inability to quote Shakespeare.