IP: "Artificial intelligence" filter blocks news -- but not smut

[Dave Farber <farber@cis.upenn.edu>]

>Date: Tue, 20 Jun 2000 09:53:37 -0400
>To: politech@vorlon.mit.edu
>From: Declan McCullagh <declan@well.com>
>
>Some pornographic images BAIR approved as OK and my Perl test-script:
>http://www.well.com/user/declan/bair/
>********
>
>
>http://www.wired.com/news/technology/0,1282,36923,00.html
>
>Smut Filter Blocks All But Smut
>by Declan McCullagh (declan@wired.com)
>
>3:00 a.m. Jun. 20, 2000 PDT
>When Exotrope Inc. introduced its BAIR smut-blocking
>software last year, everyone seemed wowed by the company's
>claims of intelligent filtering.
>
>New York Governor George Pataki applauded Exotrope's
>"state-of-the-art technology," Tucows Network gave BAIR five
>stars, and PC Magazine handed the program a coveted editor's
>choice award.
>
>But an investigation by Wired News shows that BAIR's
>"artificial intelligence" does not work as advertised.
>
>In tests of hundreds of images, BAIR incorrectly blocked
>dozens of photographs including portraits, landscapes, animals,
>and street scenes. It banned readers from viewing news
>photos at time.com and newsweek.com, but rated images of
>oral sex, group sex, and masturbation as acceptable for
>youngsters.
>
>Company representatives say they can't explain the program's
>seemingly random behavior.
>
>"I agree with you. There's something wrong," says Dave Epler,
>Exotrope operations manager. "That's not the way our image
>server is supposed to be working."
>
>Exotrope, a privately held firm based in Elmira, New York,
>claims to have developed "the industry's most advanced
>software system" for intelligently blocking sexually explicit
>images. BAIR stands for Basic Artificial Intelligence Routine.
>
>Epler said BAIR's smart-filtering, introduced in March 1999, had
>worked in the past. But he was unable to produce any version
>of the program that performed as described.
>
>Artificial intelligence experts say training a neural network to
>work the way BAIR supposedly does would be impossible.
>Anti-filtering advocates go a step further, and say Exotrope
>hoodwinked journalists and politicians into believing hype about
>advanced "artificial intelligence" and "active information matrix"
>routines.
>
>"I think all manufacturers of blocking software have suckered
>journalists and politicians to some extent by claiming it is more
>accurate than it really is," said Bennett Haselton, founder of
>Peacefire.org. "This is an unusual case because we're talking
>about a product with a zero percent accuracy rate."
>
>One reason why reviewers could have been mistaken is that
>Exotrope, like its competitors, has assembled a massive list of
>off-limits websites.
>
>If an image appears on one of those blacklisted sites, or if it
>has words like "sex" in the filename, BAIR automatically
>restricts it. It also will block access based on keywords
>elsewhere on the Web page.
>
>Wired News tested BAIR by creating a Perl program to extract
>images randomly from an 87MB database of thousands of both
>pornographic and non-pornographic photographs. The program
>then assigned each of those images random numbers as file
>names.
>
>The random number requires BAIR to evaluate the graphic with
>its "neural network" filtering to determine whether an image is
>sexually explicit or not. The sex-themed images came from
>Usenet newsgroups such as alt.binaries.pictures.erotica.female
>and alt.binaries.pictures.erotica.male.
>
>The results were dramatic: BAIR inexplicably blocked between
>90 and 95 percent of the photographs with no regard for
>whether they were sexually explicit or not. Of the ones that
>were OK'd, about half were pornographic and half weren't.
>
>BAIR incorrectly blocked photographs of Yellowstone, the
>Baltimore waterfront, Snoopy, boats, sunsets, dogs,
>vegetables and even a Wired News staff meeting.
>
>It rated as acceptable for minors -- even on the most
>restrictive setting -- explicit images of oral sex, anal sex,
>group sex, masturbation, and ejaculation.
>
>Exotrope officials say they plan to fix the errors within the
>next month. BAIR works by funneling Web connections through
>Exotrope's proxy server, which the company says is
>malfunctioning.
>
>"We're working through our image server problems as we
>speak," says Exotrope's Epler. "We'll have this thing up in less
>than 30 days. You caught us at a bad time. I know it works
>very well. I did accuracy tests on this thing 30, 60 days ago
>and it was perfect. It went south."
>
>When asked if Exotrope has a backup copy of a working
>image-recognition routine that was introduced in March 1999
>and could be installed on another proxy server, Epler said he
>did not. "I certainly don't know of a copy," he said.
>
>Epler also said he could not release a copy of the
>image-recognition routine that runs on the server, even under
>condition of a nondisclosure agreement.
>
>Dave Touretzky, a senior research scientist in the computer
>science department at Carnegie Mellon University, doubts
>Exotrope's claims.
>
>"How do you tell the difference between a woman in a bikini in
>a sailboat which is not racy and a naked woman in a sailboat?"
>Touretzky asks. "The only difference is a couple of nipples and
>a patch of pubic hair. You're not going to be able to find that
>with a neural network."
>
>"If they don't disclose the training data, there's no way to
>figure out what's going on," Touretzky says. "But anyone who
>knows anything about neural networks knows there's no way it
>can do what they're claiming."
>
>[...snip...]
>
>--------------------------------------------------------------------------
>POLITECH -- the moderated mailing list of politics and technology
>To subscribe, visit http://www.politechbot.com/info/subscribe.html
>This message is archived at http://www.politechbot.com/
>--------------------------------------------------------------------------