This document is a NOTE made available by W3C for discussion only. This indicates no endorsement of its content, nor that W3C has had any editorial control in its preparation, nor that W3C has, is, or will be allocating any resources to the issues addressed by the NOTE. Publication of a W3C Note does not imply endorsement by the W3C membership. A list of current W3C technical reports and publications, including working drafts and notes, can be found at http://www.w3.org/TR.
This is version 3.1 draft -- very stable -- of this paper. An abstract of this paper has been included in the SSRN Electronic Paper Collection. I thank Lawrence Lessig, Joel Reidenberg, Lorrie Faith Cranor, and Ralph Swick for their excellent comments and suggestions that provided insight into concepts I had not understood well or had poorly expressed. I also thank Rolf Nelson and Robert Gellman for reading and providing comments that helped me improve the flow and focus of this paper. Any weaknesses in the technical and legal sections, or in the characterization of my reader's comments, are obviously my own and may need correction. I welcome comments from other readers.
Eskimos have many words for snow; Scots have numerous words related to rain. This concept has achieved near urban myth status -- though it continues to be contentious amongst linguists [Who40]. The idea is compelling because it speaks to our belief that the mechanism of speech itself is a reflection or our world and what we wish to say. Within this paper I examine the mechanisms by which our computer agents will express and understand what we wish to say in order to form online agreements.
This paper is about semantics. Semantics is generally defined as the study of meaning. While the description seems vague, it is of immense importance to the practical field of computer protocol design. For a computer program to be useful, it needs to know what a token such as "reset" means. The process of defining semantics within a specification (the document which details the structure and operation of some technical design) is critical to the operation of the protocol. How one defines -- or is permitted to define -- those semantics can have an effect on the social context and purposes the protocol serves.
My purpose is to examine the relationship between protocol design, policy, and law in three parts. First, I focus on the definition of semantics within a protocol and two metadata technologies (XML and RDF). This technical introduction permits me to refer freely to the technical domain when I begin my policy and legal analysis. Second, I look at the processes by which metadata semantics are likely to be defined with respect to consensus, conflict, and clarity. Finally, I review contract law as related to "the unruliness of words," [KGK86] misunderstanding, and interpretation. Note, one might answer two questions when addressing the validity of computer mediated agreements: (1) computer agent competency and (2) the semantic clarity of the agreement. This paper addresses the issue of semantic clarity; agent competency and agency are addressed elsewhere. [Rea99]
This paper is intended to be of interest to technical and non-technical audiences alike who are interested in computer mediated agreements. Specifically, where XML will be used in electronic commerce applications. Readers familiar with issues of protocol and metadata schema design may wish to skim up to §2. More extensive, but accessible, explanations of the technical issues can be found at [BC98, CKR98, Las98]. When terms are first introduced and defined they are represented as below:
This section describes how meaning is communicated through information technologies.
The three critical parts of a protocol are syntax, semantics, and state.
Consider the traffic light as a very simple protocol. The tokens (letters) in the traffic light protocol are green, red, and yellow. The syntax specifies that yellow follows green, red follows yellow, and green follows red. Any traffic light that permits a sequence of lights abusing this syntax is not a proper traffic light.
In this example, the state machine is closely linked to our syntax. The traffic light is in the red state, and is expressing the red "word." However, the state machine of a traffic light has more than three states. For instance, there are numerous states behind any red light. It may be counting down (like someone descending a ladder) such that it will fall into the green state within 1 minute. Or, it may be be counting up as the density of cars on the perpendicular road increases. Given the clear specification of the traffic light's syntax and state machine, a traffic lights can sit at an intersection happily churning through its various states while emitting a series of colorful words.
What is missing from my description? The semantics. It is only when the semantics of words are imparted to the driver does the traffic light protocol become useful. Green means go, yellow means caution, red means stop.
How do we create the protocol and ensure it works? Technical specifications are written to describe the syntax, semantics, and state of a protocol. Standards are developed to ensure independent parties can understand the specification. This promotes interoperability, improves the total social welfare of the community, and increases safety. For instance, a standardized version of our traffic light protocol serves all three purposes. Wherever I go in the US, my driving skill will interoperate with the traffic lights. Traffic lights promote efficient road utilization (flow control), and help prevent collisions. Those that abuse the protocol for their own benefit -- and to the detriment of others -- are incompatible or non-compliant and may need to be policed. Boston drivers are well known for running red lights -- I can only hope they get tickets!
XML (eXtensible Markup Language) is a specification of the World Wide Web Consortium, the standards organization for the Web. XML is a method by which users specify a syntax for structuring Web documents, sometimes known as markup. (Think of an editor "marking up" a manuscript with typographic annotations.) For instance, one can use XML to specify that an HTML page has no more than one title, and that after the title one can have any number of other elements within the body such as headings, paragraphs, and links. Note, one might speak of syntax at a number of different levels. XML enables users to create their own syntax for structuring documents (e.g., one title, one body, with many headings and paragraphs). To do this, XML itself specifies a syntax that (1) allows users to express their own syntax and (2) constrain what users can do.
Why or how would users want their own structure? One might use markup elements to represent legal citations on the Web. Consider the following:
ACLU v. Reno , 929 F.Supp. 824 (E.D. Pa. 1996)
The Bluebook provides syntax rules for constructing proper citations to legal works. In this case, it tells us that the following terms must appear in the following order with the proper abbreviation and formatting norms.
One could represent this in XML as
<title>ACLU v. Reno</title>
Applications can then store, process or display this information in many ways. Note, none of the key words like "title" or "volume" are defined by XML. XML just profides for the syntax, users define the words. Another example is that of using <book>, <title>, <author>, and <summary> tags to structure a bibliographic entry.
<summary>Metadata design has
The key words between angle brackets are known as elements or tags, information within an element are equated to each other and are known as an attribute and value, and the text inside the elements is the content:
The bibliographic example of XML also gives some indication of what role the content "Eskimo Snow" or "Joseph Reagle" plays in the document. However, it is extremely important to note that XML itself did not express semantics (meaning). In our example, one gets meaning from the natural language text within the tag. The following is an equally acceptable chunk of XML:
<field35>Metadata design has
The information is still structured, but the meaning of the tag is now opaque and since it is not captured within the tag by natural language, it needs to be explicitly declared elsewhere -- or possibly inferred from the content itself.
If XML is the mechanism by which one defines a set of tags and their relation to each other, how does one actually accomplish the task? Through a "schema." A schema is the definition of the tags and their relationships. Historically one accomplishes this with a schema definition language. The Document Type Definition (DTD) language describes the rules by which one defines the syntax of user created tags. For instance, the following excerpt from a DTD states that bibliographic entries must have one title, zero or more ("*") authors, and zero or one ("?") summaries -- the summary is optional.
<!ELEMENT book (title, author*, summary?)>
The XML specification defines the permitted structures that DTDs may generate. For instance, the XML specification states that no element (such as title) may have two definitions, as this would lead to confusion.
We now know how to define schema, but why do we care? Because it promotes syntactic interoperability: any client can immediately understand the structure of an arbitrary XML document. Even if my application never encountered a recipe before, if it encounters a recipe schema in XML, it can process the recipe structure, it will know that all recipes have ingredients without having to ask the user to intervene or install anything!
How then do we communicate meaning? It is wonderful for an XML application to be able to know -- on the fly -- that all recipes have ingredients. Wouldn't it be nice if it was able to figure out that barley malt is an alternative ingredient to honey -- again without the user having to intervene?
The Resource Description Framework (RDF) [RDF99], another W3C Recommendation, constrains the productions permitted by XML with a data model, and it defines its own syntax (a special set of XML tags) so that the meaning of tag relationships can be enriched.
There are three core concepts to RDF.
The RDF data model encapsulates all of these concepts so one has a way to think about the relationships between resources, properties and values. The way the data model is commonly portrayed is as a graph. Resources are ovals, with arcs extending from the resource to a square. The arc is labeled with a property, and the value within the square is its value. In this instance, the homepage has a property of type creator, with a value of "Ora Lassila."
Consequently, through XML we have the ability to make sets of structured information available to XML applications. Furthermore, one can use the RDF data model (graphs) and syntax (special elements) to describe the relationships between resources and elements. For instance, I could create an RDF statement that said the book located at some URI has the characteristics as expressed by my bibliographic example in the preceding section. Keep in mind that RDF does not necessarily look different from XML. Rather, it is XML, but with a particular data model and predefined set of tags.
Note, RDF does not permit us to describe all semantics. Only some meaning regarding the relationship of elements and resources, such as the things within the container (<ALT> ... </ALT>) are alternatives, or that one resource describes another. However, it doesn't capture the semantics that are likely to be meaningful to an ordinary person. It is useful for a computer to know the chunk of text "barley malt" is an alternative type of "ingredient" to another chunk of text "honey." It takes a human being that understands English to know the purpose of a recipe, how ingredients fit in, what barley malt and honey are, how to get them, and what they may cost. If I wanted the computer to cook me dinner, I'm likely to get hungry.
Let us imagine a computer agent (application) that helps users shop for and build a home computer based on the competitive prices of on-line stores selling computer parts. The agent is familiar with the computer parts XML schema. This schema specifies that every computer has a type of processor, RAM, hardrive, and price. A human creator might program the agent to understanding the meaning of price: that low values are better than high ones. Perhaps she will use RDF so the application can explore different prices for alternative components, or it will be able to check out reviews or statements made by others on the Web about that computer retailer. The creator of the agent also needs to program the agent on how to purchase things: if it decides the price/performance is right, it should take the information from the <purchase> element, invoke the checkbook application, and hand the payment information to it. It is the creator of an application that associated much of the meaningful semantics about the operation of the computer (make a sound, save a file, invoke another application) with the processing of data.
I've already described the benefits of syntactic interoperability: any application can understand the structure of a document the first time it encounters it. Semantic interoperability gives us more. It allows us to take the semantics that associate invoking a payment application from a <purchase> tag and understand and share that as well. Whereas no one conceived of online grocery shopping, my agent might already be familiar with recipes, as well as buying computer parts, there is no reason it can't buy ingredients (recipe parts) on-line! The more meaning (how to agree, how to invoke, how to buy) we capture in computer understandable schemas, the more they will be able to help us. The authors of [CKR98] described this characteristic well:
And yet the ability to combine resources that were developed independently is an essential survival property of technology in a distributed information system.... In his keynote address at Seybold San Francisco [Berners-Lee, 1996], Tim Berners-Lee called this powerful notion "intercreativity".
Presently, the most reasonable way to define operations and methods as part of a schema is to rely upon Remote Procedure Calls, object oriented network repositories, or computer languages like Java. How these might be best integrated into syntax and other semantic definition languages is an ongoing area of research.
The final issue in this brief technical exposition is that of name spaces. As I already demonstrated in my description of RDF, the URI (Uniform Resource Identifier) is a way to identify and locate a resource (object) on the Web. Namespaces [BHL99 ] address the question of how to identify a schema, and how to mix various schema within one document. They accomplish this by using a URI.
Namespaces appear in XML as follows:
<book:summary>Schema design has
The prefix "book:" clearly signifies that a tag is defined by the book schema at the URI provided. Why do we need to do that? It allows us to represent the element and any semantics associated with it unambiguously. Consider the case in which a different organization defines a schema for a poetry book; they allow more than one title to appear within a book, since a poetry book has many poems, each with its own title. Rather than requiring a centralized registry for all elements, namespaces permit anyone to use any element they want in their document, and to combine schemas, as long as they qualify the element name. Note, schema designers should only use names that include URIs under their administrative control. For instance, a third party should never declare a namespace of the type "http://w3.org/bib.schema" if she does not control the identified resource because it may be deleted or changed. (See §3.4.2 for an example of why this is important.)
Before proceeding to §3, where we learn about the policy implications of schema design, I wish to reiterate a few key points.
In Judeo-Christian theology the first power granted by its God to man was the power to name, "Out of the ground the Lord God formed every beast of the field and every bird of the air, and brought them to Adam to see what he would call them. And whatever Adam called each living creature, that was its name." (Gen2:19-20). Designing a schema that others will use is -- in some sense -- an exercise of the power to name.
This section examines the policy implications of schema design. When I refer to "policy implications" I mean the substantive non-technical issues related to the schema design process and its result. A substantive issue is when a question is likely to be decided on the basis of something other than technical merit, or if the answer to that question is likely to have an effect on a non-technical domain. The non-technical domains obviously include legal, political, social, and market concerns. While this section's example application is that of privacy, the questions raised apply to the design of commerce, intellectual property, or any other schemas that capture relationships or are used to reach agreement.
"Social protocol" is a term I use to discuss protocols (or applications) that enable individuals and communities to express social capabilities. This would include tools necessary for creating rich content, managing trust relationships, conducting commerce, making verifiable assertions or recommendations, and enabling agent assisted (or automatic) decision making. The ability to make verifiable assertions, to build reputation, to solicit advice, make agreements, and defer to a trusted source are all real world capabilities. This is what real world relationships are built on. If we want sophisticated cyberspace relationships, we will need similar mechanisms.
While I do not wish to spend too much time on the concept of social protocols, (see [Rea98] for more background), the important factor for analysis is the degree to which the semantics and operation of social behavior/structure are captured within the data structures and protocol of the application. This is critical because it is at this point that social/legal concerns and methods have an undeniable effect on technical design.
My example application of schema design will be privacy. Consider a metadata schema that was designed to communicate natural language semantics of a Web site's privacy practice (e.g., purpose of data collection, whom it is redistributed to, etc.). This is one of the key components of the W3C's Platform for Privacy Preferences (P3P), its Harmonized Vocabulary is just such a schema. I will examine issues related to achieving (1) consensus in schema design and (2) semantic clarity in its definition and implementation.
The intent of the P3P Harmonized Vocabulary is to express data collection practices in a proposal such that it can be acted or mediated on by the user's computer agent. The agent does not truly understand the natural language semantics of "the purpose is for marketing activities," but it can understand that the user told it to reject proposals with purpose=4.
The goal of P3P is to enable users to operate within a sphere of policy of her choosing without necessarily having to investigate the -- sometimes confusing -- natural language privacy practices of every site she visits. For instance, a user could configure her agent to permit the release of her zip code, and remain otherwise anonymous, such that she can received geographically customized services. Also, the user would configure the agent such that when the agent encounters a site requesting personal information, the site's privacy practices must be assured by a trusted third party, and only then would the user be informed of the site's other practices and prompted for the information. She can now browse the Web confidently because she has an explicit agreement about the privacy practices of the site.
Anyone can design a schema. One merely defines it in a language (such as DTD), places it on the Web (at some URI), and everyone else in the world can now use it to structure his or her information. However, for a schema to be widely used, it must offer benefits to its users.
While anyone can design or use as many schemas as they wish, schema adoption certainly exhibits the same network effects and externalities of other technologies such as the fax or VHS standard. If 90% of the world uses one privacy schema, there is little incentive to use another. However, unlike other technologies such as fax machines or VHS, there is little technological lock in. It is technically trivial to change or adopt schemas. Syntactic interoperability ensures that when a new schema comes along, one does not have to throw away one's old applications or technology to understand the structure of the new schema. For instance, applications based on XML are like a VCR that could play VHS, beta, or any other format. Admittedly, there would still be pressure to minimize the number of "video schema." However, the success of a schema is likely to be based more on the merit of the schema design rather than success derived solely from a first mover advantage, market dominance, or technological lock in.
To rephrase the preceding paragraph: whereas many competitive schemas may exist for a given application, not all are equal. This is obviously related to the human cost of understanding and supporting other schema. (While open metadata structures lessen this cost, it remains above zero.) Consequently, parties will often be interested in standardizing on one schema when possible. This is certainly the case of privacy disclosures, which carry a great deal of natural language privacy semantics that are costly to communicate to users. (Users that have to understand different -- and often inconsistent -- terms and descriptions are not likely to truly understand the disclosures.)
Consequently, one is presented with a design scenario in which conflicting parties know it is in their interest to standardize upon one schema, but differ regarding what should be in that schema. I identify three classes of tension:
In the privacy domain, examples of (2) and (3) are readily available. One can imagine a company that sells information to third parties shying away from redistribution disclosures. One can imagine a privacy regulator wanting as many disclosures as possible, particularly with respect to a service's jurisdiction or compliance with a legal code.
However, it would be naive to assume that all users of information fall into class (2), and all regulators fall into class (3). For example, the organization TRUSTe has developed a Web icon/seal that is used to indicate membership in TRUSTe.The presence of the icon means that the privacy practices abide by TRUSTe guidelines. Originally, the icons made a privacy disclosure: there was a different icon for each of the three ways a company might distribute user data to other parties. One of the alleged reasons TRUSTe moved away from three trust marks was that it was not granular enough; companies felt they were lumped together with companies that had very different practices. A second example is the concern of regulators about the presence of some disclosures in P3P. These disclosures are about practices that are not legally permitted in a given jurisdiction -- having the ability to express an action implies the right to do so. The inclusion, exclusion, and definition of terms in a schema can be a highly political exercise for all involved.
So how does one resolve the tension between the desire to standardize upon contrary interests? The approach taken in the P3P Harmonized Vocabulary, and one I recommend for other applications, is to try to standardize as much as possible, but to be cognizant of reality: "No single rating system or schema can satisfy the whole world unless the whole world is of the same opinion." [Rea97] Efforts should be spent on developing a core set of schema definitions that all parties can agree to. Competitive or supplementary schemas may develop as a response to the core schema's inadequacies. For instance, a different group could develop a competitive schema that in time supercedes the first. Or, a group (perhaps the original one) may supplement the original schema with disclosures that were not required or well understood at the time. These multiple schema's can be easily intermixed using the XML-namespace facility.
Where consensus cannot be achieved, the environment of adoption and deployment (individuals, markets, regulators) would be the ultimate judge of which schema (or set of schema) is most appropriate -- I refer to this process as the schema ecology. Again, the charter states the group's philosophy: "... acknowledges that if two constituencies disagree about X, they disagree. And if constituencies browse the Web space of others, all of their unique concerns may not be addressed in that other space. " [Rea97]
It has been observed that philosophy of seeking unanimity on a core set of definitions can lead to minority veto power: that one party can easily keep a particular disclosure from being included; this is true. The question a schema design group is then presented with is, is this core schema at least good enough to be acceptable as a minimum? Is the lack of a disclosure so important that one cannot support a core schema without it? Only the consensus process of a group can resolve this issue. However, the important result of this section is that the consensus process need not be orientated towards reaching the complete and exclusive solution. Rather, multiple schema can exist, compete, and/or extend each other. When compared to the standardization of other technologies, the adoption of a schema will be predicated more on its merit than the brittleness of the technical or political institutions from which it came.
If one assumes complete consensus on the goal of a schema design, its implementation and deployment may still lead to confusion. This confusion is likely to result from a poor understanding of implicit (default) semantics or a conflict of semantics within or between schema.
Schema definition languages (such as DTD) allow one to define default values for elements or attributes that are not present. For instance, the P3P like statement below
<STATEMENT VOC:purpose="2,3" VOC:recipient="0" VOC:identifiable="0" consequence="a site with clothes you'd appreciate.">
could be defined in a DTD such that if the "VOC:purpose" (purpose of data collection) attribute is not included in the <STATEMENT> element, the application should assume it's present but with a value of "0." In P3P, a value of "0" means "Completion and Support of Current Activity." One can also decline to set a default, in which case the absence of an attribute means no semantic is expressed. This issue is also subtly linked to the concept of, "that which is not stated, is not done." For instance, in P3P, the specification defines the semantics of a privacy statement such that the service is obligated to make representations of all her privacy practices as they relate to Web interactions. A lack of the statement is a representation that the corresponding action is not done. Interestingly, there is a parallel to these concepts of defaults in contract law, such as caveat emptor and Expressio Unious Est Exclusio Alterius as discussed in §4.3.3.
Regardless, why would one use defaults? First, it permits brevity. The terms that are most often used can be made the defaults, and need not appear in the proposal. (Recall the proposal is the set of P3P data collection and privacy statements.) Second, as originally formulated in P3P, it compels the service to make a statement. A service can not turn a blind eye to a disclosure they do not like, since its absence commits them to a meaning regardless. If a service ignores the purpose attribute, a misrepresentation that the purpose is only "Completion and Support of Current Activity" could be fraudulent.
However, this leads to a rather interesting problem that I characterize as syntactic interoperability with semantic fault. A service, with many purposes, might not understand P3P very well, and may assume the purpose declaration is optional. If the service does not include the purpose attribute, the user is more likely to release information since the purpose seems benign. While the protocol and syntax give no indication that there has been a protocol failure the user gave out information she should not have! I generalize this observation to state unlike other failures, semantic faults are not readily manifested but have substantive impacts on the user.
If one can assume ordinality, (that a purpose of 5 is worse than 4 with respect to privacy), one can pursue a different approach: the default values should be high. This means that the default value of purpose would be 5:other purposes. This still results in services having to counter the implicit semantic, but a failure to do so doesn't harm the user, since information is less likely to be released. If one can not assume ordinality, one could make the default be that all values (1-5) apply, but the DTD syntax does not permit such a default definition given P3P's present design.
The final strategy was to keep semantic requirements separate from syntax requirements. This means that one should not use the meaning of a default value to require a service to abide by the syntactic requirement that a attribute must be present in the statement. If the purpose attribute is required (syntax), it is defined as such in the DTD. It is up to the XML application to enforce this requirement by rejecting any markup that lacks the attribute.
It is interesting to note that this technical discussion of how to structure default corresponds to the legal theory of default rules in contracts. The traditional legal approach is to define the default term of a contract -- if it is not explicitly defined -- as that which a majority of those that make contracts would want. This is akin to the technical goal of brevity and compact proposals!
Ayers and Gertner promote a different idea, one very similar to the idea of defining a default such that a service is obligated to explicitly counter it, "Penalty defaults are designed to give at least one party ... an incentive to contract around the default rule and therefore to choose affirmatively the contract provision they prefer." [AG, p. 91] However, while this may lead to the efficient establishment of contracts because it forces information to be revealed, it does not help in our example. The possible use of default rules in P3P compels the disclosure of a predefined semantic, not information that is likely to be novel or unique to the circumstance as does the legal default.
In American culture -- at least -- there is an odd convention, particularly among children and politicians, whereby someone can negate the meaning of their statement by crossing their fingers behind their back. Can one do the same with metadata? They can try. I've already recommended that when consensus cannot be reached competing or supplementary schema might develop. Consider four scenarios that might result:
To finish this section, it is worthwhile to examine how a different metadata technology, the Platform for Internet Content Selections [PICS96], addressed issues of semantic clarity and conflict -- we will return to this topic in the legal context of §4.3.3. PICS, like RDF, allows one to associate a label (metadata), in a rating system (schema) with a Web resource. PICS includes mechanisms for stating two other things:
The first mechanism is useful when more than one label (piece of metadata) exists for any given resource within a single schema. For instance, a "thumbs up" label in the Siskel_and_Ebert schema might apply to the whole Web site and sit on the home page, a "thumbs down" label exists five mouse clicks away from that home page. In this case, one would wish to say on the home page, that while this "thumb's up" label applies to the whole site, a more specific and appropriate label might exist elsewhere on the site, in which case the one on the homepage should be ignored. PICS includes syntax for distinguishing between a generic label (which can be qualified but not overridden) and a default label (that can be overridden.) P3P solves the problem of ambiguous semantics by requiring that an identifier for any proposals relevant to a URI be presented to the agent when it accesses that URI. (This is the hash of the proposal, propID). Consequently, there is no ambiguity as to which metadata applies to any given URI. If two proposals apply, two identifiers will be provided. If there are conflicts between the proposals, it is up to the P3P agent to discover them and report an error.
I've already spoken of the second mechanism, that of mandatory/optional extensions. Earlier I posed an example of Europeans requiring services to use a jurisdiction schema. How could they compel a European agent to be cognizant of that schema when intermixed with a P3P proposal? While there are no mechanisms for this in schema definition languages presently, there are extension mechanisms in protocol design. For instance, within PICS the extension mechanism can be optional or mandatory. One refers to an extension by way of a URI. If an extension is qualified as mandatory, the client must use the extension, otherwise it must reject the entire service. This mechanism was used to extend the digital signing capability of PICS. Consequently, one can create a label such that if users do not understand digital signatures, one can tell them not to risk processing the labels at all. I expect this capability will be extended to schema definition languages.
So far I have focussed on how meaning is captured within metadata. The goal is to allow such information to be readily available to users or their computer agents. The obvious question now, is what happens if there is some confusion in the design, specification, or presentation of that data? (I am not addressing the question of error in the agent, I shall largely assume a perfect agent for the purposes of this paper.) To state the question more directly, if users are going to reach agreements on the basis of metadata, what happens when an agreement was based on a misunderstanding about the meaning of the metadata?
Contract law enforces or binds parties to their agreements. It also includes standards of resolving confusion related to ambiguous or misunderstood terms, as well as providing default terms when they are not present in the contract. Courts are called upon to determine whether a contract is valid, and to whom the liability or risk in a dispute falls upon through the enforcement of the clarified contract. I will examine the relevant legal standards to shed light on how issues of confusion in schema design will likely be met by law. My goal is not to review all of contract law, but to call out salient points that will be notable when applied to computer mediated agreements.
How is a contract made? The process of forming an agreement is straightforward: there is an offer, consideration, and an acceptance (manifestation of assent). The necessary ingredients to this process are two or more parties, the capacity of those parties to reach agreement (resulting from the competency to do so or because they are not otherwise precluded by law), mutual assent based on some active consideration or other method of validation, and an end agreement that is sufficiently definite. In computer mediated interactions, these requirements seem to be easily met with careful software design.
Within this section I have the difficult task of introducing a wide body of law that pertains to the legitimacy of on-line agreements while strenuously attempting to limit the scope such that I can focus on issues relevant to schema design. Consequently, I hope those unfamiliar with the law will be satisfied with the brief introduction I provide, and I hope experts will forgive my summary treatment of the concepts. My goal is not to argue that computer contracts can or should be made, but largely assume they will be and ask what things will technologists and lawyers need to pay particular attention to in the design of metadata schema.
Contract law originated in common law, the accumulation of principles and standards resulting from prior court decisions. This is different from a legislative and statutory approach. US contract law is a hybrid of sorts. One of the features of this hybridization is two institutions that harmonize court decisions. The two institutions are NCCUSL's Uniform Commercial Code (UCC), which drafts laws that are then adopted by US states, and ALI's Restatement, which derives its authority from the soundness of its reasoning. These two texts are the basis of much contract related teaching, analysis, and law.
Interestingly, it's the non-democratic nature of these institutions, particularly the UCC, that has led to much of the protest regarding changes of UCC Article 2b (the Proposed Law on Software Transactions) with respect to online interactions. UCC working groups are generally the provinces of legal experts, state lawyers, and lobbeyists. Once an article is formulated and approved they are often adopted by state legislatures with little subsequent discussion. Advocates have protested that the crafting of UCC 2b has exceeded the integration of current case law. Instead it creates new policy, biased by special interests, and conducted in a closed forum. My inexpert opinion on the topic is that much of UCC 2b is premature. Beyond the need to provide an online equivalent of a signature and requirements for evidence, I believe the novelty of cyberspace is being used to bypass principles and balances associated with preexisting social policy and law that can and should be applied to the online environment. These pre-existing principles are what much of the following analysis is about.
Contract law has many dichotomous categories for contracts: written and oral, express and implied, and enforceable and unenforceable. Since the purpose of this paper is not to address whether computer mediated agreements are of one type or the other, I shall quickly argue (and otherwise assume) that because computer information is explicit it is more often on the left side of the dichotomy (written, express, and enforceable) than not -- with the following caveats.
First, computer information is similar to written information in that it is recordable: logs can be saved and examined by third parties. This also leads me to argue that a computer agreement is rarely implied. However, just because information can be recorded does not mean that it will be recorded. The persistence of metadata agreements and logs by the user's agent is dependent on the agent's storage capacity -- which might be limited. A technical and legal understanding needs to define the requirements for such information to be used as evidence in disputes. These requirements would likely include the format, timeliness, and security of the log file.
Second, the very purpose of metadata is to express information sufficient for communicating some semantic to another party in a way that is deterministic, persistent, and confirmable. However, while the terms of an agreement are likely to be express, the assent of the agreement may not be -- depending on the design of the protocol. For instance, one can imagine a protocol whereby assent is implied by further interaction with the service; refusal is implied by stopping interactions with a service. For instance, P3P combines both methods. Proposals are presented to an agent, which signals its acceptance by sending an (OK) code and a propID (an identifier of the proposal) with subsequent interactions. An agent need not signal non-acceptance, it can simply discontinue the interaction.
Finally, there is enforceability: will a court uphold a contract? (This is distinct from whether a contract is found to exist, or if it is voidable.) The most common example of an unenforceable contract is when its scope is subject to the Statute of Frauds. In such an instance, a court will not enforce a valid contract for the exchange of goods if it did not meet the formalities required by the Statute, such as the presence of a signature. Consequently, while it is quite possible that subsequent interactions based on a metadata proposal act as an assent to a valid contract, a court would not enforce the contract without a signature. This is obviously the motivation for much of the activity related to digital signatures.
Fortunately, I can narrow the scope of this analysis even further. Much of contract law ("mistake" and "impossibility") concerns itself with issues that are independent of the medium of expression. Namely, the effect of unforeseen occurrences or changed circumstances that undercut the assumptions implicit to a contract, neither of which are uniquely salient to metadata agreements.
A mistake is a belief about a fact at the time the contract is made that turns out to be false. This does not relate to predictions about the future (e.g., stock prices) but the facts of the contract. For instance, if I was told John Smith was a good person, and I entered into a contract with him, but later I learn it's the wrong John Smith! I made a mistake. Impossibility is where an assumption basic to the contract -- so basic it seems it need not even be stated -- turns out to be false and makes the completion of the contract impractical. For instance, if I was contracted to perform in a concert hall, and the concert hall burns down, it's now impossible for me to perform there.
Within this paper my main focus is on the issue of how the data structure design interacts with legal concepts of misunderstanding: where one or more parties to the agreement did not properly understand the terms of the agreement; where the contractual instrument did not cause a set of common semantics to be shared between the parties.
Again, the question here is not about the validity of the agreement because of a characteristic of the computer agent, but because of the schema design itself. Consequently, I will disregard the scenario where an agent acts without the interaction with the user. In my example, the agent merely presents information to the user based on a metadata statement. It makes no decisions on its own.
Accordingly, let us assume a scenario in which a service used a metadata statement to solicit information and describe the related privacy practices and consequences (benefits) of that collection -- perhaps a free calander sent in the mail. The user released information, but then realized someone made a mistake. Someone was confused about the agreement because she is now getting lots of junk mail. How, in contract law, does one challenge such an agreement, and how is it likely to be settled? There are three basic problems that both law and technology must confront:
I've already introduced the technical mechanisms of addressing some of these problems in §2; I will now integrate those mechanisms into the legal context.
While I have said I shall largely avoid issues of mistake, an instance of a mistake was where technology got its first introduction to contract law. The seminal example is that of the telegraph. Imagine that an offer is sent for one amount, but received as another. (Des Arc Oil Mill v. Western Union Tel. Co., 132 Ark. 335, 201 S.W. 273, 6 A.L.R. 1081 (1918). In this example, an error in the transmission mechanism caused a syntactical error, corrupting both the structure and the consequent meaning of information. One can imagine an equally likely scenario, where information transmitted over the Internet becomes corrupted.
However, this scenario is not all that salient to schema design. It is relevant to the network protocol design, but network services are certainly capable of providing robust and high integrity information transport.
Before information was ever sent over the Internet, meaning had to be captured in a data structure. That metadata is then transmitted -- as discussed -- and rendered to the user. At either end of the communication, in the capture and reconstruction of meaning, there is room for error and confusion. If a service states within a P3P proposal that it uses information in identifiable form, and the graphical user interface (GUI) that presents this information to the user is confusing, the medium has failed to properly transmit meaning -- semantics!
Consequently, there are three areas where errors might be introduced in the transmission of information. First, if the transport mechanism of the Internet actually garbled the metadata on its way to the user, one has a case of syntactical mistake resulting in semantic confusion. However, this error is readily detectable (i.e., message authentication and integrity checks). Second, in the case of confusion about a schema and its representation, one has a more problematic scenario of semantic misunderstanding. Recall that in §3.4.1 I said that in instances of "syntactic interoperability with semantic fault" errors are not readily manifested but have substantive impacts on the users. If a semantic error arises from confusion, neither party may know of the error until after the contract has been agreed to. Courts are then called upon to interpret the contract and discover its true meaning. Third, the GUIs might cause confusion indepdent of the schema design.
This later point of confusion introduced by the display mechanism is sometimes characterized as a mistake of display. Fortunately, the legal system has given thought to the presentation of contractual terms. The UCC had defined what it means to be "conspicuous." [UCC §2-102(a)(7)] (See [Efr97:1356] for analysis.) While I find their text to be purposefull naive (e.g., all CAPS seems like a poor remedy) I would expect the presentation of such terms on-line to be as open to review by an informed court as that of printed text.
Consequently, if substantive terms are being negotiated through a GUI, GUI designers will have to be very careful about how they present those terms. Particularly if an agent encounters a schema it has never seen before -- as they are liable to do as explained in section §2.5. In such a case, the agent designer will not have had a chance to adjust the GUI to prominently display the most relevant terms of that schema. Instead, schema designers will likely have to provide hints in the schema design itself about which terms are most important. Then, the GUI design does not then need to be familiar with the semantics of every schema and design a custom interface, instead the design can interpret and render the hints appropriately. In time, model guidelines and experience in GUIs will advance such that we will have a better understanding of what a reasonable GUI should look like for communicating substantive terms. (Lessig comments that as our understanding of what a good GUI looks like, so will our legal understanding of what is reasonably comprehensible. This then influences the understanding of what types of subjects and terms ought to be captured in such agreements.)
How might we remedy this problem of misunderstanding resulting from the schema design or GUI ? Instead of relying purely on the presentation of the metadata, users might be encouraged to print out and read the complete natural language privacy practices of a Web site.
Imagine one user makes a decision to accept a proposal on the basis of the GUI presentation of the metadata, another user prints out the corresponding natural language statement, and also makes a decision to accept the proposal. Have both users agreed to the same thing?
A problem with privacy practices and shrink-wrap licenses today is that they are essentially meaningless to the user. The text is so complex and confusing users simply do not understand the terms, even if some are capitalized. This is one of the problems metadata should be able to remedy, to capture consistent semantics across services in a way that the user can use a computer agent to assist them in understanding those terms and to act accordingly. However, if one then says, you cannot rely upon the metadata itself, but must print out and assent to the natural language description as well -- or at least that is the user's obligation -- we have not solved the problem! It may have made it more confusing!
I've presented a scenario where one questions where the "true" semantics of an agreement reside. In §3.4.2 I also spent a fair amount of time describing the technical issues associated with semantic conflicts. If more than one statement exists they can extend or supplement each other, but what if they conflict? Where does contract law stand on this question?
A key concept in contract law is that of the parol evidence rule. This rule addresses the validity of past written or oral statements (parol evidence) with subsequent statements. In general, it states that when parties form an agreement in writing and express the fact that it is the final expression of their agreement and that agreement is unambiguous, no parol evidence may be submitted to supplement, explain, or contradict it. Where this expression of finality is not present or where there is ambiguity that needs to be resolved, consistent -- but not contradictory -- parol evidence may be examined to supplement or explain those parts not completely expressed.
This type of concept (as discussed in §3.4.2) is extremely relevant to metadata statements. Can one be assured that one has the complete set of semantics before assenting? How would one negate previous semantics? Schema and protocol designs must be very clear in addressing these questions. A technical term often used in specifications is normative. Normative sections and terms govern the exclusive and complete semantics of a specification. Other parts of a specification might be useful in providing the purpose and context of the specification, but they are not necessary. This other text is similar to why the purposes of a contract are often explicitly spelled out -- all the "whereas" clauses. This other non-binding text provides additional context to the comprehension of the normative sections.
P3P, as an application of metadata, was designed to directly address questions of semantic discovery, ambiguity and conflict. However, it is not unique. Resource discovery (finding all relevant semantics) and expiration (when semantics are no longer valid) are central to protocol design. As stated earlier, the P3P specification requires an agent to be notified of all existing semantics (even if they are found in multiple proposals) by the service. The service does this by sending a hash of proposals (propID) that apply. If a service no longer wishes to operate under a past agreement, it simply makes no solicitations, nor accepts any information, under that propID.
In general, the best solution to problems of discovery and conflict is that of forethought. Contracts that clearly state that they are the complete and exclusive manifestation of the agreement (through a "merger" or "integration" clause) are less prone to problems -- a feature metadata technologies are readily capable of providing. P3P is defined as such at a number of levels. The P3P specification requires the presentation of the propIDs by the service to be the exclusive and complete set of semantics pertaining to the collection of the specified information. This establishes the contract principle of integration of all the terms of agreement into the propID. In addition, the various data collection statements within a P3P proposal would be understood in accordance with the judicial maxim for the interpretation of contracts: "the enumeration some excludes others" (known to lawyers by the latin expression: 'Expressio Unious Est Exclusio Alterius'). Basically, that which is not stated is not done. (If you do not say you collect information, then you are not allowed to do so.) However, this semantic is carefully qualified in that it applies to information related to Web interactions and the data elements specified within the core specification. A P3P proposal does not include the semantics of interactions a user has with a service over different media, such as telephone or mail. (Joel Reidenberg points out an interesting contrast to the Expressio principle as applied to risk a user carries in an interaction. Expressio promotes complete disclosure by a Web. The concept of caveat emptor, "let the buyer beware", is a default rule of warranties that allocates the risk resulting from a product's uncertain quality to the buyer/user.)
As discussed in §2.5, the XML namespace mechanism can prevent semantic conflicts in the schema definitions, such that one cannot confuse contrary schema semantics -- one can supplement or extend. Conflicts between multiple proposals, even when of the same schema, must be resolved by the application. In law, this act is known as interpretation. In the short term, computer agents will not likely be very sophisticated in their ability to interpret multiple assertions to derive the complete set of semantics. Consequently, if an agent believes there is any chance of confusion, it should reject the proposals and inform the user.
Additionally, metadata technologies can easily represent if metadata is final, complete, or if other terms are found elsewhere. In contract law, the presence of a "merger" or "integration" clause is used to specify these qualities of a contract. Communicating the completeness of a statement is the purpose of mandatory/optional protocol extensions as described in PICS. Also, it is trivial to reference other terms and semantics: to say, "You can find other relevant terms at "http://www.foo.com/bar.html" because everything appropriate would be on the net and referenced by a URI! However, while pointing to other terms is not difficult, protocol designers are always cognizant of the need to limit the amount of wandering a network application needs to do to complete its task.
Finally, our scenario of using both metadata and natural language semantics is problematic. However, as already stated, the comparison of a schema definition and its application to the natural language expression is something an informed person is capable of doing. Regardless, the P3P specification states that the propIDs accompanying an interaction represent the complete/integrated semantics of that interaction. A future email from a service that states the terms of a past agreement have changed and allows the user to opt-out of this change is not normative, meaning that a service that acts on his revised view of the agreement is abusing the contract. For an analysis of "the Battle of the Forms" outside the context of metadata mechanisms see [Efr97:1328-1350]. Effross discusses the UCC's text on which terms in differing versions of a contract are enforced or knocked out because of (1) the role of the party, (2) the type of offer (i.e., offer to buy or offer to sell) and (3) the requirements expressed for acceptance within the contract itself (e.g., "a response to this offer of contract formation is an express assent to the terms within this contract and no others.")
When courts interpret a contract they often look to what each party knew or was likely to know (both themselves and of the other party); as well as who drafted the terms of the contract. This analysis does not quite fit our metadata scenarios. In the privacy application there are two other important parties -- drafters of the contract if you will. The schema designer crafted the terms used in the creation of agreement, and a GUI designer's product most likely represented those terms to the user.
One need not know if the parties knew, or were likely to know, if a cow was barren or not. Instead one must know if the either party properly understood the GUI, that was predicated on the GUI designers understanding of the schema, that was predicated on the clarity of the schema design!
For instance, the redistribution element of a privacy schema might permit two values, (1) data is used only by the service, (2) data is redistributed to third parties. A service that gives information to a shipping company so as to fulfill a purchase order may consider this third party to be an agent and will classify his redistribution practice as type (1). A user may not share that understanding, and might be shocked to find that the shipping company also used such data for their own marketing purposes! This very example is probably one of the most difficult issues designers of the P3P Harmonized Vocabulary had to address. One must avoid ambiguity while avoiding schema so complex so as to be meaningless to the average user. How does one resolve this? Be as specific as possible and take great care with the crafting of the text. Otherwise, we will have to rely upon the consensus processes and schema ecologies mentioned in §3.3 to provide us the best designs possible.
How does one resolve semantic ambiguity introduced by these hidden drafters of the contract? The only parallel in contract law that I can think of is the law itself. Recall that the UCC defines contractual boilerplates and default terms and conditions for others to use. If there is ambiguity in the UCC itself, how is it resolved? Recall the UCC is implemented by the states. Consequently, courts could be called on to rule on their reading of the laws, legislators may clarify it through additional law, or the drafters of the UCC may revisit the issue. The process in schema design is not that different. The schema and GUI designers may be called upon to provide the context for interpretation or provide a better design. (They might even be held accountable for not providing a better design in the first place.)
Now, the difficult question! When the user starts getting junk mail and sues the service for breach of contract because the shipping company used her shipping information for marketing purposes, how will the court remedy the situation?
I suspect a court would not find the situation all that novel. A court would ask questions about the purpose of the interaction:
Also, the court might ask questions about
In addition to interpreting terms and assigning them their normative meaning, courts are also called upon to provide omitted terms. Consequently, one approach to schema design is to be cognizant of this very fact. Wherever there is likely to be confusion or contention in the definition of a term, a buzzword such as "reasonable" may be inserted, which asks a future court to apply its own semantic to the definition. This was encouraged by some involved in the design of the P3P Harmonized Vocabulary, but was largely avoided when possible. I believe it is the schema designer's obligation to be as definitive as possible, and use subjective or relative standards lightly since the world cannot wait for courts to rule before using the schema.
Finally -- while slightly out of scope -- I wish to address a pragmatic question in the privacy example. If a court finds in favor of a user, did the agreement provide them with adequate means of protection and remedy? Not likely.
The present approach of contract law to remedy is one of economic efficiency. The plaintiff should be compensated such that they are in a position similar to the one they would've been in had the contract had not been breached. (Restatement 2nd §347). Consequently, a breach of contract is said to be efficient when the cost to perform exceeds the benefit to both parties resulting from its performance. In such a case, the plaintiff is compensated appropriately and the defendant decreases her costs.
What could be wrong with this? When one considers this rationale, the relatively small value of exchanges in electronic commerce, the difficulty of determining the expected damages (particularly in the privacy example), and the transaction cost of resorting to a court for remedy, it is clear that it will often be in the interest of a service to breach a contract. The rationale of economic efficiency does not seem to act as a good disincentive against breaking agreements when applied within the scope of most Web interactions. However, there are a couple remedies. First, as part of the agreement, a clause can define in advance the damages due in event of a breach (liquidated damages). Second, legislatures and regulatory authorities may wish to establish penalties such that parties are disincented to break agreements. Unfortunately, neither of these approaches seem politically feasible in the present day context of U.S. privacy protection, and in general seem counter to the direction of present day policy on computer agreements. Consequently, the rather weak remedies left to users are perhaps class action suits and public pressure. For instance, Intel became the focus of recent attention because the inclusion and default activation of a unique identifier in every one of its new Pentium III chips. This would be used to identify users in electronic commerce and other net-based applications. Public pressure forced them to discontinue the default activation of this feature.
Elsewhere [Rea99] I use a quote from the film The Princess Bride to motivate an exploration of computer agents and legal agency: "You keep using that word, I do not think it means, what you think it means." I was speaking of confusion resulting from a concept I define in this paper, namely syntactic interoperability with semantic fault. Protocol designers and lawyers frequently use the term "agent" in the context of electronic commerce, and while the syntax is correct, the semantics behind each usage does not necessarily match. One needs to qualify each usage with a legal or technical namespace!
Otherwise we end up with confusion. Without a clear definition of semantics, we are not speaking of the same thing. If we are not speaking of the same thing, the nature of our relationship is going to be less than satisfactory. This applies as much to cyberspace as it does to real space. In fact, cyberspace might have something to offer us that we don't have otherwise: well-specified and explicit representations that our computers can help us make sense of!
This is not to say there will never be confusion. There will be a confusion while we gain experience in schema and user interface design, as we come to understand what is feasible, fair, and beneficial to the conduct or our lives online. This later process will be as political as it is technical. However, I believe will be better served by this process of computer aided decision making and schema design than its alternative. The subject of this paper permits user choice, flexibility, and transparency in the context of common law and contemporary judicial principles. The present day alternative is to permit special interests to define a single set of terms, to set the defaults, and to bias the standards of interpration and liability such that the user always loses.
[AG] Ayres, Gertner.Filling Gaps in Incomplete Contracts: An Economic Theory of Default Rules., Yale Law Journal Company. Yale Law Journal, October, 1989, 99 Yale L.J. 87
[FOLDC] Denis Howe (ed.) FOLDC: Computing Dictionary 1993, 1998. Links are provided to some technical definitions by way of the Hypertext Webster Gateway.
[Ber96] Tim Berners-Lee. Keynote Address, Seybold San Francisco, February 1996. Available at http://www.w3.org/Talks/9602seybold/slide6.htm
[BC98] Tim Berners-Lee, Dan Connolly. Web Architecture: Extensible Languages. W3C NOTE: 10 Februray 1998.
[CKR98] Dan Connolly, Rohit Khare, and Adam Rifkin. The Evolution of Web Documents: The Ascent of XML, W3J special Issue on XML, Vol 2, Number 4, Fall 1997, Pages 119-128
[CR97] Cranor and Reagle. "Designing a Social Protocol: Lessons Learned from the Platform for Privacy Preferences Project" Telecommunications Policy Research Conference. Sept. 97.
[Efr97] Effross. "The Legal Architecture of Virtual Stores: World Wide Web Sites and the Uniform Commercial Code." San Diego Law Review. Vol 34, Number 3. Spring 1997.
[KGK86] Kessler, Gilmore, and Kronman. Contracts: Cases and Materials 984 (3d ed. 1986).
[Las98] Lassila, O. Introduction to RDF Metadata. W3C NOTE: 13 November 1997.
[Mur90] J. E. Murray, Murray on Contracts § 38, at 92 (3d ed. 1990)
[BHL99] Tim Bray, Dave Hollander, Andrew Layman. Namespaces in XML (14 January 1999). http://www.w3.org/TR/REC-xml-names
[PICS96] Miller J., Resnick P. and Singer D. PICS 1.1 Rating Services and Rating Systems -- and Their Machine Readable Descriptions. (31 October 1996). http://www.w3.org/TR/REC-PICS-services
[RDF99] Lassila O., and Swick R. Resource Description Framework (RDF) Model and Syntax Specification. (22 February 1999). http://www.w3.org/TR/REC-rdf-syntax/
[Rea97] Reagle, J. (Editor/Chair) Harmonized Vocabulary Working Group Charter.
[Rea98] Reagle, J . Social Protocols: An Introduction. Lecture Notes for The Law of Cyberspace -- Social Protocols. Harvard Law School and MIT. September 1998.
[Rea99] Reagle, J . Agent: I Dont Think It Means, What You Think It Means. (19990524) LawTech'99.
[Who40] Whorf, B. L. (1940): 'Science and Linguistics', Technology Review 42(6): 229-31, 247-8. Also in B. L. Whorf (1956): Language, Thought and Reality (ed. J. B. Carroll). Cambridge, MA: MIT Press. [While not explicitly characterized or analyzed within the reference, this text, writings by Sapir, and resulting debate are the basis of a theory now known as the Sapir-Whorf hypothesis. This hypothesis posits that language affects the way we think. Note, this is different -- though related -- to my more obvious generalization that reality/context affects language. Also see the following URL on the contention associated with "Eskimo Snow" : http://www.sfs.nphil.uni-tuebingen.de/linguist/issues/5/5-1239.html ]
Michigan Law Review, March, 1992, 90 Mich. L. Rev. 1145, 20448. Legal Responses to Commercial Transactions Employing Novel Communications Media., John Robinson Thomas