Human Voices in an AI World: Diversity of Expression and Representation

Full report

On this page

The rapid digitization of human communication has precipitated a profound epistemological shift in how knowledge is recorded, synthesized, and transmitted to future generations. Historically, archives, libraries, and oral traditions functioned as the primary repositories of human culture. Today, artificial intelligence (AI) systems, particularly large language models (LLMs) and generative diffusion networks, are increasingly positioned as the central mediators of collective knowledge. The capabilities, worldviews, and outputs of these systems are entirely bounded by the datasets used to train them. These underlying data architectures encapsulate specific concepts, representations, and notions of truth, effectively acting as the arbiter of recognized human culture1. Consequently, the breadth of human-created content directly influences the systemic intelligence of AI. When digital archives and AI training repositories disproportionately reflect the perspectives of highly represented, resource-rich, and technologically integrated demographics, they risk engineering a form of epistemic collapse2. In such a paradigm, the vast and nuanced spectrum of human experience is flattened into a narrow, normative monoculture. This process systematically silences marginalized cultures, endangered languages, localized knowledge systems, and populations lacking digital infrastructure. Understanding the sociotechnical dynamics of knowledge creation, the barriers to digital participation, and the extractive nature of algorithmic curation is essential. Examining these structures provides the foundation necessary to protect meaningful opportunities for human expression and to ensure that participation in collective knowledge remains broadly accessible, authentic, and sovereign.

The Human Record Must Contain More Than the Loudest Voices

The prevailing methodology of generative AI development relies upon the indiscriminate ingestion of internet-scale data. Corpora such as the Colossal Cleaned Common Crawl (C4) and LAION-5B have become the foundational infrastructures upon which modern text and image generation models are built1. However, these datasets are not objective or neutral reflections of humanity. They carry the inherent social and political perspectives, visibility gaps, platform biases, and cultural assumptions of the systems and populations that produced them1. The sheer scale of these datasets—comprising billions of web pages and image-text pairs—often serves to mask profound demographic and representational asymmetries. The dominance of English-language content in these foundational datasets highlights a severe structural skew. English comprises roughly sixty to eighty percent of LLM training data, despite representing only eighteen percent of global internet users and a mere five percent of the world population4. This phenomenon, often termed "WEIRD" (Western, Educated, Industrialized, Rich, Democratic) bias, fundamentally shapes what machine learning models encode as typical human behavior4. Western perspectives, United States-centric references, middle-class personas, and standard English phrasings appear with such high statistical frequency that the models mathematically prioritize them. Conversely, non-Western contexts, rural realities, low-income demographics, and non-English idioms appear rarely and are thus encoded as atypical, erroneous, or statistically insignificant4. This representational asymmetry is severely exacerbated by the automated filtering heuristics employed by AI developers to sanitize web data prior to training. In an effort to remove toxicity, hate speech, and duplicate content from the Common Crawl, developers frequently rely on automated keyword blocklists. For example, the application of the "List of Dirty, Naughty, Obscene, and Otherwise Bad Words"—an open-source project originally designed for autocomplete sanitization—inadvertently erases vast swaths of perfectly benign or culturally significant content from marginalized communities1. Minority language spaces, LGBTQ+ forums, and platforms featuring heavy dialect or African-American Vernacular English are routinely misclassified as offensive or non-standard1. Because toxicity classifiers perform poorly outside of high-resource languages, conservative filtering mechanisms designed to avoid false negatives frequently discard borderline non-English content entirely, resulting in their disproportionate removal from the human digital record4. The drive toward ever-larger language models prioritizes sheer scale over careful curation, a dynamic critically analyzed in the seminal "Stochastic Parrots" framework introduced by Bender, Gebru, McMillan-Major, and Mitchell6. The pursuit of massive parameter counts encourages developers to scrape the web indiscriminately, incurring massive documentation debt8. When datasets become too large to audit post hoc, biased, toxic, or extractive data is permanently baked into foundational models without accountability8. As researchers emphasize, models trained on uncurated, static datasets from the web encode hegemonic views that actively harm marginalized populations, while creating the false illusion of comprehending human meaning8. LLMs merely calculate the probability of word sequences based on training data; they do not possess grounded understanding, context, or the capacity for meaning6. When society begins to rely on these stochastic parrots as authoritative arbiters of historical, scientific, or cultural knowledge, it risks severing public understanding from authentic human experiences8. If the human record is to survive the AI revolution intact and accurately reflect global humanity, it must be deliberately curated to contain more than the loudest, most historically privileged, and technologically dominant voices.

Ten Important Representation Challenges Backed by Research

The barriers to equitable representation in digital archives and AI systems are deeply structural and intersectional. Analyzing research on underrepresentation reveals a complex web of geographic, linguistic, economic, technical, disability-related, and cultural obstacles that systematically exclude diverse forms of knowledge from the digital sphere. Disproportionately represented communities ultimately shape recorded knowledge by establishing an algorithmic normativity—where their specific linguistic and ideological preferences become the default standard against which all other human expression is measured and often found lacking10. The following table synthesizes ten primary representation challenges and their corresponding sociotechnical impacts on collective knowledge.

These barriers demonstrate that lack of representation is not an incidental flaw of digital technology, but a structural reality arising from the composition of the internet itself and the historical inequities of global development4. Economic barriers prevent populations in the Global South from accessing the high-bandwidth hardware required to publish rich media, while disability-related barriers prevent visually or motor-impaired individuals from contributing to platforms that ignore accessibility standards4.

The Divergence of Public Representation and Dataset Extraction

A critical distinction must be drawn between an individual choosing to be publicly represented online and that individual being selected as raw material for an AI training dataset. Participating in a digital community—such as posting on a public forum, publishing an independent blog, or sharing local historical photographs—is fundamentally an act of human connection and social exchange2. It relies on an implicit social contract regarding audience, context, and intent. In contrast, the ingestion of this data into a system like Common Crawl or LAION-5B entirely strips the content of its relational context, commodifying human expression into disembodied tokens and vectors2. The limitations of available evidence regarding individual contributions to AI models further complicate this dynamic. Because model weights are distributed across billions or trillions of parameters, it is nearly impossible for an individual creator to prove that their specific text or image was used to generate a specific AI output9. The provenance traceability of data inside black-box models remains incredibly poor, leaving creators without viable mechanisms for attribution, compensation, or removal once their data has been absorbed2. This opacity forces a non-consensual relationship upon internet users, treating public digital disclosure not as an invitation for human dialogue, but as a mandatory donation to corporate infrastructure. The danger of this paradigm is severe when applied to marginalized groups. Extraction of their digital data frequently leads to misrepresentation, where cultural knowledge is synthesized incorrectly by a model that lacks lived context4. Furthermore, the normalization of scraping subjects vulnerable individuals to forced disclosure, where narratives shared within semi-private or niche public spaces are amplified and distributed globally, often increasing the risk of targeted harassment2.

The Epistemological Value of Multimodal and Multilingual Knowledge

To build a genuinely comprehensive record of human culture, digital repositories must move beyond a monolithic reliance on standard-English text. Different forms of media contribute distinct types of knowledge, capturing dimensions of the human experience that text alone cannot convey. Multimodal data—encompassing text, images, video, subtitles, audio transcripts, and annotated datasets—provides the necessary scaffolding for a deeply contextualized understanding of global societies. Multilingual websites and the preservation of regional vocabulary are paramount to maintaining cognitive diversity. Language is not merely a tool for communication; it is a framework for structuring reality. When local histories, firsthand accounts, and specialist knowledge are forced through the bottleneck of English translation to achieve digital visibility, profound epistemological nuances are lost18. Regional dialects and specialized vernaculars carry embedded historical context, ecological knowledge, and social relationships that resist direct translation18. Audio and video recordings of oral narratives offer unparalleled fidelity to the original human experience. In many cultures, knowledge is traditionally transmitted orally, relying on tone, pacing, and performative context that cannot be captured by text20. The digitization of oral histories allows for the preservation of these subtleties. However, to make these formats accessible and discoverable, they must be paired with accurate subtitles and audio transcripts15. Subtitles and transcripts serve a dual purpose: they bridge the accessibility gap for the Deaf and hard-of-hearing communities, and they provide machine-readable text that allows specialized knowledge to be indexed and referenced by researchers globally15. Images provide a dense, immediate form of cultural documentation, yet their epistemological value is highly dependent on how they are categorized. As demonstrated by the biases found in the LAION-5B dataset, uncurated images often perpetuate visual stereotyping11. Therefore, accessible digital publications must utilize robust alternative (alt) text and annotated datasets. Alt text not only ensures that visually impaired individuals can perceive the content via screen readers, but it also anchors the image with intentional, human-authored semantic meaning, preventing automated systems from misinterpreting the visual context22.

Community-Driven Knowledge Preservation: Case Studies

To counter the homogenizing effects of extractive AI and centralized digital archiving, numerous communities have pioneered voluntary, community-led models of knowledge creation. These initiatives prioritize data sovereignty, ensuring that contributors retain editorial authority and meaningful attribution over their cultural heritage, while building robust, high-quality digital resources.

Case Study 1: The Masakhane Initiative

The Masakhane initiative represents a transformative paradigm in natural language processing (NLP) research, focusing on African languages through a grassroots, participatory framework23. Historically, African languages have been severely underrepresented in global technology due to colonial legacies and a lack of institutional funding, creating a cycle where linguistic resources remain scarce24. Masakhane, which translates to "We build together" in isiZulu, explicitly rejects the "parachute research" model—where Western researchers extract data from communities without providing reciprocal benefits—in favor of African-centric, multi-disciplinary collaboration24. By mobilizing a continent-wide network of linguists, software engineers, and native speakers, Masakhane has successfully built high-quality datasets, such as MasakhaNER, and translation models for dozens of African languages, including Yoruba, Luganda, Swahili, and Shona25. Recognizing the difficulty of finding parallel corpora, the community engaged in "data archaeology," utilizing diverse sources such as translated religious texts from the Jehovah's Witness community to establish baseline models23. This participatory approach ensures that the development of AI tools is aligned with the cultural contexts and ethical priorities of the communities they serve. Furthermore, it demonstrates that "low-resourcedness" is not merely a technical data deficit, but a societal and historical problem that requires human-centered, collective action to resolve25.

Case Study 2: Local Contexts and Indigenous Data Sovereignty

The intersection of digital archiving and Indigenous rights has historically been fraught with exploitation. For centuries, cultural heritage, biological samples, and traditional knowledge were cataloged by external institutions without community consent, often displaying Indigenous peoples in a deficit-oriented manner28. To address this systemic disenfranchisement, the Local Contexts initiative developed Traditional Knowledge (TK) and Biocultural (BC) Labels, which act as digital signifiers embedded directly into the metadata of digital archives28. These labels establish Indigenous cultural authority, communicating provenance, local protocols, and explicit permissions for how historical and cultural data can be accessed and circulated going forward28. This systemic intervention operationalizes the CARE Principles for Indigenous Data Governance—Collective Benefit, Authority to Control, Responsibility, and Ethics13. Unlike the widely adopted FAIR principles (Findable, Accessible, Interoperable, Reusable), which focus primarily on technical data sharing, the CARE principles introduce a people- and purpose-oriented obligation31. They demand that any use of Indigenous data must empower the communities from which it originates, ensuring that knowledge preservation is paired with self-determination, ethical future use, and the active mitigation of harm13.

The Right to Opacity and Ethical Storytelling Framework

To encourage digital participation without treating public disclosure as an absolute obligation, digital repositories must adopt frameworks that respect the autonomy of communities. The indiscriminate scraping of the web operates on the colonial assumption that anything visible is inherently available for consumption and commodification2. This logic violates the safety of communities who rely on the internet for connection but do not wish their localized practices or sensitive histories to be absorbed into corporate models. To resist this, digital theorists and community organizers increasingly draw upon Martinican philosopher Édouard Glissant’s concept of the "right to opacity"29. Glissant argued that the preservation of cultural difference should not require communities to render themselves entirely transparent or legible to dominant, Western systems of knowledge33. In the context of digital archives, the right to opacity defends the strategic withholding of information. Silence or absence in a digital repository does not necessarily reflect a lack of knowledge or a technological deficit; rather, it often represents a deliberate act of refusal to participate in extractive data pipelines that exoticize or misrepresent marginalized experiences14. Honoring the right to opacity requires shifting the focus of digital repositories from infinite data accumulation to relational ethics14. Drawing upon established Oral History Association guidelines and the CARE principles, independent platforms must implement frameworks that protect narrators and prioritize informed consent throughout the entire knowledge lifecycle13.

Inclusive Authorship: Accessible Formats and Low-Bandwidth Publishing

A genuinely equitable knowledge repository must systematically lower the barriers to digital participation. Geographic isolation, economic disparities, and disability frequently intersect, systematically excluding individuals from contributing to the digital record. Creating an accessible publishing ecosystem requires strict adherence to international web standards while simultaneously accommodating users in low-resource or low-bandwidth environments.

Multilingual and Accessible Publication Checklist

To ensure that platforms support inclusive authorship and readership, independent websites must be designed with deep technical empathy. The following checklist synthesizes the Web Content Accessibility Guidelines (WCAG 2.1/2.2 AA) with low-bandwidth web design principles to create a universally navigable digital environment15.

Practical Strategies for Contributors Lacking Technical Expertise

Assuming that individuals inherently possess the digital literacy required to format, upload, and manage web content constitutes a major barrier to authentic representation. Independent platforms must develop asynchronous, low-friction pathways for participation that do not demand technical expertise. First, platforms should facilitate multi-modal contributions. Individuals who are uncomfortable with written text, lack access to reliable keyboards, or speak unwritten dialects should be encouraged to submit oral narratives via low-bandwidth audio formats, such as compressed MP3s or localized voice notes over mobile messaging applications. Community volunteers can then convert these into accessible text transcripts, carefully auditing them to preserve regional dialect and cadence without imposing institutional standardization. Second, the technical onboarding process must be entirely abstracted away from the user. Rather than requiring contributors to navigate complex Content Management Systems (CMS) or understand Markdown formatting, platforms should offer simple, plain-text email submission pipelines. By allowing individuals to send stories, local histories, or specialist knowledge via standard email, the technical burden of web publishing and WCAG compliance shifts from the creator to the platform's archival team. Furthermore, community-led digital workshops can serve as vital entry points, teaching foundational digital literacy while concurrently gathering vital firsthand accounts in a supportive, localized environment27.

Decentralized Discoverability: IndieWeb and Semantic Structures

As the digital ecosystem becomes increasingly dominated by algorithmic search engines and walled-garden social media platforms, independent writing faces a severe crisis of discoverability. Centralized platforms dictate visibility based on engagement metrics, advertising revenue, and algorithmic normativity, often burying nuanced, multilingual, or specialized community knowledge2. Making independent platforms usable and findable requires rebuilding the connective tissue of the decentralized web. The IndieWeb movement provides a robust technical and philosophical framework for this endeavor. By prioritizing principles such as owning one’s domain and retaining absolute control over data, the IndieWeb allows creators to syndicate their content without surrendering ownership to corporate entities41. To facilitate discoverability across the independent web, developers must integrate semantic HTML structures, specifically Microformats2. Microformats2 vocabularies—such as h-entry for blog posts, h-card for author identities, and h-feed for content streams—embed machine-readable metadata directly into the HTML of a webpage42. This semantic structuring allows independent sites to understand, parse, and categorize each other’s content seamlessly, without relying on proprietary APIs controlled by centralized tech monopolies41. Furthermore, implementing the Webmention protocol enables a decentralized conversation network. Webmention is an open W3C standard that allows a website to notify another website when it has been linked to, cited, or replied to43. When an independent author writes a response to an article on a separate site, Webmention automatically sends a notification, allowing the original site to display the response as a comment or citation41. Tools like Bridgy Fed allow these decentralized interactions to cross over into the wider Fediverse, significantly expanding the reach of independent writing while maintaining strict data sovereignty46. By integrating Microformats2 and Webmentions, independent knowledge repositories can build highly visible, interconnected communities of practice that actively resist algorithmic suppression.

Movement Connection: The Sovereign Voices Project

To fulfill the mission of protecting meaningful opportunities for human expression and making participation in collective knowledge broadly accessible, BuildToResist.com must launch a participation initiative that is fundamentally rooted in trust, autonomy, and community sovereignty. This initiative must transcend standard user-generated content models, functioning instead as a collaborative, ethical stewardship of the human record. Initiative Title: The Sovereign Voices Project 1\. Core Philosophy: The Sovereign Voices Project serves as an open invitation to global communities, independent researchers, and grassroots storytellers to archive their local histories, regional vocabularies, and cultural perspectives entirely on their own terms. Recognizing that silence in the digital sphere often stems from a justified distrust of extractive tech platforms, this initiative guarantees that participation remains completely voluntary. Contributors retain absolute copyright, editorial control, and the right to opacity, ensuring their knowledge is preserved without being commodified. 2\. Mechanisms of Participation and Control:

• Low-Barrier Submissions: Contributors can submit original text, audio, or visual work via low-bandwidth interfaces, plain-text emails, or secure messaging protocols. All technical formatting required for WCAG compliance, transcription, and IndieWeb syndication will be handled by the BuildToResist stewardship team, ensuring no individual is excluded due to a lack of technical expertise.
• The "Living Archive" Promise: Unlike static AI datasets that permanently lock in extracted data, the Sovereign Voices Project treats all publications as living documents. Contributors are provided with unique, secure credentials allowing them to edit, update, or permanently delete their contributions at any point in the future, maintaining ultimate sovereignty over their narratives.
• Contextual Metadata Tagging: Borrowing heavily from the Local Contexts framework, contributors will be empowered to apply custom digital labels to their work28. These labels will explicitly state how the knowledge should be used—whether it is open for academic citation, strictly restricted from AI scraping via aggressive robots.txt enforcement, or meant solely for community reading.

3\. Community Outreach and Growth: Rather than relying on targeted advertising or algorithmic growth hacking, the initiative will expand through decentralized trust networks. BuildToResist will establish a "Community Liaison" program, partnering directly with local linguists, oral historians, and digital activists to advocate for the platform within their specific language and cultural communities. Liaisons will run localized, offline workshops demonstrating the value of independent digital preservation, fostering environments where firsthand accounts and specialist knowledge are respected and nurtured rather than extracted. 4\. Meaningful Attribution and Discoverability: Every contribution on BuildToResist.com will be wrapped in robust Microformats2 HTML, ensuring the creator’s h-card (author identity) travels wherever their work is referenced across the internet42. Through the implementation of Webmentions, contributors will be able to see the real-time impact of their writing across the independent web, fostering a deep sense of solidarity and collective resistance against the homogenization of human expression41.

Conclusion

The vast and intricate breadth of human cultures, languages, and experiences cannot be adequately captured by the automated scraping of a commercially driven, heavily filtered internet. As artificial intelligence systems increasingly dictate the boundaries of recorded knowledge, the systemic underrepresentation of marginalized voices threatens to produce a future defined by epistemic collapse and strict algorithmic normativity. Addressing this crisis requires significantly more than simply expanding dataset parameters; it demands a fundamental restructuring of how digital knowledge is gathered, governed, and distributed. By embracing participatory research models, asserting indigenous and community data sovereignty, and respecting the fundamental right to opacity, communities can successfully protect their cultural narratives from extractive tech pipelines. Concurrently, utilizing accessible, low-bandwidth design and decentralized web protocols ensures that independent writing remains highly discoverable without capitulating to the demands of corporate algorithms. The human record is an ongoing, dynamic narrative. To ensure it accurately reflects the genuine diversity of human expression, the digital infrastructure of the future must be consciously built to resist homogenization, placing editorial authority back into the hands of the people who actually live the stories.

Works cited

1. The case of 'Colossal Cleaned Common Crawl' (C4), https\://knowingmachines.org/publications/9-ways-to-see/essays/c4

2. A Critical Analysis of the Largest Source for Generative AI Training, https\://www\.researchgate.net/publication/381197403\_A\_Critical\_Analysis\_of\_the\_Largest\_Source\_for\_Generative\_AI\_Training\_Data\_Common\_Crawl

3. Unmasking LAION-5B: Age, Gender, Race, and Emotion Biases in, https\://arxiv.org/html/2606.23204v1

4. WEIRD Bias: Western Over-representation in LLM Training Data, https\://cyberarctica.com/synthetic-audiences/05\_weird\_bias\_western\_over-representation\_in\_llm\_training\_data.html

5. Training Data for the Price of a Sandwich \- Mozilla Foundation, https\://www\.mozillafoundation.org/en/research/library/generative-ai-training-data/common-crawl/

6. Stochastic parrot \- Wikipedia, https\://en.wikipedia.org/wiki/Stochastic\_parrot

7. On the Dangers of Stochastic Parrots: Can Language Models Be, https\://archive.org/details/stochastic-parrots-3442188.3445922

8. On the Dangers of Stochastic Parrots: Can Language Models Be, https\://s10251.pcdn.co/pdf/2021-bender-parrots.pdf

9. On The Dangers of Stochastic Parrots: Can Language Models Be, https\://selfassuredpaperreads.medium.com/on-the-dangers-of-stochastic-parrots-can-language-models-be-too-big-d08edfc59fab

10. On the Dangers of Stochastic Parrots: Can Language Models Be, https\://www\.researchgate.net/publication/349754361\_On\_the\_Dangers\_of\_Stochastic\_Parrots\_Can\_Language\_Models\_Be\_Too\_Big

11. Unmasking LAION-5B: Age, Gender, Race, and Emotion Biases in, https\://arxiv.org/pdf/2606.23204

12. Auditing and instructing text-to-image generation models on fairness, https\://pmc.ncbi.nlm.nih.gov/articles/PMC12103484/

13. The CARE data principles | Research | Imperial College London, https\://www\.imperial.ac.uk/research-and-innovation/support-for-staff/scholarly-communication/research-data-management/care-data-principles/

14. August, 2026 \- Unboy | Ideas & Community, https\://unboy.org/2026/08/

15. WCAG Compliance 2026: Standards and Requirements, https\://www\.levelaccess.com/compliance-overview/wcag-web-content-accessibility-guidelines/

16. Web design guidelines \- Aptivate, https\://aptivate.org/en/services/resources/web-design-guidelines/index.html

17. AI Training Data Contributes To Its Bias \- Diplomatic Courier, https\://www\.diplomaticourier.com/posts/how-ais-training-data-contributes-to-its-bias

18. PluralizingPoliticalPhilosophy \- OAPEN Library, https\://library.oapen.org/bitstream/handle/20.500.12657/100520/9780198887560\_Print.pdf?sequence=1\&isAllowed=y

19. (PDF) Translation and Race \- ResearchGate, https\://www\.researchgate.net/publication/378021374\_Translation\_and\_Race

20. Informed Consent \- Oral History Association, https\://oralhistory.org/informed-consent/

21. Accessibility in Web Development: Importance & Best Practices, https\://www\.webdevelopmentgroup.com/insights/web-accessibility-and-web-development/

22. Web Content Accessibility Guideline Resources for Designers | WCAG, https\://www\.wcag.com/designers/

23. masakhane-io/masakhane-mt: Machine Translation for Africa \- GitHub, https\://github.com/masakhane-io/masakhane-mt

24. With Collaboration and Support From \- Masakhane, https\://www\.masakhane.io/?ref=criticalplayground.org

25. Participatory Research for Low-resourced Machine Translation \- arXiv, https\://arxiv.org/html/2010.02353v2

26. (PDF) Participatory Research for Low-resourced Machine Translation, https\://www\.researchgate.net/publication/344505986\_Participatory\_Research\_for\_Low-resourced\_Machine\_Translation\_A\_Case\_Study\_in\_African\_Languages

27. AI for social good: Masakhane \- SEWF, https\://sewfonline.com/ai-for-social-good-masakhane/

28. Local Contexts – Grounding Indigenous Rights, https\://localcontexts.org/

29. BLACK QUEER AND TRANS ANTHROPOCENES, https\://ir.vanderbilt.edu/bitstreams/99ee4088-d2f5-46fd-88d9-bf8298e4a5d9/download

30. Earth Science Data Repositories: Implementing the CARE Principles, https\://par.nsf.gov/servlets/purl/10548029

31. CARE Principles \- Indigenous Data Governance \- CASRAI, https\://casrai.org/dictionary/term/care-principles

32. A Manual of Anti-Racist Architecture Education 9782954414508, https\://dokumen.pub/a-manual-of-anti-racist-architecture-education-9782954414508.html

33. Place and Personhood in the Literature of Ikaino \- eScholarship.org, https\://escholarship.org/uc/item/7pr810xn

34. Collective approaches enriching design principles, https\://distributeddesign.eu/wp-content/uploads/2024/04/DDBook\_year6\_.pdf

35. Migrations / Mediations \- Unibo, https\://cris.unibo.it/retrieve/915ae87e-61e6-4101-a4b2-fff701fb7310/Com%20soc%20Musaro%20Peja%20e%20Papastergiadis.pdf

36. UCLA Electronic Theses and Dissertations \- eScholarship.org, https\://escholarship.org/content/qt1g22r8ht/qt1g22r8ht.pdf?v=lg

37. How to Ethically Collect Oral Histories from Communities ... \- Oria, https\://www\.missionoria.com/blogs/how-to-ethically-collect-oral-histories-from-communities-using-digital-tools

38. Considering Legal and Ethical Issues \- National Park Service, https\://www\.nps.gov/articles/000/oral-history-resources-considering-legal-and-ethical-issues.htm

39. (PDF) The CARE Principles for Indigenous Data Governance, https\://www\.researchgate.net/publication/346665037\_The\_CARE\_Principles\_for\_Indigenous\_Data\_Governance

40. Website Accessibility Standards: Build a Program that Sticks, https\://www\.siteimprove.com/blog/website-accessibility-standards/

41. Posts about the IndieWeb, https\://indieweb.org/Posts\_about\_the\_IndieWeb

42. People and Blogs: Brennan Kenneth Brown \- Zachary Kai, https\://zacharykai.net/notes/pb/brennan

43. Social Web Protocols \- W3C, https\://www\.w3.org/TR/social-web-protocols/

44. Webmention \- W3C, https\://www\.w3.org/TR/2016/WD-webmention-20160429/

45. bridgy-fed/templates/docs.html at main \- GitHub, https\://github.com/snarfed/bridgy-fed/blob/main/templates/docs.html

46. bridge to the Fediverse \- Bridgy Fed, https\://fed.brid.gy/docs