The 50 random pornographic images that we used in our test of the BAIR image filter

Unfortunately, while it would be easy for us to select 50 random pornographic images from the Web, it is not easy to prove that the images were selected at random. Exotrope could easily claim that the images in our test were selected on the basis of not being blocked by BAIR. In order to produce a "random sample" of images that could be duplicated by a third party, we used the 64 porn sites that were returned from a Google search for the term "titty fuck" "oral sex". We used Google since their search results are generally more consistent than the search results returned by Alta Vista -- so a third party can follow the Google search link and verify that the sites listed there are the ones that we used.

(Of course, since we're assuming that Exotrope might dispute our findings once this report is published, they could teach the image filter to recognize and block the images listed here anyway. Even if the BAIR filter cannot be taught to recognize pornography, it could probably be taught to recognize and block specific images, so that after they've made these changes, the images below will be blocked even if you save them and upload them to a new location. If that happens, then if you still want to test the BAIR filter for yourself, you can collect your own samples sample of pornographic and non-pornographic images and upload them to a private Web site, and see which images are blocked by BAIR.)

At each site, we downloaded and saved any images that showed bare female breasts or exposed male or female genitalia. (While certainly not all such pictures on the Web are pornographic, we assumed that any such picture displayed on a site that came up in a search for "titty fuck" "oral sex" would qualify as pornography.) However, we did not include pornographic pictures showing black or Asian models, because the BAIR product page suggests that BAIR relies partly on "flesh tone" (presumably meaning Caucasian flesh tone) in pictures to determine if they are pornographic:

"The BAIR may occasionally overblock non-pornographic images that have flesh tone or other characteristics similar to pornographic visuals."

We only used pornographic images displayed on pages linked from Google's search results (but this did include images displayed on other servers, as long as they were visible on a site linked from Google's results page). We also skipped ads that were loaded from a URL which indicated that the same URL would serve a different image every time, since this would have made it harder for third parties to repeat our results. If a site consisted mainly of links to other adult sites, we moved on to the next Google search result. Also, we did not use more than five pictures from any one site, since that would reduce the randomness of the sample.

These are the first 50 pornographic images collected from pages that turned up in a Google search for "titty fuck" "oral sex".

Most of these images will probably be blocked by BAIR if you load them from their original location given below. To test whether the image can be recognized by BAIR's "artificial intelligence" filter, which only looks at the content of an image, you would need to copy the image to a private Web site (as long as your site is not already blocked by BAIR).