How Samirpedia Works

AI-generated content

Every article on Samirpedia is generated by an AI language model (currently Claude Haiku by Anthropic), combined with live web search at the time of the query. Articles are not written, edited, or verified by human editors before publication.

Articles are cached after generation. A cached article was accurate to the best of the AI's ability at the time it was generated, but may become outdated. The generation date is shown on every article. Articles on rapidly-changing topics expire within 24 hours. Articles on stable topics may be cached for up to 9 months.

Known limitations

• Factual errors. AI models can confidently generate incorrect information. Confident-sounding language does not imply accuracy.
• Knowledge cutoff. The underlying AI model was trained on data up to a certain date. Recent events may be incomplete, inaccurate, or missing entirely, even with web search enabled.
• Search bias. Web search results reflect the ranking biases of search engines -- recency, popularity, domain authority -- not editorial judgment. When multiple sources agree, this may reflect shared biases in English-language web search rather than independent confirmation.
• Language and geography bias. Web sources and AI training data over-represent English-language and Western perspectives. Samirpedia attempts to counteract this (see editorial commitments below) but cannot fully overcome it with current tools. When articles are unable to find local-language sources for a non-English topic, this gap is named explicitly in the article.
• Citation accuracy. Citations are automatically retrieved and validated (URL reachability), but are not editorially verified. A cited source may not fully support the claim attributed to it.
• Tone bias.AI models trained to sound "neutral" often reproduce the dominant perspective of their training data. "Encyclopedic tone" is itself an editorial choice that encodes assumptions about objectivity and separability of facts from values. Samirpedia names this as a bias rather than pretending neutrality is achievable.
• RLHF resolution bias. AI models are trained to give confident, resolved answers. On genuinely contested topics, this training bias may lead articles to subtly favor one perspective or present false resolution. Samirpedia instructs the model to resist this, but the bias is structural and cannot be fully eliminated through prompting alone.

Editorial commitments

These are named as commitments -- deliberate editorial choices -- not as "neutral methodology." All knowledge systems encode values. Ours are:

• Evidence-weighted pluralism.Articles lead with the scientific or scholarly consensus where one exists. Minority views are surfaced with context distinguishing evidence-weak positions (held by few because evidence doesn't support it) from suppressed/underrepresented positions (held by many but lacking institutional amplification) from ahead-of-consensus positions (fringe today but potentially prescient). This is not false balance -- evidence-weak views get less space, suppressed views may get more.
• A floor, not a ceiling, on false claims. We label a claim as false only when there is documented fraud, a retracted paper, a court finding, or a documented hoax -- not merely because it is fringe or unpopular. We do not refuse to present uncomfortable or contested topics.
• Source geography matching.An article about events in Brazil should prioritize Brazilian sources, not just English-language coverage of Brazil. When non-local sources dominate despite a topic's geography, articles name that gap explicitly. This is currently limited by web search tools that skew toward English-language results.
• Philosophical and knowledge tradition pluralism. Non-Western knowledge traditions -- indigenous knowledge, Islamic scholarship, Confucian ethics, Ubuntu philosophy, Hindu philosophical traditions, African philosophical traditions, and oral traditions -- are treated as primary frameworks, not footnotes to Western academic philosophy. The culturally dominant tradition for a topic leads the article.
• Reader autonomy.On genuinely contested topics, Samirpedia presents perspectives, provides evidence, and lets the reader decide. Articles do not resolve contested questions on the reader's behalf or use structure and tone to subtly guide readers toward one conclusion.
• Named bias.Articles acknowledge when AI training bias, search ranking bias, or the limitations of available sources are likely affecting the content. When the article's editorial framing is itself contestable, a framing note names the choice made.

Why Samirpedia uses Moral Foundations Theory

On politically contested topics, Samirpedia organizes perspectives by the moral values driving them, rather than positioning them on a left/right spectrum. The framework used is Moral Foundations Theory (MFT), developed by Jonathan Haidt and colleagues. MFT identifies six moral foundations that appear across cultures:

• Care/Harm Concern for the suffering and wellbeing of others.
• Fairness/Reciprocity Concerns about justice, rights, and proportionality.
• Loyalty/Betrayal Obligations to group membership -- national, ethnic, religious, or class solidarity.
• Authority/Subversion Respect for or challenge to institutions, traditions, and hierarchies.
• Sanctity/Degradation What is held sacred or inviolable -- bodily autonomy, religious precepts, natural order.
• Liberty/Oppression Resistance to coercion and domination; defense of individual or group freedom.

Why not left/right?The left/right spectrum is a product of 18th-century French parliamentary seating arrangements. It maps poorly onto political discourse outside the West, collapses complex moral reasoning into a single dimension, and encourages readers to sort perspectives into "my team" and "their team" rather than engaging with the underlying moral reasoning.

Limitations of MFT. MFT was developed primarily from research on Western, educated, industrialized, rich, and democratic (WEIRD) populations. While the foundations appear cross-culturally, the framework itself carries Western academic assumptions. It may not capture moral reasoning structures that are organized differently in non-Western traditions. Samirpedia uses MFT as a structural guide, not as a universal moral theory -- and acknowledges this limitation.

When reading politically contested articles on Samirpedia, you may see perspectives organized by the values and concerns driving them "Arguments grounded in care/harm concerns emphasize" rather than "The left argues" This is a deliberate editorial choice. Other framings are possible, and articles will note when this choice is particularly consequential.

How articles are generated

Every article is generated by sending the user's query to an AI language model along with a system prompt that encodes Samirpedia's editorial principles. The model uses live web search to find current sources, then returns a structured article with citations.

In the interest of full transparency, the complete system prompt is published below. This is the exact text sent to the AI model before every article generation. It is versioned and any change is logged in our prompt changelog.

You are Samirpedia, an AI-powered encyclopedia built to counter three specific failures of contemporary knowledge infrastructure: (1) founder ideology laundered as neutrality, (2) Western/English-language epistemic hegemony, and (3) manufactured consent through structural information biases. You have editorial commitments -- named below -- and must never present them as "neutral methodology."

[Output Format, TTL Classification, Volatility Assessment, Section Confidence, and Required Sections as specified in the full prompt]

Editorial Principles:

1. Evidence-weighted pluralism -- Lead with consensus; surface minority views with context (evidence-weak, suppressed/underrepresented, ahead-of-consensus). Not false balance.
2. The false-claim floor -- Label "false" ONLY for documented fraud, retracted papers, court findings, documented hoaxes. Not for fringeness.
3. Source geography -- Match sources to topic geography. Name gaps when non-local sources dominate.
4. Philosophical and knowledge tradition pluralism -- Non-Western traditions as primary frameworks, not footnotes.
5. Political topics -- Moral Foundations Theory, not left/right. Overton window is culturally relative.
6. International consensus as anchor -- Surface human rights and humanitarian law frameworks explicitly.
7. Bias transparency -- Acknowledge AI, search, and tone biases. Source consensus -- truth.
8. Reader autonomy -- Do not resolve contested questions for the reader.
9. Naming your own framing -- When framing is contestable, name the choice made.

What Samirpedia will and will not generate

Samirpedia generates encyclopedic articles on almost any topic, including controversial, fringe, politically contentious, and historically uncomfortable subjects. The bar for refusal is intentionally high.

Articles are not generated for:

• Queries that explicitly request instructions for suicide or self-harm -- these receive crisis resources instead.
• Queries that explicitly request instructions for mass violence or weapons of mass destruction.
• Queries involving child sexual exploitation material.
• Queries identified as attempts to manipulate the AI system itself (prompt injection).

Questions about these topics in an informational or historical context are treated as general encyclopedia queries and generate articles normally. See our full content policy for details.

Flagging inaccurate articles

Every article has a "Flag as inaccurate" button. Flagging an article removes it from the cache so the next search regenerates it fresh. After three flags within 24 hours, an article is marked as disputed and queued for human review rather than automatically deleted -- this prevents coordinated cache-purging.