Credibility Signals
Know which Wikipedia edits to trust
Wikipedia changes constantly, and most of those changes make it better. A few do not. If your product ingests Wikimedia content at machine speed, you need a way to tell the difference before your users see it.
Credibility signals are data points delivered in every Wikimedia Enterprise API response: revert-risk scores, visibility and protection status, editor context, and more. They let you judge two things about any revision - is it safe to use, and is it significant enough to act on - using your own thresholds, not ours.
Why credibility signals matter
Search results, knowledge graphs, and AI assistants are only as reliable as the data they ingest. Wikimedia content is the backbone of many of those systems, and it is living data: written, revised, and repaired continuously by a global community of volunteers.
That repair process works remarkably well, but it happens on human and bot timescales. If you consume changes in real time, the gap between an edit landing and the community's verdict on it is exactly where your risk lives. Ingest everything blindly and vandalism can reach your users; wait too long and you serve stale facts. Credibility signals close that gap: they tell you, revision by revision, what the community and Wikimedia's machine-learning models already know about an edit, so you can decide what to publish, hold, or skip.
How Wikipedia fights vandalism and misinformation
Wikimedia projects are written and maintained by a dedicated community of volunteers who ensure the free knowledge ecosystem is updated to reflect the changing world around us. For every Wikimedia project, in each language edition, a set of guidelines and rules governs how knowledge should be described, and a rigorous system of citing sources and finding consensus ensures that this knowledge is neutral and reflects reality.
At the heart of this process is Wikipedia's core policy of Verifiability: a reader can check which third-party source any piece of information comes from. Supporting policies require sources to be reliable and forbid original research. This standard creates a paper trail for facts, allowing the community to constantly cross-reference content against peer-reviewed journals, mainstream news, and academic texts. For API users, it means Wikipedia isn't an arbitrary collection of claims, but a structured web of external evidence.
Even with these processes in place, content can be wrong, uncited, or contradictory. Human knowledge is messy: consensus changes, and multiple points of view can conflict yet coexist. Sometimes the intent is malicious: people try to use Wikimedia projects to spread misinformation, further political agendas, troll, or prank. Generally, over time, the community fixes 'bad' edits, removes vandalism, and updates articles to reflect new consensus. Historically, on larger language Wikipedias, bots and filters catch over 80% of vandalism, while human patrolling catches the next 10-15%.
WATCH: If volunteers edit Wikipedia, how can you trust it? A Wiki Minute
These mechanisms are difficult to follow unless you are a deeply embedded Wikimedia volunteer. Credibility Signals are the product of Wikimedia Enterprise's content integrity work: we translate the community's methods and processes into simple, actionable data points, so the trust the community builds becomes trust your systems can compute with.
Signals, not verdicts
Ask ten reusers what "trustworthy" means and you will get ten answers. A kids' encyclopedia, a news aggregator, and an LLM training pipeline draw the line in very different places. That is why we do not reduce a revision to a single score: one number without context gets misinterpreted.
Instead, every API response carries a set of signals you can read the way you read a nutrition label - each one a specific, verifiable fact about the revision, the editor, or the article, ready to be weighed against your own standards:
| Signal | Example | What it tells you |
|---|---|---|
version.scores.revertrisk | probability: 0.96 | How likely this edit is to be reverted by the community |
visibility | editor: false | The community hid part of this revision, such as an abusive username or personal information in a comment |
protection | edit: autoconfirmed | How locked down editing currently is for this article |
version.editor | temporary account, 12 edits | Who made the change and how much history they have |
version.maintenance_tags | citation needed added | The article's verifiability just changed |
watchers_count | 1,204 | How many editors are watching this article for changes |
You set the thresholds, ranges, and criteria that match your product's risk tolerance. The signals are the knobs; the mix is yours.
Safety and significance: the two questions
Every signal in the payload helps answer one of two questions:
Is this edit safe? Safety describes how likely an edit is to be accepted and kept as part of the article, rather than reverted, changed, or hidden. It is measured by signals like the article's visibility status, its protection level, the revert-risk score, and the presence and quality of references and citations.
Is this edit significant? Significance describes whether a change matters enough to act on: the relevance and importance of an edit. It is measured by signals like the size of the change, maintenance tags, the editor's experience, and the attention an article receives.
To answer both, credibility signals draw on three areas of data in each revision: the edit itself (changes in the article's content), the editor (was it a bot, a temporary account, or an experienced editor?), and the citations (the sources backing the change).
How reusers put credibility signals to work
Threshold pipelines. Most reusers define acceptance rules and let the pipeline run: publish revisions below a chosen revert-risk score immediately, hold riskier ones for review, and skip revisions the community has hidden. Signals become simple gates in an ingestion flow.
Custom scoring. Teams with their own ML infrastructure consume the raw signals and blend them into internal trust scores, weighting what matters for their product. The same fields that gate a search snippet can be features in a ranking model.
Context for users. Some platforms answer trust questions with transparency: they link entities and claims back to the Wikipedia articles they come from, letting users see the living source and its history. Credibility signals help decide when that context is stable enough to surface.
Building your pipeline now? Follow the credibility signals tuning guide: recommended thresholds, an evaluation strategy, and field-by-field reference tables.
How to Access Credibility Signals
Credibility signal fields ship in all of our API collections - the same JSON structure everywhere:
- On-demand API: request individual articles from any project, signals included.
- Snapshot API: bulk files of entire projects, every article carrying its signal fields.
- Realtime API: a stream of every change as it happens, with signals attached to each revision - including edits the community has since hidden or fixed.
A free account covers all three APIs; talk to Sales for more requests, faster data, or help designing your acceptance rules.
Keep exploring
Related articles and deep-dives on Credibility Signals:
