Version 1.3 (Revised 29-Sep-1999, Updates from Seth Finkelstein)

Rating does not work

by Kristian and Marit Köhntopp

Content rating models such as PICS have been proposed as a solution to the problem of unwanted, harmful or prohibited content on the Internet. This document contains a number of Theses which support the claim that any Internet Content Rating and Selection (ICR&S) scheme including PICS cannot work as advertised.

To our knowledge, most of the problems and objections here have not been addressed by PICS or any other ICR&S scheme.

Identifying the parties involved

This section tries to identify the parties involved in the process of running and rating the web and their roles (see below to understand why we concentrate on the web). In small installations, a single person may impersonate multiple roles.

Server side roles

The content provider is the role which is responsible for all content of a document. Often the content provider is the creator of a document, but on the Internet it is common that a content provider only provides means to create content, and does not actually create the content on a site. Examples are discussion boards, search engines, live video feeds, and similar installations. Depending on the size of this content, it is entirely possible that the content provider does not have direct knowledge of all content on a site and that much content on a site is not reviewed nor endorsed by a content provider.

The presence provider or web hoster provides the means to serve this content to the Internet by running the machines, the server software, and maintaining a network connection.

Recipient side roles

An access provider runs the systems and network connections for the recipient of content. For small office and home use, this is currently often a dial-up service, a proxy server, and similar hard- and software.

The recipient is either a person to be protected against harmful content, or an adult, which still should be able to access harmful, but not prohibited, content. The recipient’s hard- and software is maintained by system services which is a separate department in a school or library situation, in an Internet Cafe, or within a company. The recipient’s system may be a single system or a network of systems with proxy servers and intranet servers.

Internet Content Rating and Selection (ICR&S) roles

Developers of rating systems define the dimensions of a rating system and create rules how to apply values along these dimensions to content. They promote their rating systems so that they become popular and are widely used.

A rating service will apply the rating system and the rules that come with it to create ratings. These ratings, an identifier for the rated content, a date, an identifier for the rating source, and additional information (i.e. a checksum against the rated content and a digital signature) are collected to form a content label.

Content filter vendors create software which can regulate access to content, depending on local settings (filtering rules) and content labels.

Content filter control is often exercised by the party who controls a machine, that is, the adult party in a household, the dean of a school, the directorate of a library, and so on. These filter settings are then deployed, often by system services mentioned above, sometimes by an access provider located upstream.

Attackers may be content providers, recipients, or other parties who want to communicate outside of the control of a filtering system.

Methods for content selection

Principle of Operation

The basic idea behind Internet Content Rating and Selection is to attach a kind of machine-readable description, called a Content Label, to all Internet Content. The Content Label contains a set of ratings which make up a formal description of the rated content in a formally specified system of ratings. Finally, the recipient has to have a filtering mechanism before or on the recipients machine which allows or intercepts reception and display of requested content depending on the Content Label and some local configuration.

Taxonomy of Rating Systems by Source of Rating

The Content Labels may be provided by different parties. In Third Party Rating, a party that is neither the recipient nor the sender creates content labels and distributes them via a Label Bureau. Third Party Rating requires a method to uniquely identify content components, and Third Party Rating cannot be finer grained than this identification system. Currently, all identification systems are URL-based which implies that all Content Labels refer to either URLs or coarser grained objects (such as subtrees of a web server or an entire site).

In Second Party Rating, the recipient provides ratings and shares them with other recipients. This is sometimes referred to as a Community Rating process. Since the sharing of ratings again involves a Label Bureau, for the purpose of this discussion Third Party Rating and Second Party Rating can be treated alike.

First Party Rating is different because here the sender provides a Content Label with the content itself. Usually, this label is embedded into the content or sent with the content. A Label Bureau is not needed in this context.

Taxonomy of selection mechanisms by point of interception

The selection process at the recipients end can either be implemented directly on the machine of the recipient, or it can be part of a proxy solution upstream of the recipients computer. In the latter scenario, the selection process will not happen on a machine controlled by the recipient and it is much more difficult to manipulate by the recipient. A proxy-based selection requires that the recipient is forced to use this particular proxy to be able to access content at all (otherwise the recipient could elect not to use a proxy at all or to use a different proxy) and that the content can be identified and read by the proxy.

Theses

Internet Content Rating and Selection applies only to the Web

Internet Content Rating and Selection does apply properly only to the web, because of the little degree of interactivity and the large size of the communicated objects in the web. More interactive services or services with smaller objects are less likely to adopt such a system for content control with success. For example, USENET newsgroups are highly interactive and very fast communication channels with only little structure. The focus on discussion and contents, and the style of such discussions may change very quickly.

USENET articles are authored on the fly. Often they contain no keywords, subject lines of little use, or are inappropriately addressed. It is very unlikely that authors of USENET articles will take the time required to do proper First Party Rating for their articles, even if current USENET access programs offered such a feature.

Chatrooms and the Internet Relay Chat consist of even faster communication and no structure at all. IRC channels are often created on the fly and, unlike USENET, communication on IRC is not stored but only passed through by the servers. Also, the basic unit of communication in IRC is a single line of text which is much too small to rate properly. An IRC channel at any point in time is a very complex web of rapidly changing and overlapping individual and nonindividual communications that make up the overall tone and content of the channel.

Also, the public discussions in USENET newsgroups or IRC channels are only the publicly visible part of a much larger communication process which involves private communication by mailing lists, private e-mail, server based message communication, and direct client-client communication. Often the public communication channel is only used to meet other persons with similar interests, and then a more secluded communication channel is established.

Because of these properties of the more interactive communication mechanisms, Internet Content Rating and Selection will be ineffective in these type of media.

Labeling content that is not harmful nor prohibited is a requirement but cannot be enforced

The ultimate goal of Internet Content Rating and Selection is to make harmful content inaccessible to minors and make prohibited content inaccessible to all recipients on the Internet. Basically, it would be sufficient to label all harmful and prohibited content accordingly, so that it can be recognized and intercepted. For First Party Rating, this requires the cooperation of the provider of this content, so that the Label is being delivered with the content itself. For Third Party Rating, cooperation of the content provider would simplify the process enormously, for example by providing stable identifiers for the content or by organizing the content in a way that makes it easier accessible to quick and accurate rating.

It can be safely assumed that this kind of cooperation will not be available in many cases and that it specifically will not be available in the most severe cases where the content provider places content on the web with a malign purpose.

Thus, it is insufficient to block pages that are rated as containing harmful or prohibited content. Instead, all unrated pages must be blocked as well, so that the pages of non-cooperating content providers will be made inaccessible. In this case, the rating of harmful and prohibited content becomes a moot point because such content will be blocked automatically. The Internet as viewed from behind a filter will be immediately clean in such a scenario because it will initially contain no content at all. Clearly, this is not very attractive and most of the targeted audience for content filtering services will not tolerate such filters if large portions of harmless or valuable content will be made unavailable to them by the content filter.

For a Content Rating Solution, it becomes a requirement to provide Content Labels for non-objectionable content, so that at least some pages will be available to those persons behind the selection filter. Thus, all burden and cost of content rating, labelling, and label distribution is placed on the providers of perfectly legal, and in most cases, valueable content.

This situation is not different from, for example, the movie and video industry where a film without a rating is automatically rated as not suitable for minors (at least this is the case e.g. in Germany and the United States). Unlike the movie industry, many content providers have little or no incentive to have their content rated because in many cases they do not sell content or do not cater specifically to a younger audience. Specifically, we can expect the large commercial websites positioned towards a younger audience to provide content labels on a voluntary basis, while most private content and content not specifically geared towards a young audience will not have labels at all.

Consequently, it may be necessary to enforce the provision of labels with content to make a sufficiently large portion of the Internet available to minors. Some content providers will probably try to evade such a requirement by providing a “safe” default rating, such as “harmful content, not suitable for minors”, because they do not want to do the work to properly evaluate and individually label their content. Others will do this because they are insecure about the correct rating for their content, and will rate their content more harmful as it actually is, just to be legally safe. For a requirement to provide labels to be effective, it may be necessary to treat the provision of a too high label just like the provision as a label that is too low.

This turns the situation perfectly upside down: To protect against harmful and prohibited content, all work and all legal liability is at the side of the well-behaved members of the community, while it is technically completely unnecessary for the providers of harmful content to do anything.

In some countries, rating your own content cannot be legally enforced (see below).

Establishing a metric invites a dysfunction

A rating system for content is essentially a metric. The metric may have one (“suitable for age x”) or multiple dimensions (“contains sex x1, nudidity x2, violence x3 and language x4”), and within one dimension it may be non-ordered (“contains any of the following: smoking, gambling, drug abuse, violence as a method of conflict resolution”), ordered (“a violence rating of 3 means that there is more violence in a particular content than in a content with rating 2”), with discreet or continous values.

A rating system is a cultural code, too, because with the dimensions and values provided with any such metric comes a decription on how to interpret and apply them. As a cultural code, a rating system cannot be objective or valid outside a specific culture. Specifying the dimensions of the rating system, for example, defines which issues can be addressed by a particular rating system. Issues and values outside of the dimensions of the rating system cannot be addressed, and are therefore inaccessible to reflective discussion within the space of the rating system. For the purpose of the rating system, such issues and values essentially do not exist.

Values within the dimensions of the rating systems can be used to define proximity relationship if these values are ordered. Many people tend to think of content that is in close proximity as being similar, being of similar value, or promoting similar values. Some content providers insist that their content is not to be rated with some content rating systems because rating them with that particular system would place them into the proximity of other content they do not like, and with which they do not want to be lumped together [1][2]. For example, both hardcore pornographic bondage content and a documentation about torture methods in the holocaust would receive very similar content ratings within the RSACi rating system. The RSACi system is too coarse and has no appropriate dimensions to differentiate between these two types of content. More complex rating systems raise ease-of-use issues in First Party Rating situations, though. Also, more context sensitive rating systems raise other issues, too (see below).

Establishing a metric in a social context and granting rewards under this metric tends to create dysfunctional social systems. For example, measuring the efficiency of programmers by measuring the lines of code produced by them tends to work in favour of code that has been produced by cutting and pasting large segments of code. Such code is generally large, ineffective, and expensive to maintain. Thus, while the metric provides useful information when used short-term, it tends to work destructively when used for longer periods of time (Tom DeMarco, “Why does software cost so much?”, Essay 2). Essentially, degradation begins when the metric is used for a period long enough to allow for feedback loops.

A classic dysfunctional metric that can be observed on the web is rating the efficiency of a banner ad on a web page by counting hits to that banner ad. Using this metric for a longer time tends to encourage people to put content on their pages that creates page impressions and to load many banner ads on such pages. The result is your average free porn page. It also encourages spamming search engines and the network with ads for a particular page to increase page impressions. The ultimate dysfunctional perversion is to load the page with JavaScript code that opens further pages with banner ads when the user tries to leave or close the page. While the banner ads are not recognized by the user, at best, and create a negative image for the product advertised, at worst, such pages do generate lots of hits on a counter and are therefore favoured by the metric.

Because cultural and moral values cannot be expressed and measured directly like volts and ohms, any such metric will show dysfunctional behaviour and finally invalidate itself when used for a longer time. Obvious examples are things like exempting news sites from the requirement to rate content to relieve them from the burden to provide ratings for rapidly changing pages - this will provoke providers of content that rates bad under any given rating system to present their content in the format of a news site to get the bonus of exemption. Similarly, exempting web discussion forums from a requirement to provide ratings will make badly rated sites to adapt and to chose a forum-like format.

Because of the inability to address cultural or moral values objectively and outside of the context of a particular culture or moral system, it would be only logical and very tempting to create context sensitive rating systems. Context sensitivity here goes two ways: it may rate a specific content component within the presentation context, and it may rate specific content within a specific cultural context.

For example, singular pornographic images may rate bad in a certain system, but a photomosaic(tm) of girlie-power icon Lara Croft consisting of lots of such images may rate good within the same system because of artistic value and the message promoted by this artwork. So while the individual photo may still rate bad, in this particular context of presentation it will rate good. Also, a photo of James Dean leaning against his car and smoking a cigarette may rate good in some cultural contexts by showing a celebrity in a pop-art context, and bad in others as promoting environmentally bad modes of transport and nicotine drug abuse, depending on the priority of values of the culture perceiving the image. These priorities may change within the same culture or even the same recipient, depending on fashion or even personal mood.

What is more, only the process of filtering content which may appear as rating bad in a given context may actually create objectionable content in another context. For example, by suppressing all images which rate bad in one context there may be created a pattern that conveys a message of objectionable content [3].

So while context specific labelling is seductive as a way out of the dysfunctionality problem, it is also error prone, unenforceable (Would you please try to rate your web pages within the cultural context of the Amish People, the values of the current Nipponese, Spanish and Scandinavic societies?) and favouring one specific cultural context is of course an act of culture imperialism. It is also a value rating (“The context rating on your web page/for our web page in your rating service implies that you have authority to decide what Christian/Scientologian/Buddhist/… values are and that our page is bad according to what you assert these values are.”) with all problems that come with such ratings, including a ton of liability issues.

Translating from one metric into another does not work

Different metrics have different underlying cultural and moral value systems and different assumptions for judgement. Unlike algebraic spaces, the spaces created by such metrics cannot be mapped onto each other without loss (sometimes they cannot be mapped at all). Certain dimensions may not exist in the target metric and must be emulated. For example, it is impossible to translate the one-dimensional Video Standards Council metric for video films (“general viewing”, “PG”, “R”, …) into the RSACi (integer numbers between 0 and 4 in each of the categories of Sex, Nudidity, Violence, and Language), even though the target systems seems to be more expressive, having more dimensions and values.

Some proposals offer functionality that can be used to bundle presets in different rating systems into a single profile. For example, the PICSRules extension to the PICS system allows a user to define a single profile that contains an arbitrary number of selective criteria applied to different PICS compliant content ratings systems.

A user of such a PICSRule may be able to view content that rates this-or-that under SafeSurf ratings or such-and-such under RSACi ratings. Using such a system may create the impression that the this-or-that SafeSurf rating and the such-and-such RSACi rating are equivalent and that it is therefore possible to map one system onto the other. This is a thouroughly false impression.

What does work (in PICSRules) is to bundle arbitrary selective preferences into a profile, but this profile is in general (and probably in practice in most cases) not consistent or well-defined. Consider content that has been rated under two different rating systems and a majority of reviewers agreed in both cases that these ratings are correctly applied to this particular content. Assume further a recipient that is asked to define a filtering profile for his or her child in both of the rating systems reflecting the personal cultural values that this child should be able to expose itself to. Unless the page has a marginal rating (i.e. totally harmless or totally unacceptable), it will be common that the page will be rated accessible under one rating system, but inaccessible under the other, i.e. the ratings do not functionally map onto each other because of the different implied cultural values within the metric itself.

E-commerce and ICR&S are natural adversaries

The introduction of the Internet marks the ultimative shortening of the communication pipeline, completely industrializing the process of communication. Historically, the number of people and institutions involved in the production and reproduction of a specific publication has continually decreased. While many people had to copy each page of a book manually in medival times, the introduction of the printing press has dramatically simplified the process. The advent of DTP technology has further eliminated persons and institutions from this process by cutting out the typographer, the layouter, and the editor. Xerox technology has cut out printers for small editions.

On the Internet, all persons have been cut out except for the sender and the recipient. Production, reproduction and distribution of content are automated or controlled by either party. In many cases, content location has been automated, too, by using search engines, portal sites, web rings, or bookmark lists. In some cases, even content processing has been automated where automatic translation services, summary generator services, or XML postprocessing are already deployed, thus limiting the personal involvement of the receiver as well. Also, in cases where content is being produced automatically, such as in catalog systems, database driven web shops, or in community content systems like slashdot.org, the involvement of the sender in content creation is limited, too.

This does not mean that the absolute number of people involved in the creating or processing of content has decreased. In fact, there are many people needed to keep the Internet running and to update web content. But these people are no longer involved in a single specific project (working as craftsmen), instead they keep running a generalized infrastructure which is used to transport and produce many, many different works (they are working as industrial workers).

E-commerce directly benefits from this, shortening supply and retail chains to a length of 1, or where processing is completely automated, even to a length of 0. E-commerce requires that concealed, tamperproof 1:1 communication is possible, i.e. that the infrastructure can be used to transport certain transactional data, but without the ability to gain knowledge of the actual transactional content, and without the ability to change that content.

This is essentially the opposite of the requirements for ICR&S, because ICR&S is all about gaining knowledge about the transported content and from case to case tampering with it (i.e. replacing the original content with a “this content is inaccessible” message). You cannot build secure e-commerce on any network that is capable of ICR&S and you cannot build functioning ICR&S on any network that provides secure e-commerce services - see below.

Proxy-based ICR&S cannot work in an E-commerce enabled environment

In an installation where E-commerce is possible, it is impossible to successfully deploy proxy-based ICR&S, because at any time harmful or prohibited content can arrive in the form of E-commerce transactions, i.e. serving porn pages from an SSL server. The proxy cannot detect labels within such a connection because the entire connection is a military-grade encrypted, digitally signed, tamperproof connection pipe from within the browser program into the remote web server. Improvements in E-commerce, i.e. the wide deployment of S/WAN, IPsec and other services, will only worsen this situation.

Recipient-based ICR&S can only work in a cooperative setting

Because most current recipient platforms have no (Win 95, Win 98) or insufficient (Windows NT) local security, and because the recipient platform is generally under physical control of the recipient, there is essentially no way to keep the recipient from tampering with the installation of the selection mechanism.

Most local content filters can be easily deactivated by setting a few registry keys, rewriting a line in an *.ini file, or by copying some original operating system files that have been replaced by the filter back into the system. The process only has to be reenginieered once and can then be automated, requiring no expertise at all from the user of such an automated attack tool.

Third Party Rating cannot be based on URLs

All widely deployed systems for ICR&S are currently using URLs to identify content and to tie their ratings to such content. This approach is basically broken because URLs do not map to content, not even in the case of static content.

For static pages, current webservers offer a lot of facilities to negotiate the actual content which is being served in response to a specific URL. In particular, serving language specific content is very popular. For example, the homepages of multinational projects (like the Debian Linux project, http://www.debian.org ) are negotiated based on the language preferences that were expressed with a HTTP request. Also, many pages serve different content, depending on the IP-address or domain the request originated from or customize their content depending on the current time [4].

The problem worsens when it comes to dynamically generated pages, where the actual webpage which the client receives never exists on the server, but is created just as it is requested. The contents of such a page may change from one request to the next, even if all other parameters are identical. The server may take any parameters into account when creating the page, including internal state (memory of previous requests). Because internal state is kept on the server only, it is impossible to fully describe all request parameters needed to recreate a specific content without thorough knowledge of the application running on that particular server.

But even if we have perfectly static, non-negotiated content, URLs are inadequate to address content. URLs describe storage display objects such as images and HTML pages, but content can be smaller or larger than such pages.

For example, the start page of any portal, news site or discussion forum contains many small, unrelated articles which are presented in overview form. Each individual article is a semantic unit with different properties regarding a rating system, but with URLs we can only address the whole page and assign a rating to the entire page.

Conversely, a page or a site may be pieced together from individual documents which form a greater semantic unit that deserves a certain rating only when viewed as a whole. This may be true, for example, for an AIDS information site or a holocaust memorial site which may contain texts and images that deserve a very different rating when taken out of their original context.

HTML provides no mechanism to address subentities of a page or to address pages collectively. Some extentions to the XML specification will slightly improve this situation.

Third Party Rating cannot keep up

Third Party Rating is based on the assumption that a single database can list ratings for external pages, so that a recipient with an enabled filter can cover a large enough portion of the Internet for the net to become useful. Search engines work under a similar assumption, only they can automatically process and store pages, whereas proper content rating, being a process that involves moral and cultural evaluation and decision making, requires manual intervention. But a recent survey of search engines found that they cover only a small percentage of the web, ranging from 10% to 35% of all pages [5]. The survey also observed a strong growth in the general size of the web.

The problem is worsened for both ICR&S systems and search engines because there is no mechanism for web pages to notify indexing and rating services of a change. Web sites in general do not know which other services keep pointers to them and need to be notified of updates. Also, there is no mechanism to bind content identities to certain versions of that content, i.e. there is no way to work a modification date and/or a content checksum into a URL which designates certain content.

Third Party Rating is also impossible for content that is available only after passing a password protection (closed user groups, [5] talks about a “publicly indexable Web”). Some closed user groups are not so closed after all, though: With a simple registration, the protected area opens up. In such cases, the password protection essentially keeps all automated content processors and some casual browsers outside.

Third Party Rating creates privacy issues

Third Party Rating raises the issue of label distribution without cooperation of the content provider, leading directly to the creation of Label Bureaus which can distribute labels for any content, independently from the distribution of the original content. Connection between the content which has been rated and the label is being made via some content identifier, currently usually URLs. Currently, in some ICR&S systems such as PICS, there are simple provisions to test if the content that has been rated and the content that is actually served are still identical. Such a test may be a check for the creation date of the content or the distribution of a MD5 checksum for the content with the label. Usually, these provisions are optional and in many existing implementations they are not being used, leaving the system without a check for label validity (This should really be another headline, “Current implementations generally suck”).

To get labels from a Label Bureau, the client system has to ask the Label Bureau each time a piece of content is being requested. Because it can be expected that the Label Bureaus provides its services for money, we can assume that each of these requests is identified and authenticated, so that the Label Bureau can charge for its services. Label Bureaus will therefore automatically get large and detailed trails of all requests a certain user creates while surfing the web, creating a large user profile of (identity, time, URL) triples.

Third Party Rating has no standard complaints procedure

In Third Pary Rating, some group essentially applies a set of ratings to web content which directly influences the reach this content will have. Current implementations, such as PICS compliant systems, define how such a rating may be applied to web pages technically, but they define no standard procedures for rating services and labelling bureaus at all.

In current implementations, there is no standard procedure with which a content provider is being notified that content has been rated or on which grounds these ratings have been established. In fact, some (usually propietary, non-PICS) rating services do not expose their rating criteria for public inspection or even encrypt their ratings in trying to hide which services they disapprove of [6]. These services often argue that this procedure is necessary so that their rating service cannot be inverted, i.e. turned into a search engine for objectionable content. Decryption of such rating files invariably showed questionable or overly broad ratings, though. For example, in such rating files it is common to find blocking entries for websites that are critic of ICR&S or of this particular rating service. Also, there are often very vague and general blocking entries to be found, for example wildcard blocks for anything that has the letters “xxx” or “sex” in an URL [7].

Even if a content provider is being notfied that some content has been rated, there is no standard complaints procedure with which a content provider can go against a rating that he feels to be inadequate. Also, the whole rating process is intransparent and inaccessible to revision: there are no standard provisions that enable any involved party to detect why false ratings have been created: Was it a true false rating, a storage problem, a transmission problem, or a misconfiguration? On the other hand, false filtering can have substantial financial impact or can be equivalent to libel.

First Party Rating cannot be enforced

In the United States, certain First Amendment rights allow a content provider to not rate own content. Similar rights exist in other legislations.

But blocking unrated content unconditionally is open to unexplored civil damages issues.

First Party Rating does not scale down the problem enough

First Party Rating is basically a way to scale down the problem of Third Party Rating by one or two orders of magnitude.

Third Party Rating cannot keep up with the change of the web because of the sheer number of pages and because there is no way for content providers to communicate changes to Third Parties (see above).

First Party Rating puts the burden of rating content onto the content providers themselves which is good because it will automatically solve the problem of communicating changes. First Party Ratings must be controlled and validated, though, because a First Party will rate with a bias to promote own interests.

Thus, First Party Rating can be viewed as a mechanism to scale down the rating problem. Instead of n pages of content that have to be rated, a supervising authority now has the problem of controlling the correctness of ratings provided by m rating providers. We can only guess the relation of n:m, but we assume it to be in the range of approximately 100. m would still be a multi-million number.

Labels need to be tamperproof and tamperproof labels are expensive

Labels need to be tamperproof for two reasons. The first is LabelWasher, an imaginary sister program to WebWasher [8]. While WebWasher, JunkBuster and similar programs filter content to remove banner ad pages and malicious JavaScript code for advertisement free surfing, replacing the removed content with harmless dummy content, LabelWasher would do the same with content labels. That is, LabelWasher would remove content labels from incoming content and replace these labels with fake standard labels which rate this content as harmless.

The second reason for tamperproof, digitally signed labels is that the central authority controlling all First Party Raters has to have an instrument to put pressure on First Parties that do not rate according to the rating guidelines. The central authority would revoke the certificate needed to generate digitally signed labels from such parties, rendering all labels generated by them invalid.

Generating and administering a multi-million number of digital signatures is expensive, even if such signatures are low security (unfit to sign monetary transactions). Someone has to provide a database of all certificates issued, to verify identities of applications so that banned First Parties cannot apply again for a new certificate, and to perform all other administrative duties that come with running a certificate authority. Even if the cost for an individual certificate can be kept low, the other factor in the equation is still a multi-million number.a

Wildcard labels cannot be checksummed

A content checksum always applies to a distinct piece of content. A wildcard label such as “all content below http://www.site.de/ *” applies to an arbitrary number of pages with arbitrary content below a certain base URL. There is no way to calculate a checksum on such content, greatly diminishing the value of a tamperproof label: Without checksums and timestamps, it is impossible to determine which specific content the label was applied to.

Content rating cannot keep up with dynamic content creation

Dynamic content creation just gets faster and is often a parallel process. Examples for parallel content creation are discussion forum sites such as Slashdot (http://slashdot.org ), news sites such as CNN or Reuters, or catalogue and bidding systems such as eBay. Examples for fast content creation are the many live feeds for multimedia data into the Internet, such as live cameras (http://jennicam.org ) and live audio feeds.

It is impossible for these sites to rate their content, even with a general rating that applies to the whole site, because the content on the sites changes much faster than even First Party Raters can react and because in this case the First Party sometimes does not even have full control over the content on their own site.

On the other hand, exemption from rating for such sites will quickly generate dysfunctional behaviour and it will also violate the principle of equal treatment.

Dynamically creating content labels is expensive in current implementations

The Multipurpose Internet Media Enhancements (MIME) standard goes at great length to allow single pass implementation for reading and writing content. This is necessary so that dynamic content generators can read and write multimedia files, which are often very large, without buffering.

Current implementations of ICR&S systems ignore this requirement. They require content labels to be sent /before/ the actual content within the HTTP header, or to be part of the HTML document header. If the content label has to include a content checksum and a digital signature or even if it is just being generated dynamically, the label can only be generated after the actual content has been written. Content generation now becomes a two pass process, in which the content generator first writes out the actual content, creating the content checksum and the digital signature, and then begins sending content label and checksum first, then the actual content.

Buffer sizes for a busy site can easily become gigantic. For example, the Slashdot site generates much content dynamically at request time. A Slashdot page can easily be as large as 200 KB, depending on the number of discussion entries on such a page. Assuming ISDN transfer speeds, sending such a page consumes approximately 25 seconds. Slashdot can easily have 100 to 200 requests per second, or 2500 to 5000 simultaneous connections. Simple multiplication shows that this will consume between 500 and 1000 MB of buffer space in the form of either RAM or harddisk bandwith (the problem with disk buffers is often not space, but read/write capacity). For Slashdot, dynamic content with digitally signed content labels translates easily into twice the hardware it has now.

Streaming media is potentially infinite in length, so checksumming becomes a much more complicated process because checksums can only be calculated for chunks of data and must be embedded into a multimedia stream. This requires that the multimedia format used anticipates the need for such a feature or must be redefined (i.e. existing software must be rewritten). Another problem occurs with compound content in which some components are optional. Conventional checksums don’t work in this situation.

Finally, sending Content Labels after the content is not a good solution performancewise, either: It increases latency on the client side (that is, the PICS designers put the label before the content for a reason). When a client downloads content from the web, it has to decide whether to display the content or not. The client can only decide this after it has seen the label for that content. If sending the label is being delayed until after the actual content has been sent, building the display will substantially slow down because the client has to hold back content until it has seen the label for it. Incremental building of a display, as it is customary in current browsers, is out of the question in such cases. Besides, it will be frustrating and expensive for users if their client downloads a large file, only to decide to throw it away after the download has been completed due to the label for that content.

ICR&S systems may make error diagnosis more complicated and will decrease performance

The filtering component of an ICR&S system will most definitely add more components to an already overly complex platform. Adding such a filtering component will make debugging more complicated, even more so if the filtering component is designed to operate in stealth mode. In such cases it will become very difficult for support services to determine if a user encountered a true error or if a filtering component intercepted a download.

In the case of third party rating, the label bureau is a single point of failure: without access to the label bureau, no access to the network will be granted. The label bureau may also become a performance bottleneck.

Design changes at the local platform are necessary to enable mixed use (filtering and non-filtering) of that platform. For example, private web caches have to be redesigned. All current browsers have such caches and they are generally accessible via the filesystem with typically no access control. As a result, it is currently possible for somebody to access information directly in the web cache, which would not be available via a web browser due to it’s content label. Also, to associate certain access rights with certain users, it is necessary to identify that user. That is, user authentication (a login prompt) becomes mandatory on systems where ICR&S is being deployed.

False application

Problems exist at the filtering end of an ICR&S system, too. For example, on systems that see mixed use by minors and adults, it is highly probable that the adult user will not see indicators that an ICR&S system is active, and will get only limited access to web ressources. Alternatively, that person may perhaps see indicators showing that access is limited but is unable to turn the filtering system off because the password is not available to that person, or because the system is to difficult to figure out for that particular person.

In many jurisdictions (e.g. in Germany) making use of a constitutional right so difficult that persons will no longer make use of this particular right is equivalent to illegally limiting that right. This implies that the filtering component of an ICR&S system must clearly signal its presence to the user and must clearly advertise instructions on how it can be turned off.

Also, ICR&S that are difficult to turn off are unlikely to be accepted by users on systems that see mixed use. Things that become inconvenient are simply deinstalled.

False positives destroy trust in ICR&S and in the Internet

For an ICR&S system to be effective and trustworthy, it has to give access to a large number of pages, and it has to have a very low number of false positives. That is, it must not block pages which are essentially harmless due to wildcard labels to falsely applied labels. This implies that a large number of ratings must be created, and that these ratings should rather be too low than too high. It is highly unlikely (at least it seems so to the author) that any of this will be the case in the current hysteria and with the current legal situation.

A high number of false positives will not only create the impression of an unreliable and untrustworthy ICR&S system, but it will also undermine trust into the Internet as a reliable platform for communication of political ideas and social processes. If ICR&S systems are being publicly perceived primarily as a tool to leverage political or commerical interests, their value as a tool for the protection of minors is gone. This implies that ICR&S systems must have some mechanism to defend themselves against such an attack.

If we deploy now anyway, what do we get?

Summarizing the problems mentioned above inherent to ICR&S in general, and to current implementations such as PICS compliant add-ons to browsers, we can say that we cannot safely deploy an ICR&S systems on a wide scale and get anything remotely resembling protection of minors. Apart from technical issues, which are surprisingly hard to solve for a problem that seems to be trivial when approached first time, there are even more legal and management issues to solve before ICR&S systems can be deployed successfully. Most of these problems have not yet been tackled, some even have been left out deliberately up to now. Even if all the issues mentioned in this paper had been addressed by some system, the values of such a system will be questionable: It certainly is possible to deploy ICR&S systems which filter /something/ for /some/ people, but it highly unlikely that these systems will be effective, that is, that they filter /most things/ /correctly/ for /most people/.

[1]:

http://www.msen.com/~weinberg/rating.htm , “Rating the Internet”:

‘Jonathan Wallace, thus, in an article called “Why I Will Not Rate My Site” asks how he is to rate “An Auschwitz Alphabet”, his powerful and deeply chilling work of reportage on the Holocaust. The work contains descriptions of violence done to camp inmates’ sexual organs. A self-rating system, Wallace fears, would likely force him to choose between the unsatisfactory alternatives of labeling his work as suitable for all ages, on the one hand, or “lump[ing it] together with the Hot Nude Woman page” on the other. It seems to me that at least some of the rating services problems’ in assigning ratings to individual documents are inherent. It is the nature of the process that no ratings can classify documents in a perfectly satisfactory manner, and this theoretical inadequacy has important real-world consequences.’

[2]:

http://www.dcia.com/cyberbur.html , “Fahrenheit 451.2: Is Cyberspace Burning?”:

‘Kiyoshi Kuromiya, founder and sole operator of Critical Path Aids Project, has a web site that includes safer sex information written in street language with explicit diagrams, in order to reach the widest possible audience. Kuromiya doesn’t want to apply the rating “crude” or “explicit” to his speech, but if he doesn’t, his site will be blocked as an unrated site. If he does rate, his speech will be lumped in with “pornography” and blocked from view. Under either choice, Kuromiya has been effectively blocked from reaching a large portion of his intended audience ­ teenage Internet users ­ as well as adults. [ … ] Kuromiya could distribute the same material in print form on any street corner or in any bookstore without worrying about having to rate it. In fact, a number of Supreme Court cases have established that the First Amendment does not allow government to compel speakers to say something they don’t want to say - and that includes pejorative ratings.’

[3] :

http://www.mit.edu/activities/safe/labelling/0198f1.html where Lars Kongshem quotes a case where the program CYBERsitter suppresses the word “homosexual” on a screen display. Thus, the sentence “The Catholic church is opposed to all homosexual marriages.” is shown as “The Catholic church is opposed to all marriages.” on screen.

[4] :

Documentation for the Apache Web Server, mod_negotation Module, http://www.apache.org/manual/mod_negotation.html , mod_mime Module, http://www.apache.org/manual/mod_mime.html .

Both modules enable the server to deliver different static pages for the same URL, depending on other information that is part of the browsers request and that is /not/ part of the URL. It is the URL though, which ties a thrid party rating to a specific page.

[5] :

http://www.neci.nj.nec.com/homepages/lawrence/websize.html

“An estimated lower bound on the size of the indexable Web is 320 million pages.”

“The coverage of the major Web search engines investigated varies by an order of magnitude (variation will differ for different queries, e.g. more popular queries).”

“The major Web search engines index only a fraction of the total number of documents on the Web. No engines indexes more than about one third of the “publicly indexable Web”.”

[6] :

http://cgi.pathfinder.com/netly/spoofcentral/censored/index.html The Censorware Search Engine was helpful to check blacklists in commercial products which do not expose their blocking criteria. Seth Finkelstein reported this URL as being “gone and never coming back. The person who ran it lost the trust of almost everyone who does censorware-analysis work.”. Some vendors provide similar services on their websites, e.g. http://www.cyberpatrol.com/cybernot/ for CyberPatrol.

[7] :

http://catless.ncl.ac.uk/Risks/18.07.html#subj3 , Clive Feather reported comp.risks digest 18.07 of 17-Apr-1996 that AOL blocks the name of the small british town “Scunthorpe”, forcing inhabitants of that town to register as being from “Sconthorpe”

http://www.liii.com/~just4fun/news/article1.htm reports an incident where material of writer Anne Sexton is being blocked, because her name matches the string “sex”.

[8] :

http://www.siemens.de/servers/wwash/wwash_us.htm

“WebWasher® is a browser add-on that accelerates the navigation on the Web. The software runs on PCs or servers. [It r]emoves advertising on Web pages while you surf, Filters pop-up windows, animated images, referer.”

http://www.anonymizer.com/ offers services like

• Anonymizer URL Encryption - Beta: Extra Protection for the connection between your computer and our servers
• Anonymizer Window Washer by Webroot: Preserve your privacy & hide your tracks! Cleans your browser history and more...

[9] :

http://censorware.org/reports/ and http://www.peacefire.org/ contain a lot of information on conduct of Rating Services and Label Bureaus in general.

[10]:

A. A. Pitman, Scientific and Technological Options Assessment (STOA): “Feasibility of censoring and jamming pornography and racism in informatics, Draft Final Report, PE 166 658, Luxemburg, May 1997.

Some parts quoted in http://www.inet-one.com/cypherpunks/dir.97.08.28-97.09.03/msg00149.html

[11] :

http://www.research.att.com/projects/tech4kids/

Lorrie Faith Cranor, Paul Resnick, Danielle Gallo: “Technology Inventory: A Catalog of Tools that Support Parents’ Ability to Choose Online Content Appropriate for their Children”, December 1997, revised September 1998

[12] :

http://www.usembassy-china.gov/english/sandt/Inetcawb.htm “Political Security: A Closed or an Open Internet - The Great Red Firewall of China”, Brian Ristuccia, 31-Jul-1998:

[13] :

http://www.freshmeat.net/appindex/1998/07/31/901899756.html

“Anti-Filtering-Proxy-Proxy: httpd-afpp is designed to defeat the site-blocking fuctionality of censorware and filtering-proxies. The user first visits a site running the Anti-Filtering-Proxy-Proxy. Assuming this site is not blocked by the local filtering-proxy or censorware, the user is then free to browse any other web sites free of filtering-proxy or censorware restrictions.”,

[14] :

http://www.noie.gov.au/reports/blocking.html

Philip McCrea, Bob Smart, Mark Andrews: Blocking Content on the Internet: a Technical Perspective; Australia, June 1998,

[15] :

Gerry Miller, Gerri Sinclair, David Sutherland, Julie Zilber: Regulation of the Internet - A Technological Perspective; Canada, March 1999

[16] :

Technical Framework and Background:

http://w3c.org/PICS/
http://w3c.org/RDF/
http://w3c.org/DSig/
http://w3c.org/P3P/
http://mephisto.inf.tu-dresden.de/RESEARCH/ssonet/ssonet_eng.html

[17] :

http://www.iid.de/iukdg/carnegie_e.html “Misuse of International Data Networks”

Report submitted by the Expert Group to G8 Ministers and Chief Advisors of Science and Technology (Carnegie Group)

[18] :

http://www2.echo.lu/legal/en/internet/communic.html “Illegal and harmful content on the Internet”

http://www2.echo.lu/legal/en/internet/wp2en.html “Illegal and harmful content on the Internet Interim report on Initiatives in EU Member States with respect to Combating”, Version 7, (June 4, 1997) http://www2.echo.lu/legal/de/internet/resolde.html “ENTSCHLIESSUNG DES RATES ZU ILLEGALEN UND SCHÄDLICHEN INHALTEN IM INTERNET”

[19]:

http://www.securitysearch.net/search/papers/bsaprobs.htm “Key Legal and Technical Problems with the Broadcasting Services Amendment (Online Services) Bill 1999” (Australia) http://www.gtlaw.com.au/pubs/newdarkage.html “Censorship and Amendments to the Broadcasting Services Act” (Australia, April 1999)

[20]:

http://www.aclu.org/issues/cyber/burning.html “Fahrenheit 451.2: Is Cyberspace Burning? How Rating and Blocking Proposals May Torch Free Speech on the Internet”

[21]:

http://www.internetwatch.org.uk/ “The Internet Watch Foundation (IWF) was launched in late September 1996o address the problem of illegal material on the Internet, with particular reference to child pornography.”

[22]:

http://www.koehntopp.de/kris/artikel/blocking/index.en.html Kristian Köhntopp, Marit Köhntopp, Martin Seeger: “Blocking of Material on the Internet: Questions and Answers - A systematic analysis of the “censorship debate”; Kiel, Germany, May 1997

The article discusses non-PICS methods of blocking content and why they do create even more harm than PICS. It serves as an introduction to this article.