1. Introduction

A 2000 lawsuit against the pharmaceutical giant Pfizer uncovered internal memos documenting the brazen and pervasive use of ghost authors in Pfizer-funded studies of the effects of Sertraline (Zoloft). The memos listed articles in various stages of completion and included items that read, for instance,Footnote ¹

David Healy and Dinah Cattell analyzed the data from the internal memos. They found that well over half of all the articles on Sertraline published between 1998 and 2000 could be identified as ghostwritten on the basis of Pfizer's internal documents. Further more, 100% of the identifiably ghostwritten articles reported positive findings about the drug, whereas 56% of the not-known-to-be-ghostwritten articles reported neutral (7%) or negative (49%) findings. The ghostwritten articles were, on average, far longer and appeared in journals with much higher impact factors (Healy and Cattell Reference Galison, Galison and Biagioli2003). Other ghostwriting exposés followed; for instance, a lawsuit against Merck concerning Vioxx revealed documents listing completed papers with ‘external author?’ next to the titles (Ross et al. Reference Kakkar, Cimminiello, Goldhaber, Parakh, Wang and Bergmann2008; Saul Reference Poordad2008).

It has become clear that a large percentage of the most visible and prestigious biomedical research articles are in fact ghostwritten. Often such research is not merely ghostwritten but ‘ghost managed’: the research questions, trial design, statistical analyses, publication strategy, and writing process are all designed top-down by pharmaceutical company employees, for the purpose of presenting their products as marketable (Sismondo Reference Ross, Hill, Egilman and Krumholz2007, Reference Rudner2009; Sismondo and Doucet Reference Sankar and Kahn2009). In the wake of these revelations, a flurry of articles appeared by bioethicists and philosophers of science decrying ghost authorship as a shocking transgression of the norms of scientific practice (Brody Reference Brody2006; Moffatt and Elliott Reference Huebner, Kukla and Winsberg2007; Sismondo Reference Ross, Hill, Egilman and Krumholz2007, Reference Rudner2009).

Such data do indeed signal a kind of crisis of epistemic security in biomedical research. Healy and Cattell's numbers eliminate any reasonable doubt that such ‘authorless’ research is systematically biased in favor of the profit margins of the companies that pay for and design it. But in this paper, I will argue that we should see ghost authorship as an unsurprising and continuous extension of contemporary scientific norms and practices rather than as an abrupt transgression of them. The entrenched organizational practices and architecture of contemporary biomedical research have rendered much of it authorless in ways that raise serious challenges for social epistemology. Accordingly, no one is in a position to understand or take responsibility for the ways in which interests and nonepistemic values have shaped that research. While heavily managed, deceptively packaged, and systematically biased science funded by private companies may be especially ethically egregious, I will claim that ‘legitimate’ biomedical research that is not so centrally controlled poses epistemic problems that are at least as thorny and interesting.

2. Authorship in Biomedical Research

Traditional scientific authors produce public texts. Authors may not be knowers; one can be a legitimate author and yet come to incorrect conclusions. But a traditional scientific author is the locus of responsibility for the content of her texts. She generates raw data, interprets them, draws conclusions, and represents those conclusions by creating a text and taking responsibility for challenges to its contents or calls for its justification, including retracting its claims if necessary. On the face of it, the use of the term ‘author’ in Pfizer's memo is a cynical misuse; Moffatt and Elliott (Reference Huebner, Kukla and Winsberg2007), for instance, insist that attaching one's name to a completed text, without having played any role in its production or being in a position to vouch for its contents, involves deception that is fundamentally discontinuous with legitimate authorship in science.

But the traditional notion of authorship fits uneasily with how biomedical research and publication actually proceed. Large-scale clinical trials of the sort that make up much of biomedical research are typically radically collaborative, multisite projects involving many teams of researchers, each of which uses nurses, lab techs, students, departmental administrators, and many others to help them implement the study and keep track of data. Papers with dozens of authors, often flung over several continents, are the norm. Usually no one has the expertise and training to design and implement every part of such a study, which typically involves researchers with a wide variety of disciplinary affiliations. Years ago, John Hardwig (Reference Douglas1991) pointed out that many scientific experiments take far more than a lifetime's worth of collective labor hours and disciplinary training and, hence, could not be run by a single researcher no matter how diligent. In the cases I am interested in, the impossibility of a single author recreating the study runs yet deeper: not only could no one person have the time or the expertise to run the study on her own, but the data are time and place dependent. Questions about clinical effects are context specific in a way that questions in the pure sciences are not. The multisite spatial distribution of many studies along with their time frames is thus often essential to their research design.

Consider a concrete snapshot from the most prestigious corner of the biomedical research publication world. The December 29, 2011, issue of the New England Journal of Medicine includes five “original research” articles. They have 27, 6, 29, 19, and 13 authors, respectively. All but one of them is written on behalf of a much larger research group.Footnote ² Each of those four articles links to a supplementary appendix listing hundreds more collaborators or participating investigators. All five articles in that issue list authors from various countries and various institutions within each country. Each article's authors have diverse disciplinary backgrounds; for instance, Wang et al. (Reference Sismondo2011) combines authors from dozens of disciplines, including biostatistics, cell biology, oncology, and immunology, among many others. Such research projects are so dispersed and multidisciplinary that no one can have access to more than a small corner of them. The biostatisticians and immunologists presumably do not fully understand each other's contributions, researchers in France do not know exactly how subjects were recruited at the Taiwan site, and so on. This is decentralized, distributed collaborative research on a massive scale—a scale frankly unthinkable until recent developments in communication technology.Footnote ³

Whatever being an ‘author’ comes to in this context, it shares little with the traditional understanding of scientific authorship. Indeed, actual practices of assigning authorship make no pretense that authorship tracks responsibility for the production or the justification of the results. There is no presumption that the actual writing of the article, which many scientists see as extrascientific busywork, will be performed or even overseen by the lead ‘authors’. Authorship constitutes an institutional status; it does not represent a specific form of epistemic labor. Authorship is assigned and ordered according to whose grant money was used, who runs the lab, who has tenure, who needs a job, and so forth. As Sergio Sismondo points out, “Current practices in the medical sciences legitimately allow people to serve as authors on the basis of narrow contributions. Therefore many near-honorary authors find little reason to feel uncomfortable with their roles” (Reference Ross, Hill, Egilman and Krumholz2007, 1429). When researchers are accused of allowing their names to be placed on ghostwritten articles, they often insist indignantly that they played some role, such as helping edit a final draft of an article, or having enrolled some patients in the trial. One accused researcher insisted, “The fact that the draft was written by a Merck employee for later discussion by all the authors [sic] does not in and of itself constitute ghostwriting” (Saul Reference Poordad2008).

Universities and journals have decried ghostwriting without being willing to implement specific policies and sanctions to block it; they presumably recognize that it is by no means clear how to delineate legitimate authorship given current practices. But ghost authorship in collaborative, distributed biomedical research is a continuous extension of ‘legitimate’ authorship in this domain. Moffatt and Elliot write that in ghostwriting cases, “only the ghostwriter and the company representatives can vouch for the study as a whole” (Reference Huebner, Kukla and Winsberg2007, 22), but surely this wildly understates the problem. It is implausible that the ghostwriter—typically a professional writer for hire—is in any position to vouch for any part of the research process. More to the point, often no one can vouch for the study as a whole, whether or not it is ghostwritten. Many contemporary biomedical research articles, even when they accord with current norms of scientific practice, have no author at all in any traditional sense.

One possible response to my argument that such work has no traditional author is insouciance or even celebration. After all, collaborative, multisite, interdisciplinary research is clearly prized rather than regarded with suspicion by granting agencies and top journals. Arguably, radical collaboration fosters innovation and the cross-fertilization of ideas. The traditional author is dead, and good riddance to him, we might think. Individualistic models of epistemic authority and knowledge production are so passé; we all know by now that knowledge is a social product anyhow!

I believe, in contrast, that the death of the traditional author in collaborative biomedical research poses a real problem in applied social epistemology. Authors are traditionally the ones accountable for the security and justification of published knowledge claims. If there is no such accountability—if no one can tell a story about how we know what we say we know, and take responsibility for defending that story—then we have a prima facie epistemic problem on our hands. In radically distributed, collaborative research, there is no one who has a cognitive state instantiating a full justification for the claims that make it to print. The justification, if it is instantiated at all, is distributed across many agents. Anyone who is insouciant about the death of the traditional author needs to have some alternative story about how and when we can be confident that such a justification exists and how we can hold someone or something accountable for it.

3. Authorship and Testimony

Perhaps the most obvious prima facie response to my argument so far runs as follows: it's true that no one person has all the expertise and bilocational abilities needed to design and run such a study on her own or to reconstruct the justification for each of its parts. But in legitimate science, all of the authors can take responsibility for the whole study, because they have testimony-based confidence in all its parts. They understand what each of their collaborators contributed, know that each of their collaborators is competent and trustworthy with respect to her part of the whole, and can explain how those parts fit together. Legitimate collaborations actually have an embarrassment of riches: multiple authors, not just one author or no author at all.

This is apparently the view of the authors of “Boceprevir for Untreated Chronic HCV: A Randomized Clinical Trial,” which states, “Each author vouches for the fidelity of the trial conduct to the protocol and the completeness and accuracy of the results and data analyses” (Poordad et al. for the SPRINT-2 Investigators Reference Kahn2011, 1196). This is a report on a Merck-funded study “designed, managed, and analyzed by industry authors in conjunction with academic authors” (1196), and this statement is clearly designed to address recent concerns about ghostwriting and bias in such publications. Unfortunately it is radically implausible, if taken as an epistemic claim. The article has 15 authors in four countries employed by 11 academic institutions plus Merck. It is presented on behalf of the SPRINT-2 investigators. The 68-page appendix lists 173 principle investigators, based in 11 countries, who participated in the study, each of whom presumably relied in turn upon a wide variety of support staff to actually implement the study, keep track of the data, and communicate with other participants and with Merck. With the number of people playing some role in the path from study design to implementation to publication running in the hundreds or even thousands, the idea that each of the 15 authors could know that everyone involved was competent and reliable, that they adhered to the protocol in the same way, that the data set was complete and accurate, and so on is absurd.

Testimonial knowledge is based on an assessment of the reliability, in context, of the source of the testimony. In radical, distributed biomedical collaborations, testimonial knowledge securing each part of the study is impractical. Furthermore, even if someone had such testimonial knowledge, it is implausible to think that she could have the expertise and access to information she would need in order to vouch for the claim that all the competently executed parts fit together coherently into one big justified whole. Mere appeal to testimony will not solve our problem.

4. Group Authorship

Here is another possible response: while no one individual can be an author of the kinds of articles we have been looking at, the group as a whole serves as a single author, which is collectively in command of the justification and can take responsibility for the final published work. It is a substantial and interesting project to figure out under which conditions, if any, a group can serve as an author capable of the relevant kind of epistemic accountability. In Huebner, Kukla, and Winsberg (Reference Hardwign.d.), my coauthors and I explore this question and argue that the kind of radical collaborations I am discussing here are not contenders for group authorship. For present purposes, I will just point out that there is no clear group involved that has enough unity or cohesion to serve as a group author. Who would be ‘the group’ here? All of the official authors? The official authors plus the several hundred listed investigators? How about the wide range of support staff that implemented the study or the funders who posed the research question and coordinated the study? Regardless of how we answer such inclusion questions, I see no reason to think that the research process in these cases involves a unified and coordinated group with a shared set of interests or a shared understanding of the methods, goals, and background theory framing the research. Instead we have loosely gathered teams and researchers with different disciplinary training, working simultaneously at great distances from one another and with little or no direct contact.

In general, the field of biomedicine is not marked by a neat community knowledge base and shared set of methods but, rather, by a messy assemblage of tentative local consensuses, promising hypotheses, factions and disagreements, untested but entrenched practices, local and global professional and disciplinary competitions and rifts, and fluid and shifting epistemic communities and alliances. There does not seem to be any obvious contender here for ‘the’ group that can serve as an author with a unified voice and set of justificatory standards, who is in a position to impose coherence on the whole process or guarantee that it can be recovered. At a minimum, the burden is on someone who proposes the group authorship solution (a) to give a theory of what a group author is and how it manages to serve its authorial function, including taking epistemic responsibility for its claims, and also (b) to show that there could be some such group entity in place for the kinds of radical collaborations I have been examining.Footnote ⁴

5. Trust Processes, Not People

A third possible response is the one that is most interesting to me here. The line runs: it's not a problem that reports of collaborative biomedical research are authorless, as long as we have institutional mechanisms in place that ensure that each part of the whole project is run competently and that information flow and coordination between the parts is working properly. We need not treat any agent as authorial (whether it's an individual or a group), as long as we can trust the process that led to the final written product, by building epistemic accountability into the architecture of research.

A more specific, particularly optimistic version of this response runs as follows: the problem with ghostwriting and the ghost management of research is that it is biased in favor of the interests of the pharmaceutical companies paying for this research, and it hides this bias through deception, by blocking access to raw data, hiding true authorship, and so forth. If we could ensure transparency, we would not need to worry about assigning authorship. The way to do this is through rules prohibiting proprietary data, requiring publication of negative results and trial registration, punishing ghostwriting when it is discovered, insisting on detailed declarations of all conflicts of interest, and so forth. This way, even if there is no one or no group accountable for the whole project, we can be confident that its results are not distorted.

This is roughly the implicit position that journal editors have taken; there has been no particular fuss, in the publishing world, over what an author is and what it means for a work to be authored; instead the emphasis has been on transparency. Major medical journal articles, including those I described from the New England Journal of Medicine, are now published with links to enormous supplementary documents many times the length of the original articles, detailing the protocol, listing page after page of conflicts of interest, and describing the role each listed author played in the project. Most policy proposals have likewise been aimed at ensuring transparency.

I think some architectural solution has to be the right way to go. We will not and should not call a halt to radically distributed, interdisciplinary, collaborative research. Not only are such research practices deeply entrenched, but many of the most pressing research questions in medicine are, at this point, inherently such as to require multisite, large-scale trials drawing on a wide range of expertise in order to answer them. And in such a context, it is unrealistic to think that anyone or any unified group entity will be able to serve a traditional authorial role. And so our applied social epistemology must take an architectural turn.

But the optimistic view that focuses only on transparency seems hopeless. Note again that for this type of solution to work, we need to guarantee both that each part of the project is trustworthy and transparent and that all the parts fit together in such a way that there is some coherent justificatory story that leads all the way from the data and study design to the claimed results. That is, all the trustworthy and transparent bits have to fit together in a trustworthy and transparent fashion. Where there is a traditional author, this coherence is supposedly ensured top-down. But there are reasons why we cannot assume an easy bottom-up path from competence and honesty at the level of the distributed parts to coherence and accountability at the level of the whole. I want to focus on one problem that bears directly on the ghostwriting scandals: the role that interests play in shaping research and its representation.

Before we can trust a published scientific article, we need to know whether its contents have been unacceptably skewed or distorted by interests we may not share. Data like Healy and Cattell's strongly suggest that research controlled by pharmaceutical companies is skewed in this way and that its dummy ‘honorary’ authors are in no position to know exactly how. Most proposals presume that if we can ensure procedural transparency and access to all raw data, we will have eliminated or at least made visible the interested distortion of research. But the problem is more epistemologically thorny than that. I will follow many others in arguing that social interests ineliminably shape research and the representation of its results, even when there is no deception involved. Then I will argue, more boldly, that recovering and assessing the role of interests in radically collaborative, interdisciplinary, multisite biomedical research can be nearly impossible and that there is no reason to think that the interest-driven parts of such research add up to a coordinated and coherent whole.

6. Interests and Inductive Risk in Research

Private interests can lead people to do dishonest or incompetent things—to hide, forge, or misinterpret data, for instance. But even the most honest and competent of researchers must make a wide variety of interested decisions during the course of the research and writing process—decisions not compelled merely by love of truth. Putting aside the obvious fact that interests determine which topics we study, designing and implementing a clinical trial, for instance, requires settling exactly which interventions to compare, in what doses and using what delivery methods, what control group to use, how to operationalize the outcomes, how to code and interpret the data, which populations to enroll in the trial, and so on. None of these questions can be settled on purely epistemic grounds. For example, whether we compare a new drug to a placebo or to the standard of care will depend, among other things, on whether we are more interested in isolating its biochemical effect or in showing that it provides benefits not already available on the market. What dose of a drug we choose to test may depend on what doses are or are not replicable with current generics and hence unpatentable.Footnote ⁵

Sankar and Kahn write, “Conflicts of interest do not influence research only by biasing results. They can also assert themselves into research by allowing potential financial gain to distort the design and trajectory of research in ways that might ultimately undermine broad and affordable access to important therapies” (Reference Perren2005, w462). This is right, but note that there is no such thing as a ‘pure’ study design; the goal of enabling “broad and affordable access to important therapies” is just as clearly a social interest as is the desire to get results that will carve out a profitable market niche for a drug. Interests that are not purely epistemic ineliminably shape study design in multiple ways that do not go away when data are handled honestly.

The necessary role of social interests in shaping research is not limited to the study design stage. In 1953, Richard Rudner argued that scientists could not present results without making value judgments; since no scientific hypothesis is ever verified with certainty, the decision whether to accept a hypothesis in light of the data depends on a necessarily value-laden judgment about how to balance the inevitable inductive risks—that is, the risk of a false positive (accepting the hypothesis when it is in fact false) and the risk of a false negative (rejecting the hypothesis when it is in fact true). Using his examples: if our hypothesis is that a toxic ingredient in a drug is not present in lethal quantities, we will likely be highly intolerant of false positives, and this means upping our standards for hypothesis acceptance, which means tolerating a higher risk of false negatives. If on the other hand our hypothesis is that a lot of belt buckles are not defective, we will likely tolerate a far higher risk of false positives, thereby lowering our risk of false negatives. No interest-independent standards can govern such inductive risk judgments, because “how sure we need to be before we accept a hypothesis will depend on how serious a mistake would be” (Rudner Reference Moffatt and Elliott1953, 2).

So interests play a necessary role in shaping study design at the beginning of research, and at the end, when results are interpreted and represented. If the point ended there, we could perhaps just implement procedures that ensure transparency with respect to both the interests governing study design and the confidence with which results are reported. But the story becomes interesting, for our purposes, with a recent spate of work in philosophy of science that has built a case for how interest-driven inductive risk judgments penetrate all stages of the research process.

In her seminal article, “Inductive Risk and Values in Science,” Heather Douglas (Reference Burger2000) argued that there are multiple points during the course of research at which we must make inductive risk judgments in the face of uncertainty. For instance, in studies on the carcinogenic effects of dioxin, expert pathologists differed in their classifications of the same slides of rat liver tumors as benign or malignant. Confronted with borderline cases, interests will help determine how we classify and code what we see, even given baseline competence and honesty. Such in medias res inductive risk judgments are interesting because they cannot possibly be codified and made transparent as part of the protocol; expert judgment is required in moving from perception to codification. Another nice example comes from Torsten Wilholt: in testing the toxicity of Bisphenol A in rats, it turns out that 90% of government-funded studies report significant effects from low dose exposure, while 0% of industry-funded studies do. It also turns out that different strains of rats are differentially sensitive to estrogen, a compound similar to Bisphenol A, and that the industry-funded studies were more likely to use the less sensitive rats. Here the relevant interested inductive risk judgment is, as it were, embodied in the chosen rat itself. There is no correct rat with the correct level of sensitivity built into the structure of the universe.Footnote ⁶

The kind of interest-shaped inductive risk judgments that Douglas and Wilholt point us to can be expected to be especially rife in medicine, where there are often no clean, prefabricated standards for collecting, coding, and managing data. Consider how much wiggle room there is in deciding whether a research subject counts as depressed, or infertile, or experiencing unacceptable amounts of pain. Or consider how we decide whether to discard or keep the data from a subject we suspect might not have fully complied with a protocol. In such cases, we must make some judgment about the relative risk of false positives and false negatives, given our interests, and there is no correct standard written into the scientific method.

The point so far is as follows: procedures ensuring transparency are not enough to give the epistemic accountability and security that we sought from traditional authorship. This is because throughout the research process, interests necessarily shape research and the representation of its results. At a minimum, if we are to trust procedures rather than people, we need some rather more sophisticated procedures. We would need some way of making perspicuous which interests are at play and how they are affecting research, so that we can decide whether and how to trust the represented results, given our own interests. We would also need standards for deciding when particular interested choices exceed the boundaries of the socially reasonable range.Footnote ⁷

7. Messy, Decentralized Research

But I want to argue that the problem we face is yet more challenging. The literature on ghostwriting paints a picture in which a pharmaceutical company mastermind funds and controls the entire research process in order to produce a finished product—a well-placed, high-impact article—that will maximize profits. But in many cases, radically collaborative, decentralized biomedical research is not driven by a neatly unified telos in this way, nor are its parts coordinated in any such top-down fashion. Major biomedical research articles in top journals typically ‘disclose’ literally hundreds of different financial conflicts of interest. Each author typically lists dozens of sources of funding and support, ranging from shares to grants to lecturing fees and more. Disclosing a funding source does nothing to tell us how it shaped the relevant interests; we have no idea what sorts of micropressures the participants were under to keep various grants, satisfy various donors, wrap up a study, and so on. And while it is standard for articles to disclose financial conflicts of interest, people's interests, as they complete their corner of a research project, are much richer and more multidimensional than this. At any given site, postdocs are under pressure to please their advisers, scientists are attached to their pet hypotheses and concerned to build their reputations, faculty are under pressure to fill up their annual reports, there are disciplinary turf wars and grand ideological battles, professional competitions and investments, local university and community pressures, and much more. All of these factors can help shape decisions about how to proceed during the course of research.

Peter Galison writes of massively collaborative experimental physics, “It is entirely possible, even likely, that no one individual (much less a group of individuals) is entirely in control over the full spectrum of justificatory arguments that feed all the way down. … It may well be that experimental knowledge in the age of massive collaborations never comes back to a single center, but rather only to partially overlapping, complicated, inchoately bounded assemblages” (Reference Carson2003, 351). This is an apt description of biomedical research as well, but I am less sanguine than Galison about the consequences. Different collaborators have different expertise and skill sets and may well not share a set of methodological standards or be in a position to assess each other's. They may face different local pressures, or have different ideological or professional commitments, or different personal conflicts of interest. Furthermore, biomedical research—unlike some other highly collaborative and distributed research, like particle physics—typically deals with messy bodies and unruly research participants, often in far-flung regions riddled with material, communicative, and organizational challenges, and it is hard to predict in advance what sorts of uncertainties will need to be negotiated.Footnote ⁸ There will be research subjects who disappear or are noncompliant, breakdowns in linguistic communication, and so forth. It is thus difficult to predict or control in advance what sorts of uncertainties researchers will need to negotiate. Where there is no top-down authorial control over the research, I see no reason to assume that all these rough and ready assemblages cohere with one another, nor that in the end they “feed up” to a result with a recoverable, critically assessable, challengeable justification.

The values and interests that shape a research project may thus be buried so far down that they are impossible to retrieve. They “cannot be recovered in bite-sized pieces” but rather “lie in all the nooks and crannies,” to use helpful language borrowed from Eric Winsberg, in the course of his making a kindred argument about the role of values in climate modeling (Winsberg Reference Sismondo2012, 132). In these days of widespread private funding and intense professional and ideological pressures of various sorts, many of these interests are quite different from and may be in tension with the traditional interests that serve as the ideal ends of medical research, namely, the discovery of accessible and effective tools for protecting and improving health and well-being. Neat stories like the ones we tell about the testing of Zoloft, involving “bite-sized” interests, result from top-down pressures; we cannot easily tell the story of how interests shape research from the bottom up, in the absence of authorial accountability and control.

I have argued that much of contemporary biomedical research is not authored at all. The epistemic division of labor in contemporary medical research is decentralized, complex, and deeply messy, and no one is accountable for the ways in which interests shape the research process. In this context, ghost authorship of the sort that is clearly rife in privately funded research is not a sharp departure from authorial norms, but continuous with them; its pervasiveness is a symptom rather than a self-standing problem. If we are to secure accountability for biomedical research through architectural solutions, we will have to be far more creative than we have been so far in coming up with them.