“Machinery Hurtful to [Scientific] Commonality”: Automation and Scientific Publication

In the past decade, the model for academic journal publishing has shifted from subscriptions to open access forms of publication. Enabled in part by new business approaches based on article-processing charges (APCs) where authors pay for publication, open access in a commercial environment tends to reward publishers who can publish as many articles as possible as quickly as possible, and with as little paid human input as possible. Commercial publishing has therefore become about high article volume, economies of scale, and technological automation. Papers are funnelled through publishing systems with minimal friction in the hope that technological smoothness will enable the publication of ever-increasing scholarly content.

The drive for article volume means that many commercial publishers are losing track of what they publish, allowing fraudulent science to slip through the net and legitimate journals to become “hijacked” by editorial collectives of bad actors. For example, the publisher Wiley recently shuttered its open-access Hindawi imprint—a company it acquired in 2021 for £298 million—due to its having become “heavily compromised by paper mills.”¹ These paper mills make their money by selling authorship on papers that they can guarantee will be published, which they ensure through various measures such as image manipulation, data fabrication, automated content creation, journal hijacking, and citation “cartels” of people willing to review manuscripts favorably. In turn, because of their focus on automation, publishers are increasingly unwilling to hire humans to perform specialist checks around fraud and research integrity (checks that go beyond the scope of the peer review process) and are looking instead to their own automated detection services to spot the work of paper mills. These suites of software, such as Wiley’s “AI-powered Papermill Detection service,” are also being made commercially available so that other publishers can automate their fraud detection checks. Both AI fraud and AI detection are rapidly becoming big business.

Within this turn to automation, scientific publishers are also broadening out their business models from mere publishing toward research analytics, looking to amass user data in order to sell AI insights and predictive tools to research institutions hoping to understand researcher performance.² The datafication of researchers is well illustrated by the trends in research infrastructure design, particularly those designed to be walled gardens to keep users interacting with publisher services from submission to publication and beyond. I have theorized this strategy elsewhere as a process of “individuation through infrastructure” in which publishers seek to turn platform users into individual, trackable, and traceable units in order to follow their behavior across a range of services.³ This data is also being sold onto AI technologists to be used as training data, as Informa has done recently in a deal with Microsoft.⁴ According to media theorist Jeff Pooley, we have moved into an era of “surveillance publishing” whereby abstracted researcher behavior will “double back onto our work lives” and reinforce existing biases in the research process.⁵ Most pernicious is how ungovernable these processes are, not least by academics themselves but also by broader democratic society, controlled instead by sophisticated proprietary algorithms geared toward extraction and profit over scientific progress.

The way to work through the issues of automation within academic publishing is primarily a question of governance. In the face of such technological upheaval, it seems almost futile to suggest a possible solution to such a fast-moving threat to knowledge production. Regulation is currently powerless because policymakers do not know what to regulate, nor is academic publishing particularly high on the long list of concerns in tech policy.

Instead, academics should deploy tactics to both reduce the pace of commercially driven automation and increase the governance of scholarly knowledge infrastructures more broadly. There are ways to do this through collective and continuous action, but they require strategic nous and partnership with other industries impacted by automation. The scientific fraud detector Elisabeth Bik speaks of the need to “slow down scientific publishing” in order to force publishers to check their articles properly.⁶ But there is also the need to de-marketize publishing and return accountability to academic communities rather than the market at large. For example, there is a growing trend of academics resigning from their publishers en masse and taking their journals to more ethical presses.⁷ Such a trend reveals the potential of academic communities to refuse to participate in harmful business approaches to publishing and to work toward alternatives. This action could be encouraged and facilitated through universities in a variety of situations, especially in the context of the growing trend toward more ethical forms of “diamond” open-access publishing that do not rely on APCs.⁸ There is also a space for academic unions to increase governance and demand accountability over data processing by private companies, particularly through the higher education institutions that pay for these services.

None of this is to say that there are not benefits in the move to automation, but more that the harms are so profoundly obvious and out of direct community control that they require urgent action on behalf of all stakeholders in academic knowledge production. While it is commonplace to cite the Luddites in the context of generative AI, I think they are a helpful touchstone for the kinds of tactics needed to slow down the rapid development of unaccountable AI technologies. Grounded in collectivity, the Luddites set their sights not on wrecking technological development itself but on what they termed “machinery hurtful to commonality”—that is, the technological processes that allowed capitalist owners to eradicate human labor, divide workers, and pay lower wages. This resistance, as Eric Hobsbawm shows, was often a deliberate “resistance to the machine in the hands of the capitalist” rather than mechanized labor more generally.⁹ In the face of such a profound threat from AI, researchers should take a leaf out of the Luddites’ playbook by intervening and organizing wherever possible to prevent automation imposed by commercial publishers and analytics companies.

The struggle against automation is made more difficult by the fact that datafication itself is a form of individuation that actively works against commonality. When scientific publishers track a researcher’s journey across the research lifecycle, they are in fact creating that researcher as a series of data points, abstracting them from their unique situational context and reinforcing the understanding of subjectivity that is detached and quantifiable. Yet knowledge production makes sense only as a collective activity, as something emerging out of local contexts and community interactions, or what we might term a form of commons. Indeed, looking to the expansive literature on common-pool resource management is one way to understand how such collective management and stewardship of these commercial technologies could take place. Governing our AI futures then becomes about how to collectivize knowledge production more generally, allowing research communities to take control of the machinery so hurtful to commonality.