Notes — Jimmy Wales on Wikipedia, Neutrality, and the Governance of Knowledge

Notes on Jimmy Wales in conversation with Lex Fridman — Lex Fridman Podcast #385, 18 June 2023.

Four questions [Adler frame]

Q1 — What is it about as a whole? A three-hour-plus conversation with Wikipedia’s co-founder, structured around three strands. First, mechanism: how Wikipedia actually decides what belongs in an article (neutral point of view, undue weight, source quality) and whether an article should exist at all (notability, verifiability, biographies of living persons). Second, governance under pressure: recurring accusations of political bias, two specific controversies that were widely misread as ideological censorship (a ‘recession’ definition dispute, the Twitter Files deletion debate), and Wikipedia’s stated refusal to comply with government takedown demands. Third, and increasingly urgent given the interview’s June 2023 timing, Wikipedia’s relationship to large language models — as training data, as a ‘grounding’ source, and as a competitor for the reader’s attention.

Q2 — How is it argued? Almost entirely through worked example and personal anecdote rather than abstract policy statement. Wales illustrates undue weight with a small-town councilman’s DUI paragraph, notability with a BIC pen and his own mother, false neutrality with the moon-is-made-of-cheese and flat-Earth articles, and community health with Emily Temple-Woods’s path from teenage vandal to Wikipedian of the Year. Fridman brings his own experience as a Wikipedia subject — a contested MIT/employment detail, his birthplace — as a live test case, which gives the conversation an unusually concrete anchor for what ‘getting a biography wrong’ actually feels like from the inside.

Q3 — Is it true, in whole or part? Wales’s account of mechanism — how deletion debates, source deprecation, and BLP protections actually work — is a first-person description from the person who built and still steers the system, and the specific incidents he cites (the recession-definition edit, the Twitter Files ‘snowball close’) are checkable against Wikipedia’s own public edit history. His rebuttal of the political-bias accusation is less independently verified: he reports that accusers ‘normally struggle’ to produce examples, which is his own characterisation rather than an audited count [?]. The claim that the Wikimedia Foundation received zero content-change requests from the FBI is stated as fact-checked with Foundation counsel before the interview, but remains a single-source claim [?source]. The demographic-bias material (novelist coverage length by gender, editor composition skewing male) is presented as findings Wales has seen rather than studies he cites by name [?source].

Q4 — What of it? The wiki gains its first sustained, practitioner-level account of a working knowledge-governance system: the neutral-point-of-view / notability / verifiability triad that decides both what Wikipedia says and what it says nothing about, and the ad-free funding model as a structural choice that removes the engagement-optimisation incentive discussed elsewhere in the wiki with respect to Facebook and Twitter. It also captures an early (2023) first-person view of LLM grounding from the person whose content is the actual grounding corpus most models draw on — useful counterpoint to purely technical treatments of the same idea.

Glossary

Neutral point of view (NPOV) — Wikipedia’s foundational content policy: articles represent significant viewpoints on a topic in rough proportion to their prominence in reliable sources, without the encyclopedia itself taking a side. Not the same as giving every view equal airtime. [§ Wikipedia’s political bias]

Notability — the threshold test for whether a subject can sustain a standalone article at all, distinct from whether the subject is personally important to someone. Turns on whether enough independent, reliable source material exists to write from — Wales says he prefers thinking of it as closer to verifiability than to worthiness. [§ Number of articles on Wikipedia]

Verifiability (Wikipedia sense) — whether a claim can be traced to a published, reliable source, as opposed to being merely true or merely believed. Distinct from the machine-learning sense of verifiability used elsewhere in this wiki (see Verifiability, on checkable ML task structure) — same word, unrelated concept. [§ Wikipedia pages for living persons]

BLP1E — Wikipedia shorthand for ‘biography of a living person, notable for one event’: the case of a private individual (commonly a crime victim) who has no ongoing public role but generated a burst of source coverage from a single incident. Treated as a standing danger zone because the subject cannot participate in correcting the record and there is rarely enough material for an actual biography. [§ Wikipedia pages for living persons]

Undue weight — giving a minor or single-sourced fact disproportionate space or prominence relative to its actual significance in a subject’s life, as when a short biography devotes a full paragraph to a family member’s unrelated arrest. [§ Wikipedia pages for living persons]

Deprecated source — a source (e.g. the Mail Online) that Wikipedia editors have agreed is generally unreliable for factual claims, without being formally banned; contributors are expected to seek a better source rather than cite it, decided article by article rather than by blanket rule. [§ Wikipedia pages for living persons]

Grounding — in the language-model sense used in this conversation, tying a generated claim back to a checkable source, the same underlying discipline Wikipedia enforces through citation. Wales treats Wikipedia as a natural grounding corpus for LLM output. [§ ChatGPT]

Snowball close — a rapid, informal end to a Wikipedia deletion debate once it is clear the outcome has no realistic chance of going the other way (‘doesn’t have a snowball’s chance in hell’), closed by a single administrator rather than run to a full discussion period. [§ Twitter files]

Ignore all rules — one of Wikipedia’s foundational principles, stating that a contributor should not be blocked or discouraged over a minor procedural error (like footnote formatting); deliberately worded to provoke attention rather than to license genuine rule-breaking. [§ Building Wikipedia]

Key claims by section

Origin story of Wikipedia / Design of Wikipedia [§ Origin story of Wikipedia, § Design of Wikipedia]

Wikipedia followed a failed two-year precursor, Nupedia, which required a seven-stage academic review process; the first article published under that process had to be pulled for plagiarism within days, despite the review. Wales found the review process personally intimidating even as an academic himself, writing a biography of a Nobel laureate he knew well.
Wikipedia (a side project, launched from fear the Nupedia academics would reject the idea outright) produced more content in two weeks than Nupedia had in two years, because contributors could act immediately rather than submit for review.
Early technical choices were improvised and imperfect but did not block adoption: the site ran on a single-file Perl wiki script with no real login system at launch, and its link syntax (square brackets) could not be typed on a standard German keyboard — German Wikipedia nonetheless became one of the largest editions.
Wikidata now lets a single data point (e.g. a city’s census population) update automatically across every language edition that links to it, realising a structure early contributors discussed but could not build in the software of the time.

Number of articles on Wikipedia / Wikipedia pages for living persons [§ Number of articles on Wikipedia, § Wikipedia pages for living persons]

English Wikipedia held roughly 6.6 million articles and 58 million total pages as of May 2023. Wales still finds the scale startling despite knowing the figures, and recalls when 100,000 articles felt large by comparison with existing single-volume reference works.
Notability is a limiting principle, not a value judgment: a generic BIC pen can plausibly sustain an article (as a mass-produced object with a documented history); a specific pen in someone’s hand cannot, because nothing citable exists about that particular instance. The same logic makes an encyclopedia entry about Wales’s own mother — privately notable to him, but with no public source material — inappropriate.
BLP1E marks the sharpest edge of this principle: a private crime victim who generated a burst of coverage from one event should not get a standalone biography, because there is no ongoing public record to write from and no way for the subject to participate in correcting errors.
Undue weight governs proportion within an article that does exist: a long paragraph on a minor local politician’s son’s unrelated DUI, in an otherwise short biography, is disproportionate regardless of whether the underlying fact is true. Wales treats a dedicated ‘controversies’ section as poor practice for the same reason — it invites editors to go looking for material to fill it, rather than integrating any genuine issue into the article’s overall context.
Source quality is handled by editorial judgment rather than blanket bans: the Mail Online was ‘deprecated’ (flagged as generally unreliable) rather than banned, meaning editors are expected to look for a better source rather than being forbidden from ever citing it.
Community health — how the community responds to a good-faith correction or complaint — matters as much as any specific rule; Wales cites a ‘Did You Know?’ front-page item flagged as a possibly racially insensitive analogy, resolved through discussion rather than deletion or accusation.

Wikipedia’s political bias / Conspiracy theories [§ Wikipedia’s political bias, § Conspiracy theories]

Wales rejects the claim that Wikipedia leans left, arguing that specific examples are rarely produced when he asks accusers for one, and that persistent bias is more likely in obscure, non-political subject areas (his example: unusually favourable coverage of Japanese anime, because only enthusiasts bother to write about it) than in high-traffic political topics.
False neutrality is explicitly rejected as a goal: Wikipedia’s article on the moon does not present ‘some say rock, some say cheese’ as equally weighted positions, and the flat-Earth article treats the belief as a cultural phenomenon rather than a live scientific alternative.
Genuinely contested empirical questions are handled differently from settled ones: Wales cites Wikipedia’s mask-efficacy coverage, which reflects a real, mixed evidence picture, as a case where the article should — and in his reading, does — represent that the evidence is complicated rather than picking a side.
Wales’s own working distinction is not left versus right but ‘the party of the kind and thoughtful’ versus ‘the party of the jerks’: he gives his own hypothetical example of a Catholic priest and a Planned Parenthood activist collaborating productively on the abortion article, provided both approach it in good faith.
Volunteer demographic skew (Wales estimates 80%+ male, disproportionately college-educated and tech-oriented) produces coverage gaps distinct from deliberate bias: he cites early-childhood-development topics and shorter articles on female prize-winning novelists as examples where the gap in coverage reflects who is writing rather than any intentional slight.
Two controversies Wales says were misread as ideological suppression: a routine restructuring that moved the standard ‘two quarters of negative growth’ recession definition further down an article (reported in the press as changing the definition to help the Biden administration), and a Twitter Files deletion nomination that was ‘snowball closed’ within hours as a purely editorial, not political, decision.

ChatGPT / ChatGPT vs Wikipedia [§ ChatGPT, § ChatGPT vs Wikipedia]

Wales identifies ChatGPT’s central flaw as fabricating plausible-sounding but false content because the model is optimised to be agreeable rather than to track truth — demonstrated when he asked it for examples of a white athlete compared to a car, and it invented three specific, entirely fictional quotes.
He proposes ‘grounding’ — tying generated claims to Wikipedia’s sourced text — as the fix, and describes an internal weekend demo (built by Wikimedia’s head of machine learning) that used an LLM to identify relevant Wikipedia entries and then answered a question only from their content, reducing fabrication.
Wikipedia’s own visible ‘the neutrality of this section has been disputed’ banners are offered as a model of transparency current LLMs lack: Wales half-jokingly wishes mainstream news outlets carried equivalent disclosure, and speculates an LLM could summarise a Wikipedia talk-page dispute directly into such a banner.
Wales is unconcerned that people might ask ChatGPT rather than visit Wikipedia directly, provided outputs cite Wikipedia as the source — reasoning that most factual questions will still surface a Wikipedia citation at the top, preserving recognition and donations, as with search engines today.
Editorial policy on AI-assisted edits places responsibility on the human contributor: using ChatGPT to check grammar in a non-native language is fine, but the editor is accountable for anything factually wrong that a model introduces.

Twitter files / Government and censorship [§ Twitter files, § Government and censorship]

The Twitter Files article was nominated for deletion by a single editor (anyone can nominate); the nomination itself argued the material belonged under the Hunter Biden laptop article instead. The deletion debate was ‘snowball closed’ within hours as procedurally uncontroversial — not, in Wales’s account, a suppression of the underlying facts.
Wikipedia states it has never complied with a government demand to alter content, and frames its response to blocking threats as unconditional refusal (‘you’re going to lose Wikipedia’), while distinguishing this from the separate, harder case of protecting on-the-ground staff from physical risk in an authoritarian country.
Wales reports — after checking with Wikimedia counsel before the interview — that the Foundation received zero requests from the FBI or other US agencies to alter content, and states that even if such a request arrived, the Foundation structurally cannot edit Wikipedia’s content, which is written by the volunteer community.
Government engagement (e.g. discussions with Chinese officials, or public-health bodies during COVID) is framed as explaining how Wikipedia’s sourcing and community process works, not negotiating over content — Wales distinguishes ‘let’s have a dialogue’ from ‘let’s fix Wikipedia so the agency is happy.‘

Wikipedia funding / Future of Wikipedia [§ Wikipedia funding, § Future of Wikipedia]

Wikipedia is funded almost entirely by small individual donations (typically around $25), a choice Wales describes as originally aesthetic (‘I don’t want ads’) rather than strategic, later reinforced by an explicit theory that an advertising-based revenue model would eventually erode editorial independence regardless of the organisation’s stated values.
The most effective fundraising message tested over time was a fairness appeal (‘you use this constantly, you should probably chip in’) rather than a poverty or scarcity appeal.
The absence of an ad-revenue incentive is offered as the direct explanation for the absence of clickbait headlines or engagement-optimising internal links on Wikipedia — there is no advertiser whose interests would be served by steering a reader from Queen Victoria to a higher-margin topic.
Wales expects Wikipedia’s core form to persist roughly unchanged over ten years, with AI-assisted tools and a more conversational search interface layered on top; over 100 years, he expects the Wikimedia Foundation’s conservative reserve-building and a separately governed endowment (over $100 million at interview time) to have kept the project financially durable.
The least visible but, in Wales’s view, most significant growth trend is non-English Wikipedia expansion, which he expects LLM-assisted machine translation to accelerate meaningfully for lower-resource languages that legacy translation tools served poorly.