Principle 1: Consider who has “refrigerator rights” to data

This principle invites an interrogation of how data boundaries are set and enforced, and crucially, finding ways to stabilize intangible boundaries. Unlike refrigerators, datasets do not come with solid doors to clearly demarcate what is public and what is private, and therefore, when collecting and using the data of those who have historically been marginalized, or whose data is culturally sensitive, it is vital to protect and signal off-limits data in other ways. This understanding has led to several initiatives for safeguarding data, ones that acknowledge the complexities and tensions between supporting open data and scholarship and protecting the rights of individuals and groups. It is especially useful to consider the kinship rights to refrigerator space articulated by Blackman in the context of Indigenous data sovereignty, and the ways in which traditional knowledges are mediated by technologies that are often misaligned with Indigenous approaches to how information is shared. The CARE Principles for Indigenous Data Governance (Collective Benefit, Authority to Control, Responsibility, and Ethics) were developed in response to the need for Indigenous groups to set out terms for how, and for what purpose, their data is collected and used (Carroll et al. 2020, 3). In essence, these principles create ways to determine “refrigerator rights” to data, elucidating the meanings of how Indigenous communities share and protect their knowledges and intangible cultural materials.

We can similarly interrogate the boundaries we set, and those we respect when it comes to our own data. It is strange to imagine the marketing manager of a company walking into your house and going into your fridge, yet people willingly “confess” their data and give companies refrigerator rights to their personal information (Kitchin 2014, 180). The massive amount of web data, and especially the use of social media to mobilize movements such as the Arab Spring and #MeToo make these platforms a desirable source of data for researchers (Wilson 2022, 5; Jacobson and Gorea 2022, 703). While most social media platforms are completely free, they are data collection systems with varying levels of opacity. In 2018 the Cambridge Analytica scandal revealed that Facebook handed over personally identifiable information (PII) data on 87 million users to the data firm Cambridge Analytica, who then leveraged this data to profile users and send targeted information to voters during the Trump campaign (Isaak and Hanna 2018, 55–56). This realization highlighted the risks associated with user privacy on social media platforms and resulted in Meta implementing more robust privacy protections for user data. However, the truth remains that when we encounter services that require no money to join, it is likely that we ourselves, and the data we willingly offer up for collection, are the product. Online surveys, loyalty programs, memberships, and newsletter signups all require a base level of personal data to participate and become rich sources of data for companies to use, exploit, or market further products to us (Kitchin 2014, 165). It is imperative that we advocate for ourselves and critically assess these value-adds against the risks of sharing private information.

There is a reason we feel uncomfortable going into someone’s fridge. Should we not experience the same discomfort when accessing someone’s data, or in giving access to others? To whom are we giving our precious refrigerator rights?

Principle 2: In algorithms (and refrigerators), garbage in equals garbage out

This principle reflects on the ways in which the quality and accuracy of machine-learning outputs is dependent on the quality of the data chosen to feed them. Data and the systems we employ to process them are not neutral and become embedded with the biases of the humans who create them (Buolamwini 2023, 49). The embedded bias in datasets does not just appear: it begins at the outset of collection when decisions are made about what to collect or omit from collection, where to collect data, from whom, and for what purpose. The mechanisms to collect data from individuals are not consistently designed to allow users to adequately choose the gender, ethnic group, race, or sexual orientation that best describes their identity. Text boxes often do not accommodate non-Western or Indigenous names or allow for single names or multiple names. These decisions become “amplified as data are computed upon, inflated by algorithms, and distilled by visualization” (Thorp 2021, 60–61). This suggests that omissions or deliberate over-representations ensure that certain outcomes from the data are pre-determined long before we begin the process of analysis. In automated decision-making technologies, such as those used in law enforcement or banks’ use of “big data” to construct risk models to process loan applications, inequality perpetuated by embedded biases threatens an individual’s chances for equitable treatment. For example, where a person may benefit from a positive result, such as a job interview or loan application, algorithms fed data based on so-called desirable persons are designed to filter out women and people of colour (Dastin 2018). Conversely, where a positive match results in a negative outcome, algorithms used in law enforcement are fed data which skews results to include more non-white individuals, and AI used in predictive policing is designed to direct attention to neighbourhoods where crime is most likely to occur (Demiaux 2017, 25).

There have also been instances of facial recognition being based on incomplete or flawed data. In her 2019 report “Garbage In, Garbage Out: Face Recognition on Flawed Data,” Clare Garvie of the Georgetown Center on Privacy and Technology reviews the work done in several studies, demonstrating how garbage data can violate basic civil rights, for example, when a system returns false positives, or when suspects are not informed of how their image has been used to train AI models (Garvie 2019). When confronted with identification tasks, facial recognition software displays both gender and age bias, and fails to accurately recognize gender-neutral, Asian, and Black faces (Buolamwini 2023, 242).

Machine-learning systems based on flawed systems reproduce racial and socioeconomic biases, exacerbate inequality, and limit the ways in which individuals are able to succeed in the world (Datta et al. 2017, 1). In Canada, government and public sector institutions have adopted policies for data collection and use, such as the 2023–2026 Data Strategy for the Federal Public Service, to improve collection and address data gaps (Government of Canada 2023a). It includes amendments to StatsCan collection procedures to allow Indigenous Canadians to submit information under traditional names. The most current draft of updates to the World Wide Web Consortium (W3C), Accessibility Guidelines (WCAG) 3.0 include two recommendations pertaining directly to algorithmic bias. They recommended that content authors train AI models to be unbiased towards persons with disabilities, and that they conduct usability testing and ethics reviews to determine if persons with disabilities are being disadvantaged by algorithmic systems (W3C 2025).

It is understandable that for the intellectually curious DH practitioner, presented with vast amounts of data and imagining how it can be analyzed in creative new ways, we can lose sight of how the abstractions of datafication can obfuscate the individuals behind the data points. To develop a more humanistic approach to data, there are some specific ways to mitigate data biases and remind ourselves of this fact. We can interrogate available data for signs of bias before we use them in our projects. We can seek out missing or incomplete datasets, or build our own to address gaps, and include robust discussions of these gaps in our work. When we design surveys or otherwise collect data from human participants, we can follow ethical standards for data collection, such as the Tri Council Policy Statement: Ethical Conduct for Research Involving Humans (TCPS-2) (Government of Canada 2023b). Finally, because AI systems “create the illusion of both a fully autonomous process and impartial output” (Viola 2023, 10), we should abandon any notion that computers equal objectivity or that numbers equal neutrality.

Principle 3: Use best-before dates and periodically purge outdated, stale, or rotten data

This principle of data ethics pertains to the use of the most accurate and current data and considers how data collected today will be used in the future. As of July 2025, the internet is home to 1.2 billion websites, which includes innumerable databases, blogs, social media platforms, file sharing sites, message boards, gaming sites, periodicals, e-commerce sites, and search engines (Haan 2025). If we look at a database like Statistics Canada, which manages census data, it is fairly easy to see how the data is managed. Every five years, there is a new census with fresh data. It is posted on the website and older statistics are archived. Users have access to the most current data, and it is very clear what is old information.

With other sites and databases, it can be trickier. First of all, what constitutes “fresh” information? A date is the beginning, but sometimes the content cannot be considered the most current by the date alone. Historic information, such as that in the Slave Voyages Dataset or old census records, are likely as accurate as they will ever be (SlaveVoyages 2025). New information may be discovered and added later, but the data on a primary source material like a ship’s manifest is much less susceptible to data that changes with the passage of time. A feed on X, begun when the platform debuted as Twitter in 2007, is a dataset of the person or company who used it. The data on X is more personal and fluid, an online diary of events over time, registering shifts in our responses to the social climate and culture. A series of chronological tweets becomes a history of what we were thinking and saying at a specific point in time, and often these viewpoints change. As evidenced by multiple cases in the press, individuals who have had their lives derailed by the online chronicling of past indiscretions or opinions they once had but no longer do, exemplify the need for stronger policies pertaining to one’s right to be digitally “forgotten” by compelling service providers and search engines to scrub the offending material (Jones 2016, 3). In January 2022, Meta changed their data policy to allow users to review a chronological list of all their posts, likes, and comments and delete them individually or bulk-delete all at once (Instagram Help Center 2025). The intricacies of current debates surrounding digital “redemption”—including whether it curtails the free movement of ideas in the information age, to whether or not it is logistically possible to remove digital evidence entirely, even landing on a definition of what constitutes “privacy”—are more complex than the space here allows (Jones 2016, 81). One key point is that once data is scraped, it has already begun to age, and it is nearly impossible to retract information that has already been republished or used to train machine-learning systems. Buolamwini asserts that the practice of deep data deletion should involve not only removing “dirty” data (that which has been debunked, is inaccurate, or exhibits damaging bias), but also removing the AI models that have been trained on it (Buolamwini 2023, 101–102).

A historical dataset versus the real-time dynamics of social media are two extremes. Many datasets accumulate information more slowly, changing and evolving over longer periods of time. In the middle are other types of stale or downright rotten information whose effects can range from being a nuisance for researchers to causing measurable harm. These can include dead websites, link rot, or news stories containing now-debunked myths, which, if not followed by further research, can still be read as statements of fact. The New York Times digital archive turns up zero search results for articles about the doctor whose conflicts of interest with pharmaceutical companies led him to falsify data linking MMR vaccines to autism. However the article still lives in the archives of The Lancet, where it was first published, and while the document has been overlaid with the word “retracted” in giant red lettering, the existence of the document online means it can still be read, cited, and linked to (which I will not be doing here) by those with an interest in keeping its falsehoods alive. Sites may not meet the latest minimum standards for accessibility, such as those developed by the World Wide Web Consortium (W3C 2024) or be in possession of and posting culturally sensitive material that should be revisited, reframed, or repatriated to its rightful owners. Data is also financially and environmentally costly to store.

These harms are mitigated by a combination of policy and best practices for managing the lifecycle of data, including its creation or acquisition, storage and maintenance, usage and distribution, archiving and retention, and destruction (Ige et al. 2024). A formalized Data Management Plan (DMP) is frequently required by funding organizations that support DH projects, including the Social Science and Humanities Research Council of Canada (SSHRC 2025). While diligent DH practitioners make DMPs, outdated material thrives on the internet. A Pew Research study from 2024 estimated that 38% of websites operational in 2013 were no longer being updated, and that 21% of U.S. government web pages contained at least one broken link (Chapekis et al. 2024). That anyone can publish on the web is both a strength and a weakness, and as the Pew report found, best practices are not consistently employed, even on government websites. Owners of their own websites and datasets should take the time to take a critical look at their own content (Kuenn 2014). The unbridled freedom to publish online does not neatly align with legal or moral demands that the information be up to date. Should users be required to review information for accuracy, broken links, or outdated information, perhaps before they can renew their domain?

Principle 4: Obtain permission to use others’ data, as often as necessary

Asking permission before taking what does not belong to us is arguably the simplest principle to grasp, but one requiring multiple approaches to ensure we are using data permissions ethically and correctly.

In 2011, the company Reveal Digital, which works with university libraries to digitize cultural materials, secured the rights to On Our Backs, a lesbian pornography magazine published from 1984 to 2004. Five years later, the material was removed. Among the reasons for doing so was a quiet protest in the form of blog posts by an archivist and librarian, Tara Robertson, who pointed out that consideration of the historical and cultural value of the material must be weighed against the privacy rights of those who posed in the magazine, people who were members of a sensitive community. After all, she wrote, “consenting to a porn shoot that would be in a queer magazine is a different thing to consenting to have your porn shoot be available online” (Robertson 2016). What this example illustrates is one of the many facets of permission that digital practitioners must consider when embarking on projects that involve the collection and dissemination of sensitive data: permission is solicited and granted in many forms, and consent cannot be reused or assumed to be granted in perpetuity (Sutherland et al. 2023, 124). In Canada, the Tri-Council Policy Statement on Research Ethics requires specific standards to be met in the collection of data from individuals for research, but only for projects “conducted under the auspices of an institution eligible for funding by the federal Agencies” (CIHR, NSERC, SSHRC) (Government of Canada 2023b), and of course, not all DH projects are. However, permission and consent are not just about legality, but a community-based understanding of digital responsibilities.

This is the same tacit knowledge of right and wrong that governs our refrigerator conduct, and similar to permissions around ethical data collection and use, there is no particular set of rules applicable to every situation. We can assume some general and intuitive principles of permission and consent. Ask the right people, and consider the ways in which the use of data will impact all those involved. There must be clarity and transparency about what data we’re asking for and how it will be used. If we ask permission and the answer is no, then we cannot use it. If, because of the passage of time or reluctance of those whose data we are requesting, we cannot get an answer, we cannot use it. If the answer is yes, we must proceed thoughtfully and according to established ethical parameters, and then only take what we have been authorized to use. As in the On Our Backs example, if the answer is yes today, then the answer is yes today, not for another, unrelated project later; we need to ask again and make sure the answer is still yes. Consent must be ongoing, and participants given the option to withdraw their consent from projects at any time. We can also use tools like the Digital Research Ethics Collaboratory (DREC) to collaborate with colleagues in our field and adopt strategies that give people more control and agency over their information (DREC 2025). If we are not sure, or we find ourselves talking ourselves into it being okay, it is not okay.

If it feels uncomfortable to ask for permission, we need to question whether we are asking for too much.

Principle 5: Label and contextualize data

When used properly, labelling makes it clear what something is, what it is for, and who it belongs to. In digital contexts this can include the use of metadata, tagging, the practice of decolonizing archival catalogues, or applying labels.

The FAIR principles for scientific data management and stewardship state that data should be findable, accessible, interoperable, and reusable. At the crux of FAIR principles is the addition of metadata—data about data—that can be read by humans and machines, provides context about an entity, uses readable language, uses persistent identifiers to make information about entities findable and recognizable, and allows for data to be reused by researchers (Wilkinson et al. 2016). As D’Ignazio and Klein have articulated, a major problem with data downloaded from the internet is that a lack of metadata fails to provide the context required to analyze it (D’lgnazio and Klein 2020, 153). It is also not unusual that certain datasets are inaccessible by design and conceal rather than elucidate information. In 2014, the Boston Police Department was forced by a lawsuit to disclose the datasets they kept to record stop-and-frisk incidents, but the cryptically numbered labels did not describe the contents of the files and were not contextualized in a way the public could use (D’Ignazio 2019). It makes sense to legislate that open datasets be given plain language labels or annotations to allow for easier deciphering of labels.

In GLAMS (galleries, libraries, archives, museums, and special collections) where we work or conduct work, materials have been catalogued according to normative standards based in colonialism and white, male, Western, heteronormative supremacy, and therefore are embedded with inherent bias based on this positionality (Werling and Radio 2025, 84). Policies and best practices for critical cataloguing reject and disrupt labelling systems that flatten rather than specify, those that reproduce harmful categorizations and work toward further decolonizing archives and how information is organized within them (Sutherland et al. 2023, 124).

Working with Indigenous communities, the Traditional Knowledge Labels (TKLs) initiative provides a system of pictographic labels that can be used to identify Indigenous ownership of cultural materials (provenance), the conditions of its use (protocols) and who has access to it (permissions), cultural boundaries, and creating local and specific conditions for sharing of Indigenous cultural knowledge data, thereby protecting sensitive and sacred Indigenous knowledges (Local Contexts 2025).

Principle 6: Ensure barrier-free access to sharable data

If we assume that some people should have refrigerator rights to data, then how can we ensure that people do not face barriers to accessing information? Data fairness involves a fine balance between safeguarding private information and providing accessibility to data that people are entitled to, and different groups of people will require different tools to ensure they do not face physical, institutional, or socio-economic barriers to that access.

It could be about prioritizing open scholarship or providing ways to circumvent academic paywalls to provide more open access to books and scholarly journals. To keep pace with rapidly changing technologies, W3 has drafted an updated version of their Accessibility Guidelines to ensure web tools and content are fully accessible to persons with disabilities (W3C 2025). In the contexts of Western poverty, rural communities, and addressing the digital access gap between the global north and south, scholars have argued that unfettered and unmonitored access to the internet, provided free of charge, qualifies as a basic human right (Reglitz 2024; Cruft 2023).

Too often, the paternalistic controls under which so much information is tightly kept is often justified for the purposes of safety and security, such as video surveillance footage and passenger information collected by border agencies and airlines in the post-9/11 era, or the CCTV recordings of Londoners going about their day (Kitchin 2014, 178). In the earlier example of garbage in, garbage out facial recognition software, one recommendation proposed by Garvie was that defendants brought in for questioning based on a computer-generated composite should have the right to view the data inputs that created it (Garvie 2019).

Data kept under lock and key for the purposes of maintaining control over marginalized populations is why data needs policies to ensure that people do not need to stand on rickety chairs to access what they need, or worse, have no support at all.

Principle 7: Prioritize and practice reciprocity

This principle is very tightly bound with the idea of permissions. When we ask to use data, there should be more at stake than our personal gain, and a mechanism for returning the benefits of research back into the community.

For example, in 2009, the DiverseCity Project collected data on those in senior leadership roles across multiple public sectors in diverse communities in Toronto and issued a clear mandate to use the results to provide recommendations that would help ensure minorities were being served in their respective communities (Cukier and Yap 2009). There are other projects collecting data in ways that are less clear, and this raises questions about what they will do with it. A 2021 New York Times article profiled Jimmy Chen, the developer of an app called Propel, which he designed to assist those using the Supplemental Nutrition Assistance Program (SNAP), or food stamps, in planning and budgeting. Chen was quoted as saying that ad revenues for the app came from grocery stores advertising sales or companies offering employment: “the goal [is] to align the company’s financial interests with those of its users” (DeParle 2021). As the article pointed out, commenters on Twitter (now X), questioned Chen’s sincerity and wanted to know exactly how this app was supposed to aid the poor. In other words, Chen’s app, developed to fill a niche, was collecting ad revenues yet was not directly addressing the material conditions of those whose data it relied on to function and ultimately fund his personal venture. While social media critiques are arguably not the only or best way to gauge community benefit, this example shows how the power differential between food stamp users and tech CEOs (funded by Morgan Chase Bank, among others) exemplifies the inequalities that are reinforced when financial need is a factor in the decision-making of the end user of the app. Chen is ostensibly not a person receiving food stamps, and while the app may very well be helpful, even more so would be to consider how collected user data could be returned back to the community and leveraged to support food security initiatives.

Reciprocity is not limited to giving back after the fact. One of the most important ways to practice reciprocity is to engage in community outreach prior to the onset of a project, in ongoing consultation during its design, through continual evaluation throughout its lifecycle. The results of such considerations can move a project beyond the goals of its initial concept and create new opportunities for empowering and emancipating marginalized groups.

Principle 8: Acknowledge and challenge invisible labour

In digital contexts, realized work, that which we see as an outcome, is built on a vast infrastructure of invisible, unseen, and undervalued labour (Justesen and Plesner 2024, 2) that lies on various points along the ethical spectrum. Among the assertions emerging from critical DH discourses is that the very foundation of DH is highly skilled yet invisible labour (Smith and Whearty 2023, 28). In academic institutions, the paid work needed to further DH projects and pedagogy is often supplemented by colleague support in classroom and workshop environments, especially due to the employment of specialized technologies and methods (Croxall and Jakacki 2022). Many projects are supported by tasks undertaken by student research assistants who perform hours of work “under the hood,” building databases, cleaning and entering data and metadata into databases, and checking sources. In corporate environments, what Mary L. Gray and Siddarth Suri refer to as ghost work is the array of on-demand gig labour—proofreading, transcribing, captioning, inputting metadata, confirming identities—that drives the machine-learning operations of large tech companies and creates an underclass of workers lacking the employment rights enjoyed by full-time employees (Gray and Suri 2019, ix–xxiii). The work of social media content editors is largely done offshore in the Philippines and China, where “digital janitors” are (under)paid to supplement the work of algorithms in removing offensive or violent content (Cherry et al. 2016, 71).

After acknowledging the interdependencies of digital projects and the labour needed to realize them, challenging invisible labour means no longer accepting the “myth of seamlessness” (Justesen and Plesner 2024, 9). Pushing back against invisible systems, what D’Ignazio and Klein have referred to as “showing our work,” can be accomplished by interrogating the systems of power that undergird data collection, by being more inclusive of marginalized voices at every stage of project development, and by contextualizing data and its labels (D’lgnazio and Klein 2020, 184). The principles of both permissions (Principle 4) and reciprocity (Principle 7) contribute to this inclusion by revealing the human voices behind the data we collect and increasing visibility by focusing less on taking data and giving more attention to how data can be used to address power differentials. Even if we cannot remove the element of invisible labour entirely, for example, due to rigid capitalist structures, taking these additional steps and performing these different kinds of labour in our own projects can help make these structures, and the labour that supports them, more visible and valued.

Principle 9: Question our motives, not once, but throughout the lifecycle of digital projects

Getting to the root of our intentions requires reflection, and this principle can be applied to both the development of new technologies and the ways in which we collect and interpret data for digital humanities projects. The same way we can ask ourselves, “Am I truly hungry?” we can ask ourselves, “Are the people or groups represented by the data being empowered by this work? Is it furthering knowledge in a way that is respectful? Are we justifying shaky motives to meet a personal end?”

In our race to get technologies to market or to start a project that is exciting to us, we can fall into the trap of failing to critically examine our motivations and ask important questions about who our work will ultimately serve. This can lead to wilfully or unintentionally overlooking the downsides, or misreading cues about our own intentions, in much the same way we mistake feelings of emptiness or boredom for hunger. The excitement to digitize On Our Backs was rooted in the desire to make queer histories more accessible and correct the absence of lesbian stories in the archive. What was missing was the consideration that the right to be seen and heard is sometimes secondary to the right to be silent or forgotten. The original participants in On Our Backs were denied agency over their own images, how they were shared, and any involvement in the decision to digitize the archive.

The rapid development and adoption of generative AI has led many to question the safety and ethics of expediting extremely powerful technologies, often without public consultation or adequate testing. In 2023, the Future of Life Institute issued an open letter (signed by more than 33,000 people to date) urging the developers at Open AI to pause training of its models due to inadequate planning for the consequences (Future of Life Institute 2023). The letter’s assertion that “powerful AI systems should be developed only once we are confident that their effects will be positive and their risks will be manageable” exemplifies what a questioning of motives could look like, yet by the time the letter made the rounds, Open AI’s GPT-4 genie was already out of the bottle. Experts feared that by prioritizing speedy innovation over diligence in questioning the implications of AI systems, ones that were edging ever closer to artificial general intelligence, humanity was “sleepwalking into catastrophe” (Perrigo 2023).

What the On Our Backs example and Future of Life’s open letter demonstrate is that once projects are well underway, especially those with economic potential, activism after the fact has varying degrees of success. It is not a replacement for mindful and critical assessment of projects at the point of conception. If we consider that, as Johanna Drucker has articulated, knowledge production is “situated, partial, and constitutive,” humanistic/digital inquiry as a discipline is predicated on the acknowledgement that data is a malleable material in the construction of knowledge (Drucker 2011, 3). In other words, if data is neither inert nor indifferent, the principle of questioning our motives necessitates moving beyond obtaining permission. It requires asking what story we are seeking to tell, and if the data we take, Drucker’s “capta,” is being used in a way that helps us look at a situation differently, consider more people in more specific ways, move beyond our biases and assumptions, or ultimately become more generous digital practitioners.

Questioning our motives involves inserting human voices into digital projects and applying humanistic ethical questions to our technologies. The CARE principles emphasize that data and associated research must be employed for the collective benefit of Indigenous communities. For data practices to be considered ethical, Indigenous people must be the ones to determine collective benefit (Carroll et al. 2020, 6). If we question our motives and find them selfish or based on ill-conceived ideas of saving the heritage of peoples or cultures other than our own, sometimes the answer is to walk away. If our projects involve precarious permissions, then we should walk away. If the results of our research and the data we have gathered to support it do not result in some benefit to the community, if they lack reciprocity, then maybe we should walk away. If we look at our project and it does not include who it should include, in the form of missing data or a lack of representation in the project team, if the data is flawed, incomplete, or perpetuates bias, these should be remedied, or else we should not proceed. At all stages of a project, from its conceptualization to proposal to data collection to its end stages, there should be repeated consideration of whether the project or product creates harm, further marginalizes those already on the margins, creates barriers, or neglects to break those barriers down. We should be constantly questioning our motives and asking ourselves: “Am I truly hungry? Do I need this? Does the world need this?”

Editorial