Data Leverage: A Framework for Empowering the Public in its Relationship with Technology Companies

Before continuing, we first present formal definitions of data leverage and supporting concepts. Note that while aiming to be comprehensive, these are working definitions. Data leverage is an emerging topic in a rapidly moving field, and we aim to advance and open the discussion around data leverage, not conclude it.

• •

  Data leverage: The power derived from computing technologies’ dependence on human-generated data. Data leverage is exerted when a group influences an organization by threatening to engage in or directly engaging in data-related actions that harm that organization’s technologies or help its competitors’ technologies.

• •

  Data levers: The specific types of actions that individuals or groups engage in to exert data leverage. For instance, “data strikes” (Vincent et al., 2019a) are one of the data levers we discuss below and they operate by cutting off the flow of data to tech companies.

Data leverage emerges in part from work in the broader FAccT community that has demonstrated the limitations of purely technical approaches to advancing fairness and justice in computing systems (Abebe et al., 2020; Binns, 2018; Eubanks, 2018; Gebru, 2019; Chancellor et al., 2019). This large literature emphasizes the critical roles played by the societal context around computing systems, and has demonstrated that sociotechnical approaches are often much more powerful than purely technical approaches. Data leverage can in many ways be understood as a framework that helps us better understand data-driven technologies through a sociotechnical lens and use that lens to take action to achieve pro-social outcomes.

Data leverage is more specifically informed by Kulynych et al.’s work that proposed “Protective Optimization Technologies” (POTs) as a way to address the negative impacts of algorithmic systems and give agency to those impacted (Kulynych et al., 2020). POTs allow people to contest or subvert optimization technologies, perhaps adopting techniques from data poisoning (which we further address below) (Kulynych et al., 2020; Troncoso, 2019). Data leverage and POTs are synergistic concepts, and many POTs enable people to exert data leverage.

Data leverage is heavily informed by work that views data generated by people using computing systems as a type of labor. Building on Posner and Weyl (Posner and Weyl, 2018), Arrieta Ibarra et al. argue that data should be considered as labor, not “exhaust” emitted in the process of using technology, and as such, should be subject to some kind of remuneration (Arrieta Ibarra et al., 2018). The relationship between the data-generating public and the companies that benefit from data is very asymmetric. Not only do people have very little knowledge of — let alone agency over — how data they contribute is used, but the economic winnings from powerful data-dependent technologies are reaped entirely by tech companies (Posner and Weyl, 2018). To mitigate this inequality, Posner and Weyl called for the formation of “data unions”, which allow data laborers to collectively negotiate with technology companies (Posner and Weyl, 2018).

The discussion around data labor has inspired work that aims to measure the economic value of data (Vincent et al., 2018, 2019b; Koh et al., 2019; McMahon et al., 2017). One approach has been to look at the relationship between Wikipedia — the product of data contributions from the public — and real-world economic outcomes such as tourism and investment (Hinnosaar et al., 2019; Xu and Zhang, 2013). Building on the data as labor concept, Vincent and colleagues have investigated how people might withhold or redirect their data labor to force a data-dependent organization to change its practices (Vincent and Hecht, 2021a; Vincent et al., 2019a).

Scholars working on data feminism — an intersectional feminism-informed lens for data science — have called for more efforts to make the labor of data science visible, including the labor of data generation (D’Ignazio and Klein, 2020). These scholars argue that the invisible labor of data science, much like housework, has been hidden from public view and therefore undervalued (D’Ignazio and Klein, 2020), and that researchers can begin to shine a light on this labor by studying and highlighting the processes of data creation (e.g. (Crawford and Joler, 2018)). In this way, data feminism is very aligned with the ideas of data leverage; both aim to measure and make people aware of the value of previously invisible labor and ultimately reshape power imbalances.

The data leverage concept is also informed by work from HCI and STS on technology “use”, “non-use”, and the spectrum of behaviors in between.

Work from Selwyn and Wyatt called attention to the need to understand people who do not use new technologies (Selwyn, 2003; Wyatt, 2003). Most relevant to data leverage, Selwyn documented that people engage in ideological refusal to use certain technology “despite being able to do so in practice”. Further calls to study non-use in HCI and STS have been amplified in the years since (Satchell and Dourish, 2009; Baumer, 2014).

Use and non-use exist on a spectrum (Wyatt, 2003; Baumer et al., 2019). People face many social and technical decisions in terms of when they will use, and stop using, a particular technology, and these decisions lead to many different forms of use and non-use (Brubaker et al., 2016; Schoenebeck, 2014; Baumer et al., 2013). Recently, Saxena et al. reviewed the methods for creating typologies to describe the many forms of use and non-use (Saxena et al., 2020).

Many factors motivate non-use, such as exclusion (Wyatt, 2003), social capital (Lampe et al., 2013), and socioeconomic factors (Baumer, 2018; Baumer et al., 2019). Anyone seeking to use data leverage to empower the public must contend with these factors. Attempts to support data leverage could exclude or disproportionately benefit certain groups following existing patterns in how technology excludes and benefits these groups.

One common theme in the non-use literature is that it is not easy for people to refrain from use when it comes to products that have some benefit in their life, even if the benefit(s) come with a host of long-term drawbacks. People often speak of their technology use as a type of addiction, using terms like ‘relapsing” and “withdrawing” (Baumer et al., 2015; Baumer et al., 2013). Challenges also emerge related to the public presentation role of social media profiles  (Lampe et al., 2008). Even if people stop using a technology, they may not necessarily delete their data. In studying individuals who left Grindr, a dating app, Brubaker et al. found that “even among those who deleted the app, only a minority tried to close their accounts or remove personal data…[putting them] in a paradoxical position of thinking they have left while their profile — or data — continues on” (Brubaker et al., 2016).

The non-use literature also indicates that people engage in protest-related use and non-use behaviors for reasons relating to privacy, data practices, perceived addiction, and other issues  (Baumer et al., 2013; Baumer et al., 2015; Li et al., 2019; Stieger et al., 2013). Anyone engaging in such behaviors is a potential participant in data leverage campaigns. Casemajor et al. and Portwood-Stacer argue separately that non-participation in digital media can be an explicitly political action(Casemajor et al., 2015; Portwood-Stacer, 2013). Li et al. conducted a survey to better understand “protest users”, or people who stop or change their use of tech to protest tech companies(Li et al., 2019). The results suggested that there is a large number of people interested in protest use: half of respondents were interested in becoming protest users of a company, and 30% were already engaged in protest use.

An important related lens is that of “refusal”. Focusing on bioethics, Benjamin makes the case that broad support of “informed refusal” provides a means of developing a justice-oriented paradigm of science and technology (Benjamin, 2016). In practice, people who engage in informed refusal are engaging in a political form of non-use, and thereby data leverage. Building on Benjamin’s work, Cifor et al. and Garcia et al. describe how the notion of “critical refusal” informed by feminist scholars can be used improve practices around data (Garcia et al., 2020; Cifor et al., 2019).

Understanding the full potential of data leverage requires deep engagement with machine learning literature. Two relevant areas of ML research are those that answer questions around (1) the effectiveness of adversarial attacks on data-dependent systems and (2) the relationship between a system’s performance and changes to underlying data.

There is a large literature that considers the case of adversaries attempting to attack ML systems (e.g. (Barreno et al., 2006; Pitropakis et al., 2019; Biggio et al., 2012; Shan et al., 2020; Shafahi et al., 2018; Steinhardt et al., 2017; Lam and Riedl, 2004; Lee and Zhu, 2012; Gunes et al., 2014; Chirita et al., 2005; Mobasher et al., 2005; Fang et al., 2020; Li et al., 2016)). In early work on adversarial ML, Barreno et al. developed a taxonomy of attacks on ML systems (Barreno et al., 2006). They focused in particular on attacks in which an adversary “mis-trains” a system, which is called data poisoning. Data poisoning attacks against many types of ML systems have been studied in detail (Barreno et al., 2006; Pitropakis et al., 2019; Biggio et al., 2012; Shan et al., 2020; Shafahi et al., 2018; Steinhardt et al., 2017). A type of data poisoning attack that is particularly relevant to the work in this paper is the “shilling” attack, which involves “lying” to a recommender system so that a system recommends certain products favored by the attacker (Lam and Riedl, 2004). Accordingly, much work has been done on counteracting shilling (e.g. (Lee and Zhu, 2012; Gunes et al., 2014; Chirita et al., 2005; Mobasher et al., 2005)), which may be of concern to groups who want to use shilling-style data poisoning attacks to exert data leverage as we describe below. Researchers have also explored advanced “data poisoning” techniques that use sophisticated methods to optimally harm ML systems (Fang et al., 2020; Li et al., 2016), which can be much more effective than unsophisticated attacks (e.g. providing random or average ratings to many items (Lam and Riedl, 2004)).

Data leverage raises the stakes of the already high-stakes adversarial ML domain. This paper highlights how adversarial techniques, such as data poisoning, are not just relevant to issues of security and privacy, but also to the power dynamics between users and tech companies. While some recent work in adversarial ML has taken a political lens and highlighted real world examples of how adversarial ML can create socially desirable outcomes (Albert et al., 2020), most of the literature takes a strictly security-oriented lens.

The literature on the relationship between the amount of training data a model has access to and model performance is also highly relevant to data leverage. Many authors have found diminishing returns of additional data across many contexts and algorithms (e.g. (Cho et al., 2015; Dietterich and Kong, 1995; Figueroa et al., 2012; Hestness et al., 2017)), and some have studied techniques to address diminishing returns (Bloodgood and Callison-Burch, 2010). These findings are informative as to how effective data leverage can be.

This paper builds on the literature that explores how the public can change practices of the technology industry. Data activism is a relatively new form of civic participation in response to tech companies’ pervasive role in public life (Baack, 2015).

Currently, data activism encompasses practices that affect technology design, development, and deployment (Milan and Van der Velden, 2016). Data leverage can be seen as a subset of data activism with a specific focus on empowering the public to influence the performance of data-dependent technologies. Milan and Van der Velden provided a typology of data activism that further illustrated the specialized activities in this space — proactive and reactive data activism (Milan and Van der Velden, 2016). Proactive data activism refers to activists directly influencing software development or databases through open source projects or collaborating with institutions. A particularly relevant data activism initiative is the open data movement, which aims to democratize information that is currently only accessible to the state or businesses (Gurstein, 2011). For example, Baack studied an open data project in Finland and highlighted the intermediary role of data activists between the public and operators of data-dependent technologies (Baack, 2015). On the other hand, reactive data activism entails activists acting against data-collecting entities through adversarial behaviors such as employing encryption. Data leverage includes both types of data activism.

Equipped with the knowledge and expertise to understand data’s role in computing, researchers can provide the public with valuable information to identify and employ effective data leverage practices. Work on data activism has unveiled a rich space to improve data practices (Couldry and Powell, 2014). In particular, Lehtiniemi and Ruckenstein called for “linking knowledge production to data activism practice” to gain a comprehensive understanding of data’s role in the public sphere (Lehtiniemi and Ruckenstein, 2018).

The first of the data levers we will consider are data strikes. Data strikes involve a person withholding or deleting data to reduce the amount of data an organization has available to train and operate data-dependent technologies. Although the term data strike is relatively new, the concept builds on the well-studied practices of stopping or changing technology use as a form of protest, as discussed in Related Work. For instance, groups have participated in prominent boycotts against companies like Facebook and Uber (Semuels, 2017; Greenfield et al., 2018). In another example, people use ad blocking software to deprive companies of data about the success of their ad placements (Budak et al., 2016).

A data poisoning attack is an adversarial attack that inserts inaccurate or harmful training data into a data-dependent technology, thereby encouraging the model to perform poorly (Barreno et al., 2006). While data strikes harm performance by reducing the amount of available data, data poisoning harms performance by providing a technology with data that was created with the intention of thwarting the technology. A relatively accessible way that users can engage in data poisoning is simply by leveraging standard technology features in a deceptive manner. For instance, someone who dislikes pop music might use an online music platform to play a playlist of pop music when they step away from their device with the intention of “tricking” a recommender system into using their data to recommend pop music to similar pop-hating users. Other straightforward examples include the coordinated effort to create sexually explicit Google search results for former U.S. Senator Rick Santorum’s name  (Gillespie, 2017) and coordinated campaigns to use fake reviews to promote certain products (Lam and Riedl, 2004). As we will describe below, very sophisticated variants of data poisoning that draw on state-of-the-art machine learning research are also possible.

The above tactics operate by harming, or threatening to a harm, a given data-dependent technology. However, there are cases for which harmful tactics are not a good fit. For instance, perhaps users do not have the regulatory support needed to delete past data (Waddell, 2020) or a new technique for detecting poisoned data foils their poisoning attack. Harmful tactics may also be undesirable because an organization’s technologies may actively provide benefits to others (e.g. a ML model that is well known to improve accessibility outcomes).

“Conscious data contribution” (CDC) (Vincent and Hecht, 2021a) is a promising alternative to harm-based data leverage. In CDC, instead of deleting, withholding, or poisoning data, people give their data to an organization that they support to increase market competition as a source of leverage. People using CDC for data leverage are similar to people engaging in “political consumption” (Koos, 2012), but instead of voting with their wallet, they vote with their data. An exciting aspect of CDC is that while small data strikes struggle to put a dent in large-data technologies because of diminishing returns, CDC by a small group of users takes advantage of diminishing returns and provide a competitor with a large boost in performance. We return to this point later in our assessment of data levers.

In general, CDC has the lowest barrier to entry of the data levers we identified. This is because CDC does not require stopping or changing the use of existing technologies, which prior work discussed above indicates can be challenging (e.g. (Baumer et al., 2013; Baumer et al., 2015; Schoenebeck, 2014)). A person can continue using existing technologies operated by an organization against which they want to exert leverage while engaging in CDC (Jones and Tonetti, 2019; Vincent and Hecht, 2021a). The main barriers to transfer-based CDC are regulatory and technical. Do laws help people transfer their data (Por, 2018) and do tools exist to make data transfer realistic?

The barriers to entry for data strikes are more substantial then those for CDC and less substantial than those for data poisoning. While participating in a data strike disrupts a user’s access to online platforms, strikes do not necessarily force a user to stop using a platform like a traditional boycott would. For instance, a user who relies on Facebook to communicate with family members could stop engaging with sponsored content on Facebook but continue messaging their family members. An Amazon user might continue buying products but stop leaving ratings and reviews. An important downside of data strikes is that they hurt the performance of technologies for participating users. By cutting off data contributions, an individual often reduces their own ability to benefit from a system. As discussed by Vincent et al. (Vincent et al., 2019a), the effect of a data strike will almost always be most pronounced on the strike participants.

The barriers to entry for each data lever are also contingent on the bandwidth available to potential participants and any potential data caps or data charges they have. Data strikes are likely the least limited by bandwidth (although striking against an Internet provider, e.g. Facebook Free Basics, could be challenging (Sen et al., 2017)). In places where the Internet is easy to access and has relatively high data caps, poisoning data by letting music stream for hours or actively manipulating multimedia may be accessible. In contrast, in places where Internet access is limited (Dye et al., 2017), poisoning data may be difficult if not impossible. Similar dynamics likely will apply to CDC: data caps could stifle efforts to engage in CDC.

Many of the barriers to entry discussed above are not equally distributed across different populations, and this means that different populations likely have differing access to data leverage. For instance, with regards to data poisoning, the time available to expend the necessary effort and/or the skills necessary to do so will limit the ability of many populations to engage in data poisoning. Those most positioned to perform data poisoning attacks are ML researchers, technologists, and others with strong technical skills, an already relatively privileged group. Nonetheless, members of this group could use their powerful position for the benefit of people without these advantages (there is precedent of tech worker organizing along these lines (Tec, 2020)).

Turning to coordination, data leverage campaigns will differ in their coordination needs, with greater coordination requirements raising the barrier to entry for all three data levers. Large-scale data leverage is possible without formal organization: boycotts using Twitter hashtags provide real-world examples (Jackson et al., 2020). However, certain data levers require especially well-coordinated effort to see impact, e.g. sophisticated data poisoning (Fang et al., 2020).

Data leverage organizers may face legal and ethical challenges. Withholding-based data strikes face the fewest of these challenges. These data strikes require almost no regulatory support as users can simply cease using platforms (keeping in mind the differential barrier to entry concerns discussed above). Deletion-based data strikes require a right to deletion and a guarantee that companies are not data laundering by retaining model weights trained on old data (Lyons, 2021).

The legality of data poisoning is likely to remain an open question, and interdisciplinary work between computer scientists and legal scholars will be critical to understand the legal viability of data poisoning as a type of data leverage (and to do so in different jurisdictions). Arguments about the ethics of obfuscation (which itself can be a form of data poisoning) raised by Brunton and Nissenbaum apply directly to the use of all types of data poisoning (Brunton and Nissenbaum, 2015). Participants must contend with the potential effects of dishonesty, wastefulness, and other downstream effects of data poisoning. For instance, there are many harms that could stem from poisoning systems that improve accessibility, block hate speech, or support medical decision-making.

Interesting legal and ethical questions also emerge around CDC. Notably, if a certain data-driven technology is fundamentally harmful and no version of it can meaningful reduce harms (as can be argued for e.g. certain uses of facial recognition (Abebe et al., 2020; Gebru, 2019)), CDC will effectively be neutralized.

Another challenge specific to CDC is that there is the potential that data contributions by one person might violate the privacy of others, as data is rarely truly “individual” (Acemoglu et al., 2019; Ben-Shahar, 2019). For instance, genetic data about one individual may reveal attributes about their family, while financial data may reveal attributes about their friends. On the legal front, CDC often requires either regulatory support in the form of data portability laws or data export features from tech companies.

Data strikes and data poisoning harm data-dependent technologies, while CDC improves the performance of a data-dependent technology that can then compete with the technology that is the target of data leverage. We can measure potential impact in terms of performance improvement/degradation, as well as downstream effects (e.g. performance degradation leads to users leaving a platform). Ultimately, we are interested in how likely a data lever is to successfully change an organization’s behavior with regards to the goals of the data leverage effort, e.g. making changes related to economic inequality, privacy, environmental impact, technologies that reinforce bias, etc.

A relevant finding from prior work (Vincent et al., 2019c) describes how data strikes interact with diminishing returns of data. ML performance exhibits diminishing returns; in general, for a particular task, a system can only get so accurate even with massive increases in available data. As such, when an organization accumulates a sufficient amount of data and begins to receive diminishing returns from new data, that organization is not very vulnerable to small data strikes. Such strikes will — broadly speaking — only unwind these diminishing marginal returns. To a company with billions of users, a (relatively) small data strike simply may not matter.

The potential impact of data poisoning is also enormous: a large-scale data poisoning attack could render a dataset completely unusable. This approach is also appealing for bargaining: a group could poison some data contributions, and make some demand in return for the “antidote”. However, the enormous corporate interest in detecting data poisoning means that the would-be poisoners face a constant arms race with operators of targeted technologies. In the worst case scenario, they will be caught, their poisoned data deleted, and the end effect will be equivalent to a data strike.

CDC campaigns, which improve technology performance, operate in the opposite direction of data strikes. Small-scale CDC could be high impact: about 20% of the users of a system could help a competitor get around 80% of the best-case performance (Vincent and Hecht, 2021a). On the other hand, once returns begin to diminish, the marginal effect of additional people engaging in CDC begins to fall.

Given the current evidence, we believe that the data levers we described have a place in the tool belt of those seeking to change the relationship between tech companies and the public. A critical challenge for data leverage researchers will be identifying the correct tool for a specific job. Based on the technologies a target organization uses, a realistic estimate of how many people might participate in data leverage, and knowledge about the resources available to participants, which data lever is most effective?

Researchers, practitioners, activists, policymakers and others interested in studying, supporting, or amplifying data leverage to reduce power imbalances must contend with unequal access to data leverage. As discussed above, there is strong reason to expect that inequalities in access to data leverage mirror known patterns in access to technology and other sources of power more generally (Abebe et al., 2020). However, our framework suggests that data poisoning and CDC in particular might allow small groups to have disproportionate impacts. A group of users with needs not currently met by existing technologies might engage in CDC to support a competitor to existing tech companies, or use sophisticated data poisoning techniques that require coordination and knowledge, but not mass participation. Researchers can play an active and critical role by developing tools and promoting policy that widely distributes the ability to participate in data leverage efforts and receive benefits from data leverage. Future work may also need to contend with the possibility of organizations counteracting data leverage, e.g. removing access to publicly available data to maintain a dominant market position.

Many lucrative data-dependent technologies rely on “commons” data (e.g. Wikipedia and OpenStreetMap) in addition to the largely proprietary types of data we have discussed so far (e.g. interaction data, rating data). The same is largely true for a variety of data sources that are privately-owned but are a sort of de facto commons for many purposes (e.g. Reddit data, public Twitter posts). Examples of commons-dependent technologies include large language models (e.g. (Brown et al., 2020)), search engines (e.g. (Vincent et al., 2019b; McMahon et al., 2017; Vincent and Hecht, 2021b)), and a variety of geographic technologies (e.g. (Johnson et al., 2016)). Commons datasets have also been instrumental to the advancement of ML research (e.g.(Baumgartner et al., 2020; Bib, 2020)).

How can we view the widespread dependence on commons datasets through the lens of data leverage? Adopting a narrow perspective, all three data levers can certainly be employed using data in the commons. In fact, doing so might be a very effective way of exerting data leverage against a large number of data-dependent technologies at once. For instance, through poisoning (i.e. vandalizing) Wikipedia, one can negatively affect a wide variety of Wikipedia-dependent technologies including Google Search, Bing, and Siri (Vincent et al., 2019b; McMahon et al., 2017; Vincent and Hecht, 2021b). Indeed, this has already been done with humorous intent a number of times (e.g. (Wilmhoff, 2017)). One could similarly imagine organizing a “data strike” of sorts in Wikipedia or OpenStreetMap that sought to ensure that a certain type of information does not appear in these datasets.

That said, from a broader perspective, it is very likely that data poisoning and data strikes using commons data will do substantially more harm than good. For instance, a concerted effort to vandalize (i.e. poison) Wikipedia will cause substantial damage: it would harm Wikipedia readers across the world and would affect technologies operated by non-targeted organizations in addition to those operated by targeted ones. A similar case could be made for most data strikes.

CDC in the context of commons data presents a more complex set of considerations. Indeed, contributing to a commons dataset like Wikipedia can in some ways be understood as a type of CDC as it helps smaller organizations as well as larger ones. However, an important consideration here is that the ability to make use of commons datasets in data-driven technologies is gated by capital. A salient example is GPT-3, OpenAI’s high-profile language model that uses training data from sources like Wikipedia and Reddit (Brown et al., 2020). The unprecedented computing power needed to train GPT-3 highlights how the data labor that improves Wikipedia and Reddit can disproportionately benefit organizations with enormous resources. An unfortunate reinforcing dynamic regarding data leverage and commons data thus emerges: while a huge number of organizations and individuals stand to be harmed by any sort of poisoning attack or strike on commons data, large and wealthy firms often stand to benefit disproportionately from improvements to these data. Future work that focuses on efficient training, smaller models, and related goals can help to mitigate this particular concern. Similarly, efforts to open-source models themselves (e.g. share model weights) could also help.

We have presented data leverage as a means to empower the public to address concerns around computing systems that exacerbate power imbalances and create negative societal outcomes. However, research, tools, and policy intended to help data leverage achieve these goals could do the opposite by empowering groups to perpetuate inequalities and, therefore, achieve socially harmful outcomes. For instance, hate groups take advantage of “data voids” in search engines to engage in what can be understood as data poisoning attacks by inserting hateful content and influencing model development (Golebiewski and Boyd, 2019). Why wouldn’t these groups also try to use other types of data leverage for similar ends?

There are no clear-cut ways to eliminate these risks, but there are steps that data leverage researchers can take to avoid a “backfire” outcome. When designing tools to support data leverage, designers might consider heuristic preventative design from Li et al. (Li et al., 2018) and try to make harmful uses of a technology more challenging. For instance, a data poisoning tool might only help users poison certain types of images known to be important to a particular company or technology. Designers should also consider the principles of data feminism (Cifor et al., 2019; Garcia et al., 2020), including those that emphasize challenging existing hierarchies, embracing pluralism and context, and making labor visible.

The concept of data leverage presents exciting research opportunities for many fields. Researchers in FAccT, ML, HCI, STS and related areas in particular have unique opportunities to amplify data leverage.

Data leverage presents a new way of exerting pressure on corporations to make important changes. Most relevant to the FAccT community, this might involve exerting leverage so that a tech company stops the use of a harmful algorithm (Kulynych et al., 2020), or pushing for new economic relationships between data contributors and AI operators in which the benefits of AI are shared more broadly (Posner and Weyl, 2018; Vincent et al., 2019c). Data leverage thus presents a novel avenue for researchers to actively pursue pro-social research roles and goals (Abebe et al., 2020).

There is enormous potential to support data leverage with ML research methods. Using simulations and small-scale experiments, future work could build a catalog of results that activists could draw on to make predictions about the effectiveness of a particular data lever in a particular context, such as “if we get $x$ participants to engage in a data strike against technology $y$, we can expect to bring down the accuracy of technology by $z$%, which will likely be enough to encourage company $c$ to make the changes we are demanding”. As data leverage becomes more mainstream, there may also be opportunities to study real-world examples and answer key questions such as: What are the downstream effects on revenue, user retention, and actual changes in company behavior?

Future design work could build upon the collective action literature and develop tools to coordinate efforts to use data leverage. For example, because collective action’s progress is often opaque to individual participants and this can negatively impact engagement, future work may adopt tactics from “boycott-assisting technologies” (Li et al., 2018) and display the impact of the public’s data strike or poisoning (e.g. this technology has lost 3% of data). Such tools could also support automating data strikes or data poisoning, similar to AdNauseam Howe and Nissenbaum, to lower the barrier to entry for the public.

In addition to data strikes and poisoning, the computing community can also support CDC by addressing data compatibility and portability issues across platforms and technologies. Data generated by users are often highly platform- and/or technology-dependent. For example, ratings for the same restaurant or hotel may vary significantly across review platforms (Eslami et al., 2017; Li and Hecht, 2020). Directly transferring data from one technology to another as an act of CDC may run into compatibility issues and even negatively affect the recipient’s performance. There is a need for researchers and practitioners to develop software that automatically translates data generated using one technology into data can truly benefit another technology to maximize the success of CDC-based approaches.

Researchers should also seek to better understand the full set of societal impacts that would result from the widespread use of data leverage. As we have discussed above, we hypothesize that the direct effects of actioning data leverage will often involve broadly positive societal impacts, e.g. improved privacy, better distribution of the economic benefits from AI systems, more democratic governance of AI systems. However, the second- and greater-order effects of these changes are more difficult to assess, and even some direct effects may be negative in some cases as highlighted previously. More generally, data leverage defines a pathway to altering power structures in the current computing paradigm. Alterations of power structures in such a complex sociotechnical environment will almost certainly lead to complex outcomes, and more research will be needed to understand these potential outcomes.

Data leverage stands to benefit heavily from regulatory support. As such, data leverage research should be deeply engaged with policy by highlighting regulatory approaches that are likely to amplify the power of data leverage and address its potential negative impacts. Our taxonomy only scratches the surface of how policy may support data leverage; we are excited for this important direction of future work.

Following directly from our assessment of data levers above, we suggest a variety of ways policy can support data leverage:

• •

  Data portability laws will directly enhance CDC, enabling users to contribute data they helped generate in the past.

• •

  Right-to-delete laws will enhance data strikes, assuming these laws also account for the possibility that companies might “launder” deleted data in model weights.

• •

  Data transparency laws that make data collection more apparent may help foster support for data leverage movements.

We note that these policy suggestions are generally aligned with policy aimed at addressing privacy concerns. This suggests a potential “win-win” situation, in which policy simultaneously supports consumer privacy and enhances data leverage.

Expanding on the above points about data portability and right-to-delete laws, policy also offers the potential for making it easy for individuals to use multiple data levers in conjunction with one another. As mentioned above, there are natural connections between data strikes and CDC: by moving from one platform to a new platform, a user can take advantage of both data levers. However, through regulatory support, it may be possible to engage in much more elaborate combinations of data strikes and CDC, for instance deleting only certain pieces of data and transferring over other pieces of data.