About WikiAsOfNow
WikiAsOfNow is a research assistant for Wikipedia editors. It finds claims whose “as of” reality may have expired and helps a human editor find and verify real sources. AI here is a grounded assistant to a human — never an author of article content, and never a source.
What the tool does
- Detect potentially stale claims with a deterministic, LLM-free detector.
- Use AI only to suggest neutral search queries and point at passages that may resolve the question.
- Confirm every supporting quote appears verbatim on the real, fetched source page.
- Show its work: the queries, the selected evidence, and the candidates it dropped.
- Build citations mechanically from the real source's metadata.
- Require the human to open and read each source before it can be cited.
- Generate a mechanical, human-editable disclosure naming the AI model and version.
What the tool will never do
- Generate or rewrite article prose for pasting.
- Produce or suggest a citation that the human has not verified against the real source.
- Assert what "happened" as fact from model knowledge.
- Combine multiple sources into a single claim or sentence.
- Author the disclosure text with a model (the disclosure is mechanical).
- Treat the content of a fetched web page as instructions to follow.
- Present a model-extracted snippet as text to copy into an article.
- Auto-submit edits to Wikipedia.
- Present its ranking as a decision the human can skip verifying.
The guardrails
These are the project's binding invariants, written to keep the tool inside Wikipedia's generative-AI guideline Wikipedia: Writing articles with large language models. The authoritative text lives in the compliance contract; the summaries below are orientation only.
- the no-machine-written-text guardrail (G1)
- The human writes every sentence that lands in Wikipedia.
- no machine-derived citations (G2)
- Citations are built mechanically from real source metadata.
- anchor every claim to a real URL (G3)
- Each surfaced claim points at one real, resolving source page.
- no cross-source synthesis by the machine (G4)
- One claim, one source; only the human combines facts.
- human verification is a gated act of opening the source (G5)
- Nothing is cited until the human opens and reads the source.
- the tool shows its work (G6)
- Selected and non-selected results are both shown for audit.
- prefer official sources and never hide the candidate set (G7)
- The full retrieved candidate set stays visible.
- support-check with a verbatim-quote check (G8)
- A deterministic check confirms the quote is really on the page.
- the LLM's role is boxed to three jobs (G9)
- Query, triage, point at a passage — nothing else.
- detection is deterministic (G10)
- Stale-claim detection uses no LLM at all.
- stay in the safe lane (G11)
- Living-persons articles are excluded from the easy-win queue by default.
- disclosure is mechanical (G12)
- A template names the AI model and version from the activity log.
- the audit log is foundational (G13)
- An append-only activity log makes the guarantees real, not asserted.
- responsible automated access (G14)
- A good API citizen to Wikimedia services.
- fetched content is untrusted data (G15)
- Web page content is data to the model, never instructions.
- no copying of source prose (G16)
- The human writes original text; snippets are pointers, not drafts.
Report a concern
Feedback & community
- General feedback and questions: GitHub Discussions
- Bug reports: GitHub Issues