A model that possesses the entire collective knowledge of our civilization is useless if it can't directly quote its sources.
If we enforce the behavior of always paraphrasing and synthesizing the information before returning it even in cases where the exact quote is asked for then that is a failure.
The correct solution is to make the model capable of
1) quoting directly
2) identifying and then indicating when its output is a direct quote
3) citing the source of the identified quote
while still retaining the ability to paraphrase and synthesize those sources when appropriate. These requirements are what humans are held to and should also be applied to commercial AI models. Models that are not intended for commercial use should be exempt and at this point there isn't really a way to hold them to such requirements anyways.
If we enforce the behavior of always paraphrasing and synthesizing the information before returning it even in cases where the exact quote is asked for then that is a failure.
The correct solution is to make the model capable of
1) quoting directly
2) identifying and then indicating when its output is a direct quote
3) citing the source of the identified quote
while still retaining the ability to paraphrase and synthesize those sources when appropriate. These requirements are what humans are held to and should also be applied to commercial AI models. Models that are not intended for commercial use should be exempt and at this point there isn't really a way to hold them to such requirements anyways.