Technical Communication's Fight Against Extractive Large Language Modeling by Applying FAIR and CARE Principles of Data

Findable and Accessible

FAIR's principle that data must be findable and accessible coincides with the challenge in technical communication for the public to understand or interpret the LLMs in AI technologies such as Grammarly. The origins, processing, and applications of data in these LLMs remain inaccessible, which prevents others from verifying the source and integrity of the data, identifying biases, or ensuring the reliability and trustworthiness of commercialized AI. Companies like Grammarly have a vested interest in ensuring that their data are inaccessible so a product cannot be reproduced. Indeed, AI companies often limit or block computational systems from finding their proprietary data.

At the user level, finding and accessing user content data that the application has collected can be a difficult, time-consuming process. First, a user must navigate through the application's Terms of Services (ToS) to find what data have been collected and held by the company, where they can find their data, and how they can access the data. Some policies are prominently displayed and accessible on the company's website, but Grammarly has at least a half-dozen important policies distributed across its site with no search feature at hand. These ToS issues obfuscate users from knowing their data rights—a long-standing power problem between companies and users (Böhme & Köpsell, 2010; Fiesler et al., 2016). If businesses and individuals want to know more about the specific data that are sampled and used, they must request personal data reports from Grammarly; however, Grammarly only provides aggregated summary statistics in a PDF or text file (see Figure 2). Overall, Grammarly does not provide users with findable and accessible data or a means to understand the scope and scale of what user data it collects and retains.

OPEN IN VIEWER

Figure 2.

Partial example of Grammarly's personal data report.

Interoperable

Data are interoperable if they can be integrated within other technological systems and used by others for their goals. The data referred to in Grammarly's accessible user summary are not interoperable because the file's information is unstructured. Grammarly could instead offer a structured data set in interoperable file formats, such as CSV (comma-separated values). These data are interoperable because the provided values are not an agreed on provision that users get to define.

Reusable

The prior two principles of the FAIR framework ensure that data can be reused by people who did not create the data. The principle of reusable data assumes an open sharing of disaggregated data, so Grammarly's business values directly conflict with it. This principle also assumes a closer relationship between invested parties because the metadata of reusable data strive to document the data's exigencies and common goals (GO FAIR, 2023). If users from an Indigenous community wanted to better understand what data are collected and used by Grammarly's LLM, the company's aggregated per-document statistics, weekly email reports, and personal email reports do not meet the standards of the reusable principle. In sum, Grammarly's data practices obfuscate what Indigenous knowledges and practices they are collecting and using to develop their LLM without building the capacity to foster more equitable partnerships between the company and the community.

Collective Benefit

Collective benefit puts the onus on companies to adopt a different relationship with Indigenous communities and nations. This principle compels companies to ensure that they work with Indigenous communities to develop an equitable and agreed-on set of benefits defined by the communities. We use the following three subprinciples—inclusive development and innovation, improved governance and citizen engagement, and equitable outcomes (Carroll et al., 2022)—to assess Grammarly's modeling practices.

C1. Inclusive Development and Innovation. Carroll et al. (2022) suggested that research teams consider how their data protocols “actively support the use and reuse of data by Indigenous nations and communities” (p. 4). Grammarly maintains almost total access and control over a significant portion of user data, and users play little to no active role in creating and sharing reusable and interoperable data. Instead, Grammarly tasks teams to create what company linguists Ng and Nahorna (2022a) called “gold data” by hiring data labelers to tune their models. In describing this complex structural problem of extractive design and power, we do not mean to single out any particular Grammarly employee; rather, we cite Ng and Nahorna's findable and accessible information as representing Grammarly's modeling practices. Grammarly does not transparently communicate its collective and community approaches when explaining how potential users benefit from its capture of user data to develop its LLM. This issue conflicts with IDSov: Indigenous self-determination in relation to data.

C2. For Improved Governance and Citizen Engagement. This subprinciple asks how the data should inform and improve the relationships between user communities and institutional and commercial entities, such as Grammarly. Because Grammarly's modeling centers on the more extractive work conducted by internal teams, individual users can only decide to use the product and generate personal data reports on user content. This relationship seems strange and troubling if we consider that Grammarly has access to vast amounts of highly personal, private user content data that it shares with other entities, such as external service providers (Grammarly, 2023).

C3. For Equitable Outcomes. Prospecting “gold data'’ from users reinforces extractive practices of white-settler colonialists that steal land and its resources for the benefit of the few privileged enough to mine and model such data at a global scale. Based on the available information, Grammarly policies do not support IDSov, and its broad capture and use of user data impede data equity efforts.

Authority to Control

When users sign up for Grammarly, they relinquish control over how their data are leveraged by Grammarly and who has access to it. Grammarly is authorized to use user content for limited purposes, including refining its products, training the AI model, introducing new features, and promoting new services (Grammarly, 2023). As users employ the app on their desktop or browser plugins, Grammarly has access to emails, text messages, and documents—both individual and shared. When users agree to these terms, they may be sharing not just their own information with the company but potentially the data of their friends, family, and community.

A1. Recognizing Rights and Interests. Grammarly addresses the protection of personal data in its policies. Yet Grammarly appears to have a broader latitude in how it uses data sourced from user content. Grammarly's removal of personal identifying information (PII) from user content data can still result in misconceptions or improper use. Such considerations become even more vital for Indigenous and local communities, whose data might be of a sensitive or culturally sacred nature and have significance not meant to be shared with others outside of specific contexts.

A2. Data for Governance. Carroll et al. (2020) stated that “Indigenous Peoples must [emphasis added] have access to data that support Indigenous governance and self-determination” (p. 6). Grammarly does not permit users to gain access to the data that they extract from their use. While much of our analysis has centered on Grammarly capturing user data to train its LLM, it is important to note that data about seemingly mundane tasks, such as how users at tribal organizations write emails in intercultural and intracultural ways, can be useful for tribes to understand and analyze their own communication practices. Thus, capturing this data with no means for users to access and use it for their own self-determined goals is harmful.

A3. Governance of Data. In the same way that an LLM can train in a dialect by analyzing a large data set, which often comprises scraped data from the web, and pulling out linguistic patterns, an LLM can pull knowledge and information from regional text inputs, which could perpetuate harm to communities. Yet there is no communicated method to discover how Grammarly uses user content and who is using it. All users must permit Grammarly to store and display their content, as well as modify and develop new versions of it in order to improve and market their products. This authority is also extended to Grammarly's (2023) external service providers. Overall, users waive their authority to control their content, and their prospective data “only used in aggregate” can still be harmful if not tempered by CARE-informed governance policies.

Responsibility

Responsibility bridges the previously applied FAIR and CARE principles. It emphasizes relationships and requires that those who hold positions of power enact a culturally humble approach to data practices. As part of our own CER project, we defined a cultural humility framework (CHF) for community research partnerships (Itchuaqiyaq et al., 2023), which we also apply to commercial research and development. The CHF aligns with the responsibility principle of CARE because it calls for entities like Grammarly to understand how their scaling efforts are built on the imbalanced power dynamic between themselves and the communities from which they extract data.

R1. For Positive Relationships. Cultural humility promotes positive relationships. Mutual trust requires sharing data and insights and submitting to users’ authorial rights, needs, and direction. As previous FAIR and CARE evidence demonstrates, Grammarly maintains a position of power over user communities rather than imagines equitably balanced authorial control by submitting to, or even sharing leadership with, users. Taking a humble approach to modeling is necessary to repair the historical disrespect of Indigenous sovereignty in research and data work.

R2. For Expanded Capability and Capacity. Since Grammarly controls access to all disaggregated data and LLMs, the company does not responsibly facilitate Indigenous self-determination. If Grammarly's mission “improves lives,” then users and communities must lead them. Users, such as the tribal members from our CER team, are interested in tools such as Grammarly to expand their own capability and capacities as researchers. But using these tools would put tribal knowledges at risk. Our tribal partners represent sovereign Indigenous nations, and for them, the benefit of easily writing grammatically correct sentences does not outweigh the cost of tribal sovereignty.

R3. For Indigenous Languages and Worldviews. Our CER project supports Indigenous languages and worldviews, and if the use of generative AI were ethical, it could assist our team with those efforts. But Grammarly's issues of access cannot foster responsible relationships. Our CER project supports Indigenous languages and worldviews, and the use of ethical generative AI could assist our team with those efforts. Yet, again, the risks to IDSov outweigh the potential benefits of using its tools. Cultural humility recognizes the responsibility that Grammarly has as stewards, not owners, of Indigenous knowledges and language practices.

Ethics

CARE defines an ethic in which Indigenous Peoples “inform the use of data across time in order to minimize harm, maximize benefits, promote justice, and allow for future use” (Carroll et al., 2022, p. 2). Grammarly (2023) states that it makes decisions about its services based on a set of ethical principles dubbed as EAGER: Ethical, Adaptable, Gritty, Empathetic, and Remarkable. Although the company attempts to define these values, the language is vague and highly interpretive based on a person's background, perspectives, and experiences. As Edenfield (2018) discussed, companies develop vague values and mission statements (Lytvyn, 2021) to construct a uniform culture that inspires order but remains flexible enough so that the companies can adapt to corporate challenges and expand. This strategy is employed by Grammarly. A Grammarly business statement (Lytvyn, 2021) states that the EAGER values underscore every business decision the company undertakes, but the ambiguity of the statements makes it difficult for the company to develop more actionable ethical data practices. Conversely, the FAIR and CARE principles can remedy this difficulty.

Considering Grammarly's operations, the responsibility for data ethics extends beyond individual employees and must be reflected in the corporation's policies and procedures. As Walwema et al. (2022) emphasized, “Ethical frameworks help us understand how to enact justice and identify the behaviors, actions, and policies that should be considered just or unjust” (p. 259). Grammarly claims that EAGER informs its mission to foster inclusive and learning-centric environments for its team, offer empathetic user support with integrity, and create products that connect people. But as we will now discuss, users pay for a service that also collects and commodifies their knowledge and language practices in order to realize Grammarly's goals, definitions and methods of writing, and monetary gain.

E1. For Minimizing Harm/Maximizing Benefit. This subprinciple demands that businesses never define Indigenous data as deficient or trapped in a Western colonialist historical narrative. Using Grammarly and data labeling are existing means for users to develop an LLM. This transactional relationship, however, does not minimize harm or maximize benefits for communities. Instead, it serves to prospect Grammarly's gold data in a way that minimizes its costs and maximizes its profits.

E2. For Justice. For LLMs to potentially redress colonial power imbalances, Grammarly must be led by Indigenous Peoples. Whether community representatives are sought after as leaders of LLM development is unclear and unlikely, but Grammarly's Ng and Nahorna (2022b) do recognize that “everyone has their biases” and discuss their approach to annotation practices that they claim will “eliminate bias” (31:45). They suggest that modelers can “ensure diversity” by “hiring a large enough number of annotators.” Yet by “diversity” they do not mean a diverse population; rather, they are referring to diverse labeling tasks, such as not tasking labelers with correcting a single type of grammatical mistake. But as long as Grammarly teams retain control of the process, just LLMs cannot be developed.

E3:.For Future Use. Regardless of its EAGER principles, Grammarly falls short in truly minimizing harm and maximizing benefits for any users, let alone Indigenous peoples in relation to CARE, because it does not allow user access and control over what content is being sampled to enhance the product. For a just future, regulations and policies must replace the prospecting of gold data with CER practices that take up FAIR and CARE principles.