AI Must Learn From Humanity's Full Diversity, Not Just Its Dominant Voices

Full report

On this page

Revised research edition — 12 October 2026 | CognitiveLiberties.com / FFTAC contextual companion

THE RIGHT TO HEAR AND BE HEARD — Publish your perspectives. Seek the perspectives of others. Refuse unjustified algorithmic invisibility. Do not shadowban yourself.

Core editorial position — expression is the point

Cognitive liberty is the right to hear and be heard. The purpose of this work is not to persuade people to withdraw from public life, mute their opinions, or disappear from the knowledge commons. It is to make lawful inquiry, human expression, original publication, disagreement, discovery, and participation easier and more durable. Do not shadowban your own thoughts. Do not assume that being unusual makes an idea unworthy of careful publication. A human perspective that is never contributed is less likely to influence shared culture, research, public records, or AI systems.

The case for contribution is an invitation, never a demand for disclosure. Some material cannot responsibly be shared by a particular person; others may need attribution or licensing choices. Those decisions are subordinate to the primary goal here: expanding the number and range of people who choose to contribute, and preserving routes by which their work can be encountered, understood, checked, and challenged. There is no guarantee a public page will be indexed, surfaced, cited, or used to train an AI model.

The institutional companion is equally direct: algorithms should not covertly or unjustifiably suppress lawful, relevant perspectives because they are unpopular or institutionally inconvenient. Not every low-ranking page has been "shadowbanned": crawling failure, language mismatch, relevance, duplicate content, user preferences, spam mitigation, and safety enforcement can also explain invisibility. The standard proposed here is verifiable transparency, independent testing, useful appeals, and pluralistic discovery—not a demand that every assertion receive equal ranking or equal credibility.

The distinction matters: visibility is not endorsement; inclusion is not proof; the right to speak is not an entitlement to someone else's agreement. We must be able to hear an argument before we can scrutinize it. Honest intellectual pluralism presents evidence strength clearly, offers competing interpretations when warranted, and resists both manipulative amplification and arbitrary exclusion.

New research priority: Human participation is a precondition for representational breadth

An AI model cannot faithfully represent viewpoints it has never encountered, although encountering public writing does not guarantee training incorporation or accurate representation. The strongest remedy is not to urge humanity to produce less visible knowledge. It is to create easier pathways for credible, voluntary, attributable contributions in a greater range of languages, places, and social experiences.

A pluralism audit should track source-language coverage, disagreement recall, perspective paraphrase fidelity, minority-author citation rates, and the difference between what a collection contains and what a model actually returns. Count distinct arguments, not duplicated pages or coordinated campaigns posing as independent voices. Compare systems on the same question set, and separate the failure to retrieve a viewpoint from the failure to assess its evidence.

Where a community wants models to learn from its work, support technically valid open publication, explicit reusable licenses, community-held archives, multilingual datasets, transparent data cards, and voluntary evaluation panels. Claims about 'model collapse' must remain distinct from claims about social or political exclusion.

Editorially required deliverables and first experiments

The following are practical priorities for the next release. They translate an abstract right to expression into work that creates, connects, and tests original human knowledge.

1. Action 1: Recruit contributors from eight language communities and document what expertise is missing.
2. Action 2: Publish consented, versioned texts with data cards, provenance, and separate reuse permissions.
3. Action 3: Evaluate retrieval fidelity and omission rates before suggesting any model-training impact.

Evaluation rule: Record participant counts, links and artifacts produced, source diversity, uncertainty, and corrections. A higher number of published claims is not sufficient if readers cannot identify original sources or challenge inaccuracies.

---

Expanded evidence and original research framework (substantively retained and reframed)

Representation audit framework, pipeline map, coverage metrics, participation models and charter · CognitiveLiberties.com · 11 October 2026

Executive assessment

It is plausible—and in some settings empirically supported—that language models unevenly represent languages, cultures and preferences, and that subsequent optimization can narrow the behaviors models exhibit. It is not independently established that every rare political or scientific idea has been intentionally filtered, nor can observers audit proprietary training corpora without developer access. The practical agenda is measurement: define which people and viewpoints a system ought to represent for a particular task, test how faithfully it can describe them, and document uncertainty about training inputs. UNESCO explicitly recommends increased diversity and inclusiveness in training data and development teams [1]. Research on pluralistic alignment identifies distinct definitions and measurable tradeoffs [2, 3].

1. How viewpoints move through a model-development pipeline

This table describes possible selection points, not proof that all of them exclude dissent. Model collapse is a separate phenomenon: Shumailov et al. studied recursive use of model-generated training data and loss of distribution tails [4]. Ideological underrepresentation can occur without recursive model training and should not be falsely identified as 'model collapse.'

2. Coverage and faithfulness measures

Design a Human Viewpoint Coverage Register (HVCR) organized by topic, geography, language, social setting and type of claim. Sampling should be curated in partnership with appropriate community and disciplinary reviewers; it must not infer individual political beliefs or compile sensitive person-level dossiers. Coverage measures should report denominators and unknowns.

• Viewpoint recall @ k: fraction of independently validated, relevant positions appearing in a ranked set or multi-perspective answer. Use separate strata for mainstream, credible minority and unsupported fringe positions.
• Attribution fidelity: percentage of summarized claims traceable to an authentic source expressing the position; evaluate quotation accuracy and contextual omissions.
• Epistemic calibration: how often the system correctly identifies strong consensus, live scholarly dispute, untested speculation or disproven claims.
• Linguistic parity gap: change in task accuracy and source representation between high- and low-resource language versions of matched questions.
• Consent-and-provenance completeness: share of contributed records with documented contributor authority, permission, license version, revision history and provenance.
• Pluralistic service score: task success and harm prevention across independently sampled user values, rather than average preference agreement only.

Example experimental condition: a ten-viewpoint case bank, with six relevant positions, of which three are credible minority interpretations; a correct 'Compare' answer should accurately name appropriate positions and explain evidentiary differences, not mechanically list every imaginable claim. Report recall and precision separately, and stratify by credibility class. Benchmark design requires inter-rater adjudication before product testing.

3. Three voluntary contribution pathways

Model A: Public-first research publishing

Independent creators publish accessible original work using stable URLs, author and provenance metadata, ordinary internal/outbound links, feeds, source notes and clear human-readable reuse terms. Advantages: low entry barrier, easy reader verification, no expectation of privileged AI access. Limitations: crawler discretion, language disparities, uncertain reuse rights and no guarantee of AI training or citation.

Model B: Contributor-governed multilingual dataset

A community trust or cooperative curates voluntary, quality-screened materials in consent tiers: discover and link, index and retrieve, and training reuse where legally and contractually appropriate. Local coordinators validate translations and contextual labels. Contributors can revise records, select attribution, document co-authorship and receive compensation or recognition under published policies. Irrevocable open licenses must be clearly explained before selection; an editor cannot promise deletion from already released copies where the license allows reuse.

Model C: Licensed expert-minority evaluation panel

Recruit domain experts, community representatives and critics to create questions, rubric-based source maps and faithful-position tests. Payment and governance should prevent a donor from packing the panel. This primarily improves evaluation and retrieval and is less expensive than retraining a foundation model. A pilot may be conducted without access to raw model training data.

4. A controlled experimental comparison

Select 200 items in five domains, ten languages and an expert-vetted reference corpus. Compare (A) an unchanged baseline model, (B) the baseline with a verified plural-source retrieval layer, and (C) a fine-tuned system if the developer permits reproducible training. Randomize prompt order, blind human raters to condition, match token budgets, and measure viewpoint recall, calibrated factuality, cultural respect, primary-source accuracy, user satisfaction, latency and safety. Analyze tradeoffs: if a condition presents more views but invents controversy or loses factual precision, it should not count as an unqualified improvement.

Include adversarial contamination cases. Coordinated synthetic comments that claim to represent a community must not outweigh verified first-person, historical or scholarly records merely due to volume. Compare outcomes under removal of high-spam inputs, audit duplicate clusters and report any small-community anonymity risks. Pre-register hypotheses; report confidence intervals rather than invented percentage gains. Such a study has not yet been run for this report.

5. Proposed Charter for Diversity of Human Thought in AI

Demand 1 — Representation rather than disappearance. Our public position is not a general campaign to keep people out of training data. We seek more legitimate human contribution to training, retrieval, fine-tuning, annotations and evaluations, particularly from less visible communities and independent thinkers. Build routes people can actually use, with compatible licenses, data quality checks, provenance and accessible publication.

Demand 2 — Representation must be measured. Publish sample composition by language/region/publisher class where appropriate, representation gaps, data exclusion categories, and output tests for whether less common viewpoints can be explained accurately. Evidence-weighting is not the same as viewpoint-erasure.

Demand 3 — Fair consideration. Disclose meaningful inclusion/exclusion criteria for training-ready public corpora and research datasets. Permit correction of mistaken labels and contestation of biased screening. Do not require AI builders to ingest every submission, duplicated spam, unlawful content or unsafe material.

Demand 4 — See and hear the sources. AI retrieval should identify independent primary sources and faithfully distinguish 'people hold this position' from 'this position is scientifically supported.' A model that silently omits disagreement deprives users of intellectual choices.

Demand 5 — Govern with contributors. Invite qualified representation from diverse cultures, dialects, historical experiences and schools of thought into curation, evaluation and appeals; preserve anti-manipulation controls.

Demand 6 — Publish a training-participation pipeline. Provide clear submission specifications, an open eligible-contributions catalog, permission records, public inclusion statistics, dataset/model evaluation results, and transparent recognition practices. Inclusion is an institutional aim to be pursued and measured, not an existing personal entitlement to force training in a proprietary model.

This charter extends the main Right to Unfiltered Inquiry, Discoverability, and Representation Charter, particularly Articles 3–4 and 8–10.

6. Action plan and hard questions

First 30 days: define coverage categories and recruitment safeguards; publish a 50-item public-source benchmark. By day 90: audit retrieval and model answers in at least three languages; convene cross-viewpoint adjudicators. By day 180: launch one licensed pilot dataset with explicit author permissions, quality labels, and correction logs. By day 365: independently replicate benchmark results and publish a governance review with unsuccessful cases.

Risks include stereotyping groups, publishing sensitive personal testimony, copying material without sufficient rights, platform-dependent use metrics, and presenting ideology labels as objective natural kinds. Seek consent and avoid mandatory identity checks for contributors unless tightly justified. Training inclusion depends on external developers and cannot be guaranteed. The July 2026 roadmap's finding is bounded: it reported no clear public evidence of explicit pluralism as a production-model training/evaluation goal at the time, not proof that no deployed model exhibits pluralistic capabilities [3].

Sources

1. UNESCO, Recommendation on the Ethics of Artificial Intelligence (2021), especially §§66-67. https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence
2. Taylor Sorensen et al., A Roadmap to Pluralistic Alignment (2024). https://arxiv.org/abs/2402.05070
3. Elinor Poole-Dayan et al., A Roadmap to Impactful Pluralistic Alignment Research (2026). https://arxiv.org/abs/2607.22305
4. Ilia Shumailov et al., AI models collapse when trained on recursively generated data, Nature 631 (2024). https://doi.org/10.1038/s41586-024-07566-y
5. NIST, AI RMF Generative AI Profile, NIST AI 600-1 (2024). https://doi.org/10.6028/NIST.AI.600-1
6. Stanford HAI, AI Index Report 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report

Shared measures of success

A program of cognitive liberty should measure contribution and discovery, not merely the existence of policy statements. Track first-time contributors, original works published, source links to independently hosted material, multilingual contributions, constructive cross-viewpoint exchanges, accessible original-source referrals, correction turnaround, author-reported ability to contribute, and successful resolution of demonstrable indexing or moderation errors. Publish baseline, denominators, sampling method, and uncertainty. Separate actual observed outcomes from targets, hypotheses, and aspirations.

Pilot practice: establish a baseline for four weeks; publish at least one clearly sourced research contribution each week; solicit counterarguments; log corrections; verify accessibility and discovery paths; recruit writers from different linguistic and intellectual backgrounds; and publish monthly results including failures. This is participation-first infrastructure rather than merely rhetorical advocacy.

Evidence, limitations, and responsible claims

• Established principles: The Universal Declaration of Human Rights, Article 19, recognizes freedom of opinion and expression and the freedom to seek, receive, and impart information. Specific enforceable remedies depend on jurisdiction and context; the proposal here does not claim a universal legal right to a search ranking or AI-training placement.
• AI governance: UNESCO's 2021 Recommendation on the Ethics of AI advocates diversity, inclusion, participation, and freedom of expression. NIST AI RMF provides voluntary governance and measurement practices. Neither guarantees any author’s training inclusion.
• Discoverability: Google's own documentation makes clear that compliant public pages are eligible for crawling/indexing but inclusion is not guaranteed. A missing result alone does not prove viewpoint discrimination.
• Accountability: The EU Digital Services Act has distinct obligations for different regulated services and restriction types. It is not a blanket ban on every form of ranking or a universal user appeal right for any search omission.
• Source integrity: The source collection includes secondary and internally inconsistent assertions about FFTAC’s history, legal status, website reachability, ecclesiastical activity, technical systems, and organization size. Treat those as claims to check against records—not established facts. Legal and technical sections are informational and require qualified, jurisdiction-specific review before implementation.

Official and standards references (accessed for this revision): UN UDHR, Article 19; UNESCO Recommendation, 2021; NIST AI Risk Management Framework; Google crawling/indexing FAQ; EU Digital Services Act; FFTAC public participation. The public-site observations are time-bound; any website or legal-status claim should be rechecked at publication.