It doesn't have to be this way, and individual fields can break away (to a greater or lesser extent). For instance:
In Natural Language Processing / Computational Linguistics, the professional society (Association for Computational Linguistics, ACL) was its own publisher, with no profit motive, and so authors for its conferences and journal never signed over copyright (merely granted permission to ACL to publish the work). For years, it was quite standard for nearly all of the authors to post PS or PDF versions of their papers on their own websites. Then ACL started accepting PDF instead of camera-ready, and just posted the PDFs themselves; and then they started scanning the back-catalogue.
The result of this is that the vast majority of all NLP/CL papers ever written (excluding only those published elsewhere, e.g. in AAAI, and a very few missing proceedings from fifty years ago) are available online, for free, in PDF, at http://aclweb.org/anthology-new/ .
In Natural Language Processing / Computational Linguistics, the professional society (Association for Computational Linguistics, ACL) was its own publisher, with no profit motive, and so authors for its conferences and journal never signed over copyright (merely granted permission to ACL to publish the work). For years, it was quite standard for nearly all of the authors to post PS or PDF versions of their papers on their own websites. Then ACL started accepting PDF instead of camera-ready, and just posted the PDFs themselves; and then they started scanning the back-catalogue.
The result of this is that the vast majority of all NLP/CL papers ever written (excluding only those published elsewhere, e.g. in AAAI, and a very few missing proceedings from fifty years ago) are available online, for free, in PDF, at http://aclweb.org/anthology-new/ .
This is how science should be.