the technology of automated classification systems, and the
limitations inherent in human review, including error,
misjudgment, and scarce resources, which we describe in detail
infra at 58-74. One failure of critical importance is that the
automated systems that filtering companies use to collect Web
pages for classification are able to search only text, not
images. This is crippling to filtering companies' ability to
collect pages containing "visual depictions" that are obscene,
child pornography, or harmful to minors, as CIPA requires. As
will appear, we find that it is currently impossible, given the
Internet's size, rate of growth, rate of change, and
architecture, and given the state of the art of automated
classification systems, to develop a filter that neither
underblocks nor overblocks a substantial amount of speech.


The government, while acknowledging that the filtering
software is imperfect, maintains that it is nonetheless quite
effective, and that it successfully blocks the vast majority of
the Web pages that meet filtering companies' category definitions
(e.g., pornography). The government contends that no more is
required. In its view, so long as the filtering software
selected by the libraries screens out the bulk of the Web pages
proscribed by CIPA, the libraries have made a reasonable choice
which suffices, under the applicable legal principles, to pass
constitutional muster in the context of a facial challenge.
Central to the government's position is the analogy it advances
between Internet filtering and the initial decision of a library
to determine which materials to purchase for its print
collection. Public libraries have finite budgets and must make