1. Introduction

Many philosophers of biology accept that there is a well-understood, causal account of information defined in terms of systematic causal dependence. This account is thought to be derived from information theory and adapted by Dretske. It is noteworthy because of its role in supporting the parity thesis—the position that information cannot be used to differentiate between the role of genes and the environment in development. In turn, the parity thesis is significant because it counters the idea that genes are special in development because only they play a distinctive informational role—a view thought to lend support to genetic determinism.

I argue that, despite its utility in resisting genetic determinism, the acceptance of the causal account of information by philosophers of biology is a mistake. The causal account and the arguments based on it are deeply flawed—hampered by significant misreadings and conflations between distinct accounts of information. The philosophical arguments that use the casual account of information, and specifically the argument for the parity thesis, beg the relevant philosophical questions by illicitly reducing necessary informational elements into merely causal ones. As such, we need to reject the causal account of information and rethink the arguments based on it.

My plan in this article is to begin by setting out the view of causal information as systematic causal dependence and showing that it is widely accepted by philosophers of biology. Section 3 introduces what I will argue is the proper reading of the mathematical account of information. In section 4, I argue that causal information defined as ‘systematic causal dependence’ cannot be read as a version of the mathematical account. I introduce a modified version of the causal account of information in section 5. In section 6, I discuss Dretske's adaptations of the mathematical account. I argue in section 7 that neither one of Dretske's adaptations of the mathematical account can save causal information. Finally, in section 8, I consider what an accurate application of Dretske's account of information would look like when applied to the parity thesis argument and argue that the difficulty of applying informational concepts in biological contexts forces a reexamination of informational arguments in the philosophy of biology.

2. The Causal Account of Information and the Parity Thesis

The causal account of information is first introduced in the philosophy literature in Griffiths and Gray (Reference Griffiths and Gray1994). Drawing on earlier work in the developmental psychology literature (Johnston Reference Johnston1987), Griffiths and Gray articulate and defend the claim that the well-understood sense of mathematical information cannot be used to make a principled distinction between the roles of genes and the environment in development.

The core idea in their description of a mathematical account of information is the idea of systematic dependence. Something carries mathematical information about something if the former systematically depends on the latter. Thus, we can define causal information as follows:

This account has gained broad acceptance (Sterelny and Griffiths Reference Sterelny and Griffiths1999; Godfrey-Smith Reference Godfrey-Smith2000a; Maynard Smith Reference Maynard Smith2000a, Reference Maynard Smith2000b; Griffiths Reference Griffiths2001; Downes Reference Dretske2006; Godfrey-Smith and Sterelny Reference Godfrey-Smith, Sterelny and Zalta2007). Griffiths and Gray use the causal account of information to argue that an inherent symmetry in this account of information makes it impossible to apply it only to the role genes play in development. The mathematical theory of information breaks the world down into sources, receivers, and channel conditions linking the two. Griffiths and Gray argue that the symmetry between the source and channel conditions, that is, the fact that you can treat a source as the channel conditions and the channel conditions as a source, rules out applying information to only the role that genes play in development. One can label genes as the informational source and the environment as the channel conditions, but someone else can always invert this designation and label the environment as the informational source and the genes as the channel conditions. They set out a claim that becomes the received view in the philosophy of biology—under the mathematical account of information, both genes and environments carry developmental information.

The philosophical consensus holds that this account of information explicitly traces its history to the information theory developed by Shannon (Reference Shannon1948) and later “sharpened up philosophically by Dretske (Reference Downes, Pfeiffer and Sarkar1981)” (Godfrey-Smith Reference Godfrey-Smith2000a, 202). According to the standard view, causal information is derived from Shannon's work and made more precise by Dretske. This view is not confined to the philosophers of biology; biologists writing in the philosophy-of-science literature also endorse roughly the same account of information (Maynard Smith Reference Maynard Smith2000a, Reference Maynard Smith2000b).

The consensus on causal account of information is not complete. Winnie (Reference Winnie2000) offers a critique of Griffiths and Gray (Reference Griffiths and Gray1994) that highlights significant problems with their causal conception of information and in their original argument. In addition, Sarkar (Reference Sarkar and Sarkar1996a, Reference Sarkar1996b) clearly differentiates and correctly distinguishes between the mathematical account of information and other related informational notions. Bergstrom and Rosvall (Reference Bergstrom and Rosvall2009) also take causal information to task as an interpretation of mathematical information, although they focus on the failure of causal information to capture the aim behind Shannon information—how to engineer the efficient transmission of information.Footnote ¹ But, despite these countervailing views, much of the philosophical community still accepts what I will argue is a mistaken view.

3. The Mathematical Account of Information

The causal account of information is fundamentally confused; it fails to capture any of the relevant senses of information set out by information theory or by Dretske. It neglects the core idea of mathematical information and conflates three distinct senses of information—the sense found in information theory and two related notions developed in Dretske (Reference Downes, Pfeiffer and Sarkar1981). Specifically, causal information conflates Shannon's mathematical quantitative account (MQ information), Dretske's philosophical quantitative account (PQ information), and Dretske's philosophical account of information content (PC information). It is important to keep these accounts distinct from the account offered in Griffiths and Gray (Reference Griffiths and Gray1994) and developed in Griffiths and Knight (Reference Griffiths and Knight1998) and Griffiths (Reference Griffiths2001). I will refer to this account as the causal account of information content (CC information). As a result of misrepresentations and conflations between these accounts, the CC account of information is deeply confused. A clear understanding of each of these distinct accounts will help clarify the key issues in the debate on biological information. We begin with the mathematical account of information as used in information theory.

The MQ account of information grew out of the engineering work of Claude Shannon. In Shannon's account, we start with a basic model of communication that consists of sources, receivers, and a communication channel connecting them. Sources initiate messages, and receivers receive them. Picture a telegram message traveling between two telegraph stations connected by a wire. The source or sender is the originator of the message, while the receiver consists of the telegraph operator on the decoding end. The wire itself and any local interference make up the communication channel.

The amount of information in a message is determined by the amount of uncertainty reduced by a message sent from the source to the receiver. For example, compare three telegraph operators trying to warn their receivers of an impending invasion. One operator has two available options—one dash for an invasion by land and two for an invasion by sea. Each message received will reduce uncertainty by communicating that one of the two possible outcomes has occurred. The idea is that a one-dash signal reduces uncertainty because the receiver knows that one out of two possible outcomes has occurred. The uncertainty about the possible invasion routes is reduced by the communication of the message. Information is measured in bits calculated with a base 2 logarithm. For example, a message that reduces uncertainty from two possible states to one actual state, as in this case, generates one bit of information since log₂(2) = 1.

Another operator can send four messages; one dash for a land invasion, two for a sea invasion, three for an airborne attack, and four for no attack. A one-dash signal indicates that one out of four possible outcomes is occurring. Since the receiver can rule out three possible outcomes on the basis of a one-dash message, more uncertainty is reduced and two bits of information are generated since log₂(4) = 2. In this case, each message received will reduce more uncertainty and thus contain more information.

Imagine that a third operator has 100 potential signals to choose from, and each signal indicates a different state of affairs; that is, one dash means that the first division invades by land, two dashes means that the second division invades by land, three dashes means that the first division invades by sea, and so on. The same one-dash message now picks out one state of affairs out of a hundred and thereby provides more information because it reduces uncertainty by even more. In this case, the receipt of a message produces roughly 6.6 bits of information since log₂(100) = 6.643856189775.

The key element of the mathematical theory of information is the quantitative measure of the average amount of uncertainty reduced by the transmission and receipt of a message. Essentially, it is a measure of the degree of freedom one has in the selection of possible messages. Greater freedom of choice in messages yields more information in a system. If we assume that each message picks out a particular state of affairs, a system that can transmit one of a hundred different messages has a greater capacity to reduce uncertainty than one that can only transmit one of two messages.

It is important to distinguish between how much information can be associated with a given state of affairs and how much information can be transmitted about that state of affairs by a given message. It is not the case that every message is as probable as the underlying state of affairs it represents. For example, we could have a communication system that has 51 possible states and 100 possible messages, each picked out by a number between 1 and 100. Every even number could pick out a distinct state of affairs, while every odd number could pick out one single state of affairs. In this article, I will assume for the sake of simplicity that each message in a communication system picks out one and only one distinct state of affairs.

One important feature of this definition of information is that mathematical information is a measure of the average uncertainty reduced across an information channel and not of the content of any particular message. While the examples above are discussed in terms of the information content of specific messages, the engineers who developed this account of information were interested in measuring the average of all possible messages carried over a communication channel. The MQ account of information was developed to help electrical engineers design efficient communication systems. As such, they were interested in the ‘carrying capacity’ of different communication systems, not in the amount of information transmitted by a particular message. The MQ account of information is a measure of this carrying capacity. Thus,

For example, if we return to our telegraph examples above, the coding systems are set up in a way in which each possible message carries the same amount of information, making the average amount of information no different from the amount carried by each message. Not all coding systems are set up this way. Some systems are built to be redundant and reserve many messages for one state of affairs. In these systems, the average uncertainty reduced by any particular message varies. The important point is that mathematical information is a measure of the average amount of uncertainty reduced by all possible messages in a coding system.

It is important to be clear about what is measured by the MQ account of information. This account measures the average degree of freedom of the coding system. In the examples above, it measures the amount of information carried by the conventions of differing communication systems. In a system in which there are only two possible messages corresponding to two states of affairs, the receipt of a message carries less information than a system with expanded communication conventions. It should be clear that just knowing that two telegraph stations are causally connected and that one station received a one-dash message from the other does not tell someone how much information, if any, is carried by that message. It is impossible to measure the carrying capacity of that causal interaction without knowing the relevant communication system and particularly the set of possible messages at the sender. Generally, the quantity of information is determined by the degree of freedom with respect to possible messages, not by simple causal correlation.

4. CC Information ≠ MQ Information

Griffiths and Gray's ‘systematic dependence’ characterization of biological information does not capture the two key features of the MQ account of information. To recap,

First, the ‘systematic dependence’ account is an account of information content, not a quantitative measure of the reduction of uncertainty as one should expect with the MQ account. A quantitative measure would describe information as a quantity, but Griffiths and Gray characterize information as about developmental outcomes. This reveals that they are thinking of information as having content and not as a quantity. If they were thinking of information as a quantity, they would ask how much information is carried by genes and environments.

An accurate discussion of the MQ account of information will refer to information as a quantity. The key idea behind the Shannon account of information is that information is to be understood as a quantity and not in terms of meaning or content.Footnote ² Sterelny and Griffiths (Reference Sterelny and Griffiths1999) and Griffiths (Reference Griffiths2001) acknowledge the quantitative nature of the mathematical account of information. The problem is that they fail to also acknowledge that this point reveals that the Griffiths and Gray (Reference Griffiths and Gray1994) version of this argument fundamentally mischaracterizes the MQ account of information. The lack of any mention of quantity in their account means that there is no way to read the argument in Griffiths and Gray (Reference Griffiths and Gray1994) as about information quantity and not information content.

Second, the ‘systematic dependence’ characterization does not mention the idea that information involves the reduction of uncertainty. Not only do Griffiths and Gray (Reference Griffiths and Gray1994) mistake a quantitative concept for a qualitative one, they also distort the basic meaning of the account. For the mathematical account of information, systematic correlations are only important insofar as they bear on the relevant reduction of uncertainties. Systematic dependencies that do not involve the reduction of uncertainty are not informational under the mathematical account. The symmetry between sources and channel conditions is irrelevant if Griffiths and Gray have fundamentally mischaracterized the MQ account of information. As a result, the central argument for the parity thesis fails as stated. The problem is not that it is impossible to apply the mechanisms of information theory to environments. Environments can legitimately act as potential information sources; that is, messages from the environment could reduce uncertainty.Footnote ³ The problem is that Griffiths and Gray fail to accurately characterize the mathematical account of information, which undermines their claim that it is impossible to use it to draw informational distinctions between the role of genes and environments in development.

To characterize mathematical information as such is simply inaccurate and is problematic for Griffiths's argument. This distortion allows Griffiths to claim that there is symmetry between genes and environments concerning mathematical information because both are causally connected developmental outcomes. Griffiths uses evidence that there is a causal connection between environments and developmental outcomes as evidence that mathematical information applies, but this relies on a misreading of the mathematical theory of information.

5. The Modified Parity Argument

Sterelny and Griffiths (Reference Sterelny and Griffiths1999) and Griffiths (Reference Griffiths2001) invite us to read the argument in terms of Dretske's account of information (perhaps recognizing some of the problems with the first formulation of the causal account). Can this account of information rehabilitate the core parity argument? Griffiths and Knight (Reference Griffiths and Knight1998) reiterate the view that the entire set of developmental resources is properly thought of as the source of developmental information from Griffiths and Gray (Reference Griffiths and Gray1994) and quoted above. However, Sterelny and Griffiths (Reference Sterelny and Griffiths1999) reintroduce the same argument with a significant change in definition and emphasis. They say, “The idea of information as systematic causal dependence can be used to explain how genes convey developmental information. The genome is the signal and the rest of the developmental matrix provides channel conditions under which the life cycle of the organism contains (receives) information about the genome” (102).

Here, the earlier ‘systematic dependence’ is now ‘systematic causal dependence’. This passage marks two significant changes in the argument. First, the emphasis is not on the MQ account of information proper but on Dretske's (Reference Downes, Pfeiffer and Sarkar1981) philosophical treatment of this theory. Second, CC information is now defined in terms of systematic causal dependence and not mere systematic dependence. This shift marks a change in definition; previously, correlations alone carried information. It this version, correlations are too weak; now only causal dependencies carry information. The basic argument remains the same. Both genes and environments carry developmental information when understood as systematic causal dependence.

The move away from information theory proper to Dretske's account of information is clearest in Sterelny and Griffiths (Reference Sterelny and Griffiths1999) and Griffiths (Reference Griffiths2001). Both works explicitly distinguish between the mathematical theory of information and Dretske's account. Griffiths says,

Here, Griffiths properly differentiates between the quantitative mathematical account and Dretske's qualitative account of information. Griffiths uses Dretske's specialized account of informational content to bolster his argument that information applies to both genes and environments in development. The basic structure and point of the argument remains the same. Griffiths nicely sums up his key point about parity and causal information. He says, “So far as causal information goes, every resource whose state affects development is a source of developmental information” (Reference Griffiths2001, 398). Although the description has changed from the mathematical account of information proper to Dretske's modified account, the key inference remains the same. Since information is a matter of systematic causal dependence between states, any developmental resource whose states affect development carries information.

6. Dretske's Adaptations of MQ Information

Dretske (Reference Downes, Pfeiffer and Sarkar1981) develops the fullest philosophical treatment of the mathematical theory of communication. I dub this account the PQ account to highlight the fact that this is a philosophical reworking of the standard MQ account of information. Recent versions of the causal account explicitly cite this work, so it should help clarify the parity argument. Unfortunately, there is ambiguity about what aspect of Dretske's account is relevant in this context because Dretske provides both a subtle reinterpretation of the orthodox account and a substantially different account of information content. Godfrey-Smith describes the mathematical account as “sharpened up philosophically by Dretske (Reference Downes, Pfeiffer and Sarkar1981)” (Reference Godfrey-Smith2000a, 202). It is unclear whether he refers to Dretske's adaptation of the MQ account of information for philosophical usage or to Dretske's substantially different account of information content. This section focuses on Dretske's reinterpretation of the standard MQ account of information, the PQ account. The next section will discuss Dretske's account of information content—which I label the PC account—that highlights the fact that it is both a philosophical account and an account of information content as opposed to the other quantitative (MQ and PQ) accounts.

Dretske outlines a modified version of the mathematical theory of communication that is crafted to better suit the needs of the philosopher and less suited for the communication engineer. Specifically, Dretske reframes the mathematical account of information in terms of the content of specific messages and not an average of messages. The average measure is useful for engineers but does not get at what philosophers want to know. Dretske says, “One of the most obvious respects in which communication theory is unprepared (I do not say unequipped) to deal with semantic issues is that meaning, truth, reference, and allied semantic notions apply to particular messages or acts of communication while the theory is preoccupied with average amounts information (entropy)” (Reference Downes, Pfeiffer and Sarkar1981, 47). Philosophical interests dictate that Dretske focus not on average uncertainty but on the actual uncertainty reduced by particular messages.

Dretske identifies the relevant quantities for an account of the amount of information carried by particular messages as the “amount of information generated by a particular state of affairs and the amount of information carried by a particular signal about [the state of affairs]” (Reference Downes, Pfeiffer and Sarkar1981, 52; formulas omitted). Although more philosophically satisfying than the average measure, this account makes it difficult to apply in the real world. Dretske acknowledges this and sets out what is needed to calculate the PQ information in a given state of affairs. He says, “How, for example, do we calculate the amount of information generated by Edith playing tennis? And how much information about her activities is contained in the light reaching Herman—a casual observer? In order to answer questions like this … one needs to know: (1) the alternative possibilities (e.g., Edith's eating lunch, Edith's showering); (2) the associated probabilities of all these alternate possibilities; (3) the conditional probabilities (given the configuration of photons reaching Herman's visual receptors) of each of these alternate possibilities” (53). Sometimes the precise informational content will be impossible to calculate because not all of these conditions are met. Still, even in cases of imperfect information, comparative judgments can be made. Thus,

The key idea in Dretske's independent account of qualitative information content (PC information) is that one can identify the content of information by associating a signal with the particular state of affairs it indicates at a source. He defines his account of information content in terms of a conditional probability. He says, “a signal r carries the information that s is F = the conditional probability of s's being F, given r (and k), is 1 (but, given k alone, less than one)” (Reference Downes, Pfeiffer and Sarkar1981, 65). Dretske uses the expression “s is F” to represent a particular state of affairs, namely, that state s has property F. The term k refers to background knowledge. In plain English, Dretske's definition can be paraphrased as “a signal carries information about a state of affairs if the conditional probability of that state of affairs having a particular property is 1, given the signal and background knowledge, but is less than 1, given just the background knowledge.” For example, one lantern in the old North Church carries the information that the British are invading by land if the conditional probability of the British invading by land is 1, given the lantern in the window of the old North Church and relevant background knowledge, and less than 1, with just this background knowledge alone.

The relevance of background knowledge is an important element of Dretske's account. In the context of discussing this aspect of his account, he says, “This constitutes a relativization of the information contained in a signal because how much information a signal contains, and hence what information it carries, depends on what the potential receiver already knows about the various possibilities that exist at the source” (Reference Downes, Pfeiffer and Sarkar1981, 79). This point is relevant for two reasons. First, and most obviously, it identifies an important contextual feature of Dretske's account. The informational content of a signal is always relative to some body of background knowledge. Specifically, it is relative to knowledge about the possible variations in the state of the source. Recall that Dretske's account of information content is built on his account of information quantity. As such, the knowledge of the relevant possibilities is necessary for determining the information content. Second, it highlights the continued significance of the existence of differing possibilities at the source. Dretske's definition of information content assumes that the source has a positive amount of information in the PQ sense. This means that there must be genuine possible alternatives to any given state at the source.

The causal account of information frequently conflates three distinct ideas: the mathematical account of information (MQ information), Dretske's philosophical interpretation of mathematical information (PQ information). and Dretske's account of information content (PC information). To review,

The common understanding of the causal account does not accurately capture any of these concepts. Even papers that correctly distinguish between the mathematical account of information and Dretske's account of information content still neglect to address the explicitly information-theoretic aspects drawn from Shannon's model and the context-sensitive aspects of Dretske's own model.

7. CC Information ≠ either PQ or PC Information

The account of information offered in the modified parity argument is no closer to Dretske's PC or PQ accounts of information than the unmodified mathematical (MQ) account. With respect to determining the information content of particular messages, the causal account does not match Dretske's modification of the actual mathematical account of information. Any application of Dretske's PQ account of information fails in the same way as the MQ account. A purely quantitative account of information does not capture what Griffiths thinks the biologist means by information. Dretske's philosophical modification of the mathematical account only turns an average measure to a measure of individual messages. It still remains a purely quantitative measure and as such does not work in the biological context because it is merely a measure of the amount of information associated with a state of affairs and never a measure of information content.

In addition, Dretske's conditions highlight what is missing from Griffiths's argument—a discussion of the relevant alternate possibilities and the related probabilities/conditional probabilities associated with a given state of affairs. Since the mathematical account of information is defined in terms of the reduction of uncertainty associated with states of affairs, Griffiths needs to frame his parity argument in these terms, not in merely causal ones.

These features are absent in Griffiths's (Reference Griffiths2001) application of Dretske's account. Although the paper offers a reasonably accurate description of Dretske's account of information content, it fails to include these features in the actual argument. Recall that the basic argument in Griffiths (Reference Griffiths2001) remains the same as earlier versions. For Griffiths, information is a matter of causal dependence, and both genes and environments carry information in the causal sense since they both share causal dependencies with developmental outcomes. The characterization of Dretske's account of content as dependent on mere causality is a misreading. There is more than mere causal dependency in his account of content. The specifically informational elements, like the reduction of uncertainty, the existence of variant states of the source, and the requirement of some sort of knowledge of relevant possibilities at the source, cannot be eliminated.

None of the reviewed accounts of information are properly characterized in purely causal terms. The key idea in each is the reduction of uncertainty. They are largely mute on the significance of causality. It is possible to have a causal connection with no MQ, PQ, or PC information. Generally, information is carried by causal signals, but the concepts are distinct. There is a whole class of systematic dependencies (invariant dependencies) that do not carry information in the MQ, PQ, and PC senses of information because they lack alternative possibilities at the source. Griffiths's argument overlooks this class of systematic causal dependencies that by definition cannot carry information. For example, persistent environmental features like gravity causally affect developmental outcomes but should not be thought of as carrying information in the MQ, PQ, and PC senses. The problem is that since there are no alternatives to the causal impact of gravity, it is impossible to view its effect as a reduction of uncertainty. Without a possible set of alternate messages, the Shannon information framework does not make sense. Recall that it is impossible to determine the quantity of information on the basis of a causal connection between telegraphy stations in the absence of a communication convention. A simple causal connection does not necessarily carry information in any of the existing accounts of information when properly understood.

There is an explicitly causal element in Dretske's PC account of information, but this is a necessary, not a sufficient, condition. Dretske says, “The transmission of information requires, not simply a set of de facto correlations, but a network of nomic dependencies between the condition at the source and the properties of the signal. The conditional properties used to compute noise, equivocation, and the amount of transmitted information (and therefore the conditional probabilities defining the informational content of the signal) are all determined by the lawful relations that exist between source and signal” (Reference Downes, Pfeiffer and Sarkar1981, 76–77). Clearly, causality, or as Dretske puts it, lawful relations, is a necessary condition for a signal to carry information. But it is a mistake to suggest that any causal connection, whether systematic or not, carries information in Dretske's sense. The mere presence of a causal connection is not a sufficient condition for carrying information. The passage above specifies that the explicitly informational items (noise, equivocation, and amount of transmitted information) are the key elements in Dretske's account. Merely knowing that two things interact causally does not give one grounds to conclude that there is an informational connection between them. This point undermines the key inference in Griffiths's argument.

A critical assessment of the arguments in Griffiths and Gray (Reference Griffiths and Gray1994), on one hand, and Sterelny and Griffiths (Reference Sterelny and Griffiths1999) and Griffiths (Reference Griffiths2001), on the other, begins with a clear understanding of how the two versions of the argument differ. The basic structure and conclusion of the arguments remain the same, but the description of the key term changes. In the most recent version, causal information refers to Dretske's account of information content and not the actual mathematical theory of information. This redefinition has not been widely noted in the philosophy-of-biology literature; most authors appear to accept the conflation between these distinct accounts of information and Griffiths's argument.

8. Can PC Information Save the Parity Thesis?

Is there any way for Griffiths to salvage his core argument? Griffiths's argument is based on fundamental misreadings of the mathematical account of information and Dretske's interpretations of it. Because of this, as stated, the arguments are all unsound. It is simply false that the MQ account of information or the PQ account or Dretske's PC account of information content defines information in terms of mere causal dependency. I see no way to defend the interpretations of the various accounts of information in the argument. Griffiths needs to have a causal account of information to make his argument work. Without it, he cannot make the inference from both environmental and genetic elements having a causal role to them both having an informational role, but this causal account does not match the standard accounts of information. Griffiths could save the core argument by going back and offering up a detailed application of one of the relevant accounts of information in enough detail to address the explicitly informational aspects of the accounts. Specifically, he needs to show that, if genes can be shown to have PC information, for example, then environments can also provide the appropriate kind of information.

Griffiths is free to define an independent causal account of information without reference to either the mathematical theory of information or Dretske's account, but this move will not salvage his position. The problem with this move is that it begs the question. Everyone acknowledges that genes and environments have a causal role in development. The question is whether they also have an informational role. Simply defining an informational role as a causal role without some sort of independent justification for this definition is question begging. It would merely establish the relevant conclusion by definition. If we are to assume that information is a matter of causality, then it follows as a matter of course that both genes and environments have an informational role. The argument purports to avoid this problem by using well-known preexisting accounts of information, but if the reading of these accounts is flawed, the entire argument becomes trivial without independent justification of a causal account of information.

What would PC information look like when applied in an appropriate biological context? Since much of the debate focuses on the question of whether genes and environments carry information, I will apply Dretske's PC account of information to a canonical instance of biological information—protein synthesis.Footnote ⁴ The first step would be to identify the source, channel, and receiver. Second, one would specify the possible variations in the source, that is, set out the variations in states of affairs at the source that can be transmitted over the communication channel. One would need to determine, as Dretske specifies, the “alternative possibilities,” the “associated probabilities of all these alternate possibilities,” and the “conditional probabilities of each of these alternate possibilities” (Reference Downes, Pfeiffer and Sarkar1981, 53). In addition, there would need to be some way to address the context sensitivity of the relativization to background knowledge. How is this feature to be understood in a naturalistic setting?

There are a number of possible ways to delineate sources, channels, and receivers. In general terms, genes and environments are sources, while the biological processes of gene expression and background conditions are the channel and proteins the receiver. The general picture becomes a good deal more complicated when the variations in source or the set of possible messages are filled out. There are a number of different ways to parse DNA as a source. Is the set of possible messages determined by the variations of base pairs in DNA sequences? If so, as Winnie (Reference Winnie2000) points out, each DNA sequence of the same length contains the same number of possible messages. The advantage of this approach is that it generates a measurable set of possible messages determined by sequence length, but it fails to match biologists’ distinction between coding and noncoding regions of DNA. But this is hardly the only option. Modifying Godfrey-Smith's (Reference Godfrey-Smith2000b) analysis of ‘coding’, the tripartite sequences that code for amino acids could be considered the set of possible messages, as could the sets of the tripartite sequences that code for the primary structure of amino acids in proteins. There are still more options: as Winnie (Reference Winnie2000) proposes, each stretch of DNA could be parsed into sets of messages on the basis of the complexity of its shortest description. On this account, we would have a set of possible messages ranging from simple to complex, in which the complexity is measured by the length of algorithm necessary to specify the sequence.

It is a good deal more difficult to partition the sets of possible messages with environmental sources. Sequences of DNA are fairly easy to parse as sets of possible messages; many of the environmental factors of interest consist of gradients that are difficult to parse into sets of messages in principled ways. It is possible to partition gradients into discrete units by arbitrarily limiting the units of measurement. For example, a temperature gradient can be converted to discrete units by only using tenths of a degree and none smaller.Footnote ⁵ What is trickier is that not all environmental factors fall under a single type of uniform sourcelike gradient. There is a great deal of heterogeneity in types of environmental cause. A full defense of environmental parity in terms of information needs to set out all of the relevant types of environmental causes and come up with a principled scheme that partitions environmental causes into sets of messages. The clear delineation of the set of possible messages is a necessary starting point if PC information is to save the parity thesis. It is a challenge to even set out a coherent account of the possible set of environmental messages.

Second, if viable genetic and environmental candidates meet the PC information definitions for delineating sources and sets of possible messages, we then need to consider the necessary probabilities. We need to know the probabilities associated with each possible message, that is, the odds of each given message being sent, given the set of possible messages—roughly the likelihood of any particular message being sent, given all possible messages. In the gene case, we would need to know the probability of each genetic message, depending on which of the above attempts were used to partition DNA into discrete messages. If we parse DNA on the basis of the tripartite coding model, we would need to know the relative probability of each coding triplet being sent as a message, given all possible triplets. Each system of parsing messages will require an independent account of that system's relative probabilities. The same holds true for environmental messages. In addition to outlining the set of possible environmental messages, we would also need the probability of each particular message, given the entire set of messages.

Third, we need to know the conditional probabilities of the receiver being in a particular state, given the transmission of each message—that is, probabilities that a particular amino acid or protein is made, given each possible message. Recall that Dretske's PC model of information content required knowledge of the ‘nomic dependencies’ between the conditions at the sources and the properties of the signal. In the biological context, this would require knowledge of the causal connections between the sender of each possible message and its receiver. In the DNA case, this knowledge amounts to a complete picture of the causal pathways from each sequence of source DNA to each resulting receiver protein. Each different set of possible genetic messages (in the senses outlined above) would require different collections of causal knowledge. The same requirements hold for the environmental information cases as well. We would need a complete causal picture connecting the source of each possible message with the resulting protein receivers for each possible way of conceptualizing environmental causes as a set of messages.

Finally, is there a way to understand what could count as a relativization of each set of biological messages in the way Dretske outlines? It is difficult to know how to apply this requirement in the context of biological information. Is there a meaningful way to capture the background knowledge necessary to interpret a message as a message in strictly biological terms? On one hand, this seems to require an explicitly intentional aspect in that it seems to require that biological systems have some sort of knowledge. On the other, a clever operationalization of the relativization of knowledge by the receiver might allow for a simplification of these exceedingly complex PC models of biological information. Simply determining all of the possible sets of possible messages, their relative frequencies, and their nomic dependicies with outcomes is a dauntingly complicated task. It is possible that the possible outcomes could be narrowed by picking the right set of relativizing conditions. Still, it should be clear that actually defending the parity thesis with a principled application of Dretske's PC model of information is a daunting task that would require considerable more elucidation than I have tentatively sketched here. It is unclear whether PC information can save the parity thesis.

9. Conclusion

This article suggests that, for the most part, the current literature in the philosophy of biology does a poor job in clearly explicating accounts of information and applying them in arguments. The parity thesis arguments both conflate and misrepresent the relevant accounts of information. The key lesson is that philosophical applications of concepts borrowed from information theory need to keep a robust account of their informational structure. Philosophical applications of information cannot reduce information to mere causality. The causal misreading of “causal” information underemphasizes the significance of a reduction of possibilities and the context dependence of the informational receiver. Consequently, there is the need and opportunity to reevaluate the utility of these accounts of information as tools to explain information talk in biology. Much of what the philosophy-of-biology community thought was settled is now open for reevaluation.

My focus in this article has been on the use of informational concepts by philosophers of biology, but I hope that it is also clear that these same arguments apply to all those who think that biological entities carry information because they carry Shannon information or some suitably modified version of Shannon information. The claim that biological macromolecules carry information in this sense can only be defended by a careful explication and application of all of the information-theoretic aspects of information theory. We have yet to see an appropriate defense or justification. As such, we should be rightly skeptical of much information talk in biology until we do.