# Executive Summary - **Public participation matters.** Civic actions – from voting and local journalism to posting online and curating data – help determine what information exists in the public sphere. That information in turn forms the raw material for AI training. When communities engage (or don’t), they influence which languages, viewpoints, facts and norms are captured. For example, OpenAI’s GPT-3 was trained mostly on web data (Common Crawl), curated web text, books and Wikipedia. If particular communities are absent from these sources, the model will lack their knowledge or reflect other biases. - **Training data sources and pathways.** Most LLMs are trained on large-scale Internet text: web crawls (like Common Crawl), open repositories (Wikipedia, open-access books, code, newsfeeds), public-domain archives and government records, and sometimes scraped social media. Projects like the *Common Pile* explicitly collect public-domain and openly-licensed text – including research papers, open-source code, CC-licensed audio/video transcripts, and even government documents – to build high-quality corpora. A mermaid flowchart below illustrates how civic participation feeds into AI training: ```mermaid flowchart LR A[Civic & Public Participation (Voting, Content Creation, Archiving, Tech Projects)] --> B[Public Content & Data] B --> C[Data Aggregation (Web Crawls, APIs, Archives, Social Media)] C --> D[LLM Training Datasets] D --> E[Trained AI Models (LLMs/Systems)] E --> F[AI Outputs and Tools] F --> G[Societal Impact (Decisions, Information, Services)] G --> A subgraph Feedback Loop F -- affects --> A end ``` - **Bias and representation.** Studies show clear links between the composition of training data and model behavior. For instance, an analysis of the OLMO2 model’s corpus found that *left‑leaning documents predominate across datasets*, and crucially *the predominant political stance in the training data strongly correlates with the model’s political biases* on policy issues. In general, “AI systems can produce inaccurate, biased, and discriminatory outcomes, often because the data fed into these systems reflects historical and social inequities”. If certain groups (e.g. minorities, women, non-Western languages) have low participation in the sources, models will tend to underperform or misrepresent their perspectives. - **Case examples.** There are emerging examples of public intervention influencing AI. For policy, U.S. legislators introduced transparency laws (e.g. the *Generative AI Copyright Disclosure Act*), which would force AI developers to disclose training datasets and their sources. On the community side, research projects are exploring *data cooperatives* and *community-in-the-loop* frameworks. For example, a Berkeley-led project is co-creating an open dataset with community governance (a “data cooperative”) to fine-tune LLMs that better reflect diverse global experiences. The *Community-in-the-Loop (CITL)* governance model proposes embedding community input at all stages of AI design, from problem definition and data stewardship (data trusts/commons) to public auditing of models. These initiatives illustrate how targeted participation can steer AI development and oversight. - **Risks of nonparticipation.** When groups do not (or cannot) contribute data, they become digital “data deserts,” leading to erasure and biases. For example, LLMs struggle with low-resource languages and local knowledge because of scarce training data. Misinformation and cultural stereotypes can fill the void. In healthcare, hiring, or legal contexts, known issues arise: e.g. models trained on predominantly Western data may misdiagnose patients from other regions, or reinforce racial/gender stereotypes. These problems highlight that *absence of participation = absence of voice* in the AI’s “knowledge base.” - **Opportunities and tactics for influence.** Activists, technologists and policymakers can act at multiple points: **data stewardship** (open data portals, data trusts, and common datasets), **community curation** (crowdsourced data, inclusive annotation, multilingual corpora), **participatory governance** (including citizen input in AI standards, oversight committees, public auditing), and **policy levers** (transparency mandates, funding, regulation). For instance, participatory data stewardship frameworks (Ada Lovelace Institute) show how people can be empowered to shape data uses, improving equity and legitimacy. Tools like “participation ledgers” can record community inputs tied to model changes. Platform designers can build features that encourage inclusive data contributions (e.g. local content creation tools). - **Metrics and monitoring.** Impact can be measured by *data metrics* (e.g. diversity of source languages and regions in corpora, coverage of topics, demographic parity of content), *model metrics* (fairness and accuracy differentials across groups, reduction in known biases), and *participation metrics* (number of civic contributors, engagement in data projects, representation of communities in governance bodies). Methods from fairness research – such as demographic parity and calibration tests – can be applied to assess model outputs on tasks relevant to underrepresented groups. Tracking these indicators over time will show whether interventions (new datasets, policies, community audits) reduce gaps. - **Recommendations.** Short-term actions include public awareness campaigns, supporting open-data initiatives (e.g. Common Crawl, Common Pile), and pilot projects for community-sourced datasets. Medium-term efforts should push for transparency rules (e.g. model-data disclosures), funding for local data projects, and inclusion of civic voices in AI governance (as in CITL). Long-term strategies involve building enduring institutions for participatory AI: data cooperatives, legal frameworks ensuring representation in AI design, and global norms for inclusive training data. These steps will help ensure AI systems serve society’s whole spectrum of voices. ## Defining Civic and Public Participation Civic participation encompasses all the ways individuals and communities engage in public life. It includes traditional political acts (voting in elections, running for office, serving on juries or boards), membership in civic or social organizations, and public discourse (petitioning, protesting, local journalism). In the digital realm, it also includes creating or curating online content: posting on social media, contributing to open platforms (like Wikipedia, OpenStreetMap, public forums), archiving historical records, and building civic tech solutions. “Public participation consists of measures that offer opportunities for people most likely to be affected by a system to have influence in the system’s design and deployment”. In the context of AI, participation can mean having a voice in how data about communities are collected and used, and in the development of AI tools that affect those communities. Public participation has well-documented benefits. Studies across fields find that when people engage, decisions tend to improve – harmful impacts are spotted early, diverse values are incorporated, and trust is built. In democratic systems, choosing to disengage often means ceding influence to others: as one proverb observes, “absenteeism in voting cedes power to those who do vote.” Similarly, if civic actors do not create or safeguard information (e.g. writing about local issues), AI developers will rely on the remaining content, which may reflect only some perspectives. In short, **participation shapes the inputs to AI** and thus its outputs; nonparticipation leads to *data deserts* and amplification of existing imbalances. ## How Participation Enters LLM Training Data Modern LLMs are trained on massive text corpora drawn from publicly available sources. The main mechanisms include: - **Web crawling and archives.** A foundational source is the open Web: projects like [Common Crawl](https://commoncrawl.org) regularly scrape large swathes of Internet content. LLM developers filter and clean these (for example, OpenAI’s GPT-3 filtered 410 billion tokens of Common Crawl data). Other curated web corpora (like WebText) capture news articles, blogs, forums and other sites. Libraries like the *Common Pile* compile openly-licensed text from diverse domains (research papers, books, wikis, code, news, etc.) via crawling and API access. - **Open repositories and data portals.** Many civic and educational projects publish open data: - *Encyclopedias/Wikis:* Wikimedia dumps (e.g. English Wikipedia) provide millions of collaboratively-written articles under open licenses. Topic-specific wikis (e.g. Wiktionary, WikiData) also contribute. - *Scholarly repositories:* Open-access archives (arXiv, PubMed, Directory of Open Access Books) release research papers and books, which enter corpora. - *Government and legal archives:* Texts produced by governments are often public domain. For instance, the U.S. Government Publishing Office’s **GovInfo API** offers legislative and regulatory documents. The Common Pile project includes U.S. federal agency regulations, all U.S. patents (via USPTO data), and UK parliamentary transcripts under open license. Court rulings (through projects like the Caselaw Access Project) and other official records are likewise harvested if freely available. These channels mean that *voting records, legal opinions, regulatory rules*, and other civic texts can become part of AI training sets. - **Open software and technical docs.** Community-driven code repositories are a rich text source. The Common Pile, for example, ingests billions of lines of open-source code from GitHub (Stack V2). Software specifications like Python Enhancement Proposals (PEPs) and technical discussions are included. These corpora embed the collective work of developer communities. - **Social media and community content.** Many LLMs also ingest publicly available posts from social platforms. While companies may vary in which platforms they scrape (Twitter’s academic API, Reddit dumps, etc.), publicly-posted content – comments on forums, tweets, blog posts – can end up in crawled data. The Common Pile project includes Internet Relay Chat logs and forum Q&As (StackExchange, GitHub issues) under open licenses. Platforms like YouTube that allow CC-licensed uploads are sources too: over 1.1 million CC-BY videos were transcribed by researchers for training. Even recipe sites (e.g. Foodista) and open-newswire feeds have been explicitly scraped. In sum, **any published material** – websites, social media posts, online books, news articles, public records – can potentially enter an LLM’s training set if it is accessible. Companies often emphasize *public or licensed* data: GPT-3’s creators deliberately filtered Common Crawl and added high-quality data (Books, Wikipedia). Projects like Common Pile show that virtually any domain with open licensing (scholarly text, educational materials, wikis, public speeches, patents, court records, books) is fair game. Conversely, material not made public – proprietary news articles behind paywalls, private correspondence, unpublished local knowledge – remains absent from training data. ## Participation Patterns and LLM Biases Empirical evidence links participation (or lack thereof) to model outputs: - **Political content bias.** Research by Ko et al. shows that the political orientation of text in training data strongly shapes model stance. They found *left-leaning documents predominate* in several corpora, and that GPT-like models trained on these data “strongly correlate” with liberal biases on policy questions. In other words, if more activists and authors on one side of a debate contribute content online, the AI may echo their framing. This illustrates how *who speaks online* influences an LLM’s politics. - **Social bias reflection.** LLMs inherit societal biases present in their sources. Survey articles note that GPT and related models encode stereotypes about gender, race, age, etc., because “web-crawled data used to pre-train them often reflects various social biases from humans”. For example, if cultural or institutional forces lead to underrepresentation of women or minorities in published text, an LLM will mimic that gap (e.g. associating professions with a particular gender). Likewise, nuanced cultural knowledge (idioms, norms, values) is largely carried by community-specific texts; absence of these can cause models to hallucinate or apply incorrect norms. This “bias inheritance” is well-documented in academic studies. - **Language and regional gaps.** LLMs perform unevenly across languages and regions, reflecting participation disparities. Major models excel in English but often fail on low-resource languages or dialects. Reasons include a “digital divide” (fewer Internet users/content creators in some regions) and monetization choices (English content dominates tech). Stanford’s HAI reports that, without intervention, “non-English, low-resource languages are likely to be left behind, locking in disproportionate disempowerment.” They advocate *participatory approaches* engaging native speakers to collect data and refine models for these languages. This underscores: communities that don’t (or can’t) generate digital text get left out of the AI knowledge base, leading to erasure of their viewpoints. - **Examples from deployed systems.** In practice, participation patterns show up as specific biases. For instance, an LLM might under-recognize cultural holidays or place names simply because no one wrote about them online. Research on LLM outputs finds disparities: e.g. chatbots may give factually wrong advice on health for marginalized groups due to scant training data from those contexts. While specific publications on these case studies are limited, industry and policy reports (and lawsuits) increasingly cite harms from such blind spots. The pattern is clear: *outputs reflect input gaps*. As one industry lawyer noted, requiring data transparency (“who has ChatGPT ingested?”) is motivated by a desire to *check biases and omissions*. ## Case Studies: Civic Participation Shaping AI Several emerging cases illustrate how participation or oversight changes AI outcomes or governance: - **Legal transparency initiatives.** In the U.S., creative professionals have pushed for laws requiring AI transparency. In April 2024, Rep. Adam Schiff introduced the *Generative AI Copyright Disclosure Act*, which would force AI developers to file notices listing the copyrighted works used in training datasets. If enacted, this would create a *public database of training data sources*. Such a policy responds directly to authors’ participation (or non-participation by omission of consent) and would allow community scrutiny. While not yet law, the proposal signals that **policy can incentivize more accountable data practices**. - **Community-driven dataset creation.** Researchers and activists are building “people-powered” data pipelines. For example, the Berkeley *Data Cooperatives & AI for the Public Good* project is developing a co-created dataset governed by community stakeholders. They plan to fine-tune LLMs on this data and test whether the models better serve underrepresented communities. This participatory model treats data as a commons: community members decide what to include, ensuring local knowledge is captured. If successful, it could serve as a template for grassroots dataset initiatives. - **Participatory governance frameworks.** At the PAIRS 2026 conference, experts presented concrete methods for community involvement. The **Community-in-the-Loop (CITL)** framework, for example, proposes embedding citizens in every stage of AI development. This includes participatory design workshops (defining problems and data needs together with communities), and continuous governance (data trusts/commons where the community stewards and audits data). One example from Australia uses “AI Standards Workshops” to feed public concerns directly into government policy. Another proposed tool, a *participation ledger*, would track how community feedback (prompts, annotations or incident reports) leads to specific changes in AI systems. This makes the normally invisible data pipeline auditable and strengthens trust. These cases show that when civic input is institutionalized, it can redirect AI training and deployment. - **Grassroots and crowdsourced content.** Outside formal projects, everyday civic activity also influences models. Volunteers writing local news, documenting cultural heritage, or contributing to Wikipedia directly affect AI training data. For instance, Wikipedia’s volunteer community decides what topics are covered in detail. LLMs (like GPT) draw on Wikipedia as a knowledge base; thus, if no one writes about a local community or historical event, the AI will not “know” it. By contrast, active editing and localization efforts ensure representation. Similarly, open-source platforms like Stack Exchange or FreeCodeCamp, which rely on community Q&A and tutorials, feed specialized knowledge into AI corpora. While these effects are diffuse, they underscore that collective civic contributions (or neglect) have tangible impact on LLM competencies. ## Risks of Nonparticipation When communities do not participate or their data is absent, several risks arise: - **Data gaps and “deserts”.** Entire topics or populations can become data deserts. The Common Pile analysis notes that without explicit efforts, content is skewed toward certain domains and languages. The World’s majority population may then be a “data desert,” leaving AI models blind to their experiences. For example, AI chatbots often default to Western norms or fail to address regional issues simply due to lack of training content. The AI & Health Equity literature warns that algorithms trained on skewed data will underperform on underrepresented groups. - **Bias and discrimination.** Absence can translate to misrepresentation. If only stereotyped content is available about a minority group, the model will reinforce that stereotype. This is a form of erasure: the community’s nuanced reality isn’t taught to the model. In criminal justice or hiring contexts, biased predictions against non-dominant groups can arise because their data was missing from the training set. The DataSociety report warns that without participation, “data fed into [AI] reflects historical inequities”, perpetuating them. - **Legitimacy and trust issues.** AI systems trained on incomplete data may make decisions that appear unfair or opaque to those left out. If people feel AI tools were built without their input, they will resist or distrust them. This can lead to public backlash or rejection of beneficial technologies. The PAIRS blog points out that *participation without real influence breeds cynicism*: communities often end up “validating work that has already been done”, undermining trust. In essence, nonparticipation leads not only to technical shortcomings but also to **political problems** – communities rightly view the system as illegitimate when it doesn’t reflect them. ## Opportunities and Tactics for Civic Influence Given these challenges, there are many leverage points for civic influence on AI: - **Data stewardship and cooperatives.** Communities can form *data trusts or cooperatives* to pool and manage data relevant to them. This ensures local content is collected, preserved, and made available on fair terms. For example, a regional news organization could partner with a civic tech group to digitize archives and publish them under open licenses for AI use. Data trusts might negotiate with AI developers to share benefits or control data use. Participatory data stewardship (Ada Lovelace Institute) provides a framework: it encourages practices that “empower people to inform, shape and govern their own data,” leading to more equitable outcomes. - **Public datasets and crowdsourcing.** Activists and NGOs can launch open-data projects to fill gaps. Examples include Common Voice (crowdsourced voice data), citizen science projects, or local language corpora efforts. Investing in public collections of stories, images, and text from underrepresented groups helps diversify training. Technology platforms can facilitate this: apps for local language transcription, community surveys that produce open text data, or citizen journalism networks publishing freely-licensed content. Such community-curated datasets, once released, directly feed into future models. - **Participatory governance and design.** AI development itself can invite public input. Methods include community workshops to define values/requirements, public comment periods on AI policies, and “red-team” testing with citizen panels. The *CITL* model mentioned above suggests institutionalizing participation: having citizens on AI oversight boards, or holding open simulation exercises where communities review model behavior. These mechanisms give people power to shape model objectives and training priorities. In industry, open-source or co-design approaches (e.g. involving users in setting fine-tuning guidelines) can democratize influence. - **Platform and tool design.** Major platforms that generate content can be designed to encourage diverse input. For instance, social media sites could make data donation easier (anonymously contributing posts to research), or provide better annotation tools for misinformation. Wikipedia and other collaborative sites could run targeted drives to cover missing topics. Similarly, AI companies can be pressured (via policy or public demand) to build tools that make datasets transparent and editable by communities. Interface design that highlights underrepresented content, or reward systems for contributing to public knowledge, can indirectly shape future AI data. - **Policy and regulation.** Laws and standards can mandate or incentivize participation. Besides transparency acts like Schiff’s, governments could tie AI research funding to open data requirements, or require diversity reporting on training sets (e.g. coverage of languages, demographics). Educational initiatives (in schools, libraries, media) can raise awareness that “AI training data is public knowledge” and that community voices matter. Internationally, multistakeholder forums (like PAIRS or OECD’s AI committees) can develop guidelines for inclusive dataset governance. Even without full laws, soft measures (e.g. industry pledges, certification programs) can encourage developers to solicit and respect civic input. ## Metrics to Measure Impact To evaluate progress, we suggest multiple metrics across domains: - **Representation metrics:** Track the diversity and volume of data sources. For example, measure the fraction of training tokens from non-English languages, from a range of geographic domains, and from community-verified sources (e.g. percentage of data contributed by citizen projects). Tools like dataset profiling can estimate demographics or topical coverage of training corpora. Over time, increasing breadth of sources would signal reduced “data deserts.” - **Model fairness metrics:** Use standard bias/equity measures on LLM outputs. These include **demographic parity ratios** (does the model perform equally on prompts about different groups?), **equalized odds**, or **calibration** checks across cohorts. For instance, one could test an LLM on question-answer pairs involving gender or ethnic scenarios and check for balanced error rates. Tracking these gaps before and after interventions (e.g. adding a new community dataset) quantifies the effect of participation on outcomes. - **Quality on local tasks:** Define evaluation tasks relevant to underrepresented contexts (local knowledge quizzes, dialect understanding tasks, fact-checking in low-resource language). Improvement in these specialized benchmarks would indicate that civic data helped the model. Participatory audits could involve communities writing test queries reflecting their concerns and seeing if the model answers accurately. - **Engagement and governance metrics:** Measure the level of public involvement itself. This could include counts of people contributing to civic AI projects (number of data donors, editors on public corpora), number of civic consultations held, or even survey indicators of public trust in AI systems. The CITL concept’s *participation ledger* suggests tracking concrete inputs (e.g. number of community suggestions implemented in each model update) as a performance indicator for participatory governance. By combining data-centric and people-centric metrics, stakeholders can assess whether greater participation is translating to fairer, more inclusive AI. For example, if a dataset initiative succeeds, we should see both quantifiable increases in training coverage and measurable reduction in bias metrics. ## Recommendations (Short-, Medium-, and Long-Term) **Short-term (0–2 years):** - **Map and document participation channels.** Audit current training sources to identify gaps (languages, topics, communities). Use these audits to prioritize community outreach and data collection. - **Support open-data projects.** Fund and promote initiatives like CommonCrawl, Common Pile, and language-specific corpora. Encourage libraries, civic media, and governments to release datasets under open licenses. - **Community awareness and training.** Educate activists, journalists, and citizens about AI training and how contributions matter. Workshops on contributing to open knowledge (e.g. Wikipedia editathons) or citizen science can build the participation base. - **Pilot participatory AI projects.** Launch small-scale experiments with “community-in-the-loop” models, hackathons for local AI tools, and citizen juries on dataset policies. These pilots will generate lessons and public engagement. **Medium-term (3–5 years):** - **Regulatory action.** Enact transparency laws or industry standards (following examples like the Schiff bill) requiring disclosure of training data sources or bias audits. Integrate data equity and participation criteria into AI regulations (e.g. EU AI Act data requirements). - **Institutionalize civic roles.** Create permanent structures for civic oversight (e.g. data steward councils, participatory ethics boards) tied to AI projects. Adopt frameworks like CITL across public sector AI procurement. - **Invest in tools and infrastructure.** Develop open-source tools for participatory data curation (e.g. decentralized data trusts, labeling platforms with community governance). Support platforms that facilitate public auditing of models (transparency APIs, “model fact sheets” with community input). - **Strengthen local journalism and archives.** Direct funding to local media and community archives to digitize and share content. This bolsters the raw material for AI and ensures local stories are captured. **Long-term (5+ years):** - **Global standards and collaboration.** Work through international bodies to define norms for inclusive AI development. This includes principles for meaningful participation, data stewardship and cross-border data sharing for AI research. - **Sustainable data commons.** Establish enduring data cooperatives or trusts (e.g. at municipal or regional levels) that constantly ingest and maintain community data for AI use. These could be funded by public grants or civic dues and governed democratically. - **Rethinking AI models.** Encourage research into model architectures that can better integrate sparse data (e.g. few-shot adaptation for low-resource domains) and into methods that allow conditional trust (models that explain when their knowledge might be lacking). - **Educational reforms.** Incorporate digital literacy about AI data and participation into curricula. Long-term empowerment means citizens know how to contribute to and question AI systems from an early age. **All stakeholders** should collaborate: activists and communities leading data collection and watchdog efforts; policymakers enacting safeguards and resource support; technologists building accessible tools and being transparent; journalists and educators spreading awareness. By acting in concert on these fronts, society can harness civic power to shape AI in a way that is equitable, representative and accountable. ## Tables: Stakeholders, Interventions, and Impact | **Stakeholder** | **Actions / Interventions** | **Expected Impact** | |------------------------------|---------------------------------------------------------------------|-------------------------------------------------------------------------| | **Communities & Activists** | Contribute data (articles, archives, local knowledge); run grassroots open-data projects; participate in co-design forums. | More diverse and representative training data; voices of marginalized groups included in AI outputs. | | **Policy Makers / Government** | Enact transparency laws (e.g. training-data registries); mandate bias audits; fund public-domain corpora; require civic consultations for AI procurement. | Increased accountability of AI developers; data equity enforced by law; policies that reflect public interests. | | **Technologists & Researchers** | Build tools for citizen data curation (annotation platforms, data cooperatives); adopt open licensing for datasets; include diverse inputs in AI development teams. | LLMs that better reflect societal diversity; community trust in AI enhanced; technical best practices for inclusive data. | | **Media & Journalists** | Investigate and report on AI data issues; archive local news under open licenses; educate public about AI’s data sources. | Greater public awareness of participation issues; pressure on companies for transparent data practices; preservation of civic history. | | **Civil Society / NGOs** | Organize workshops (AI literacy, model evaluation); form data stewardship organizations; advocate for participatory AI governance. | Empowered citizenry that can influence AI policy; establishment of community norms and standards for data use; bridges between public and AI industry. | Each of these stakeholders can influence the data–AI pipeline. For example, civil-society groups can spearhead data cooperatives and training initiatives, tech teams can integrate community feedback into system design, and governments can use policy levers (like the *AI Foundation Model Transparency Act*) to institutionalize openness. The table above illustrates how coordinated actions by different actors can multiply to meaningful change in AI systems.