Who decides what is true in a fact-check?

When you verify a claim online, the outcome depends on the sources used. See how measuring primary evidence replaces editorial bias with verifiable data.

Fact-checking · Published 2026-09-04 · 7 min read · 1,547 words

AI-generated illustration. A split computer screen showing a social media post on one side and a list of linked reference documents on the other.

When you read a disputed claim online, the most important question is often not about the claim itself but about the verification process: who decides what is true? For years, social media users relied on institutional organizations to verify viral posts. Today, the landscape has fractured. The answer to how a claim is verified has become a measurement of evidence rather than a statement of editorial posture.

The traditional fact-checking industry has faced increasing scrutiny over its selection mechanisms. As The Alamo Post reported, major newsrooms and platforms turned fact-checking from a narrow correction service into a political labeling operation that protects favored sources. This shift has eroded public trust. A detailed analysis by PolitiFact Bias: 2026 highlighted a sharp drop in harsh ratings for certain political figures, suggesting that organizations lose interest in issuing critical ratings evenly. Academic audits and industry reporting, summarized by Factually.co, reveal selection effects that complicate the narrative of neutral gatekeeping. Conservative outlets have long argued that these organizations act as partisan filters, an observation echoed by PJ Media. These disputes eventually led platforms like Meta to abandon their third-party fact-checking programs entirely, as documented by ONE.WORKS.

Replacing human editors with artificial intelligence does not automatically solve the problem of source selection. Recent studies show that automated retrieval systems develop their own preferences. According to research published on alphaXiv, neural retrievers are biased toward content generated by large language models over human-written texts. This bias stems from the supervision in retrieval datasets rather than the models themselves, as explained in a 2026 arXiv study. Researchers have traced this training-induced bias across different stages of model development, a phenomenon detailed in another alphaXiv paper. To build a fair system, developers must actively optimize source selection for balanced knowledge retrieval, an approach outlined by researchers at Salesforce.

The wyper approach removes the editorial shortcut entirely. Sources are never filtered by origin. There is no mainstream bonus and no preference for official sources. The tools return a truth score of 1-10 alongside an evidence chain of real, linked sources. We never call the truth score proof of trustworthiness, and we never say the tool judges intent. Both judgments stay strictly with the reader. You can test this verification process in the free wyper web app to see the evidence chain for yourself.

The standard check uses a Dual-AI cross-check featuring Gemini and Grok. Grok searches for itself in this configuration, ensuring the two models do not share a single evidence set. However, our own published measurements show that two models trained on overlapping text are correlated, not independent. When we tested 30 claims with frozen sources, Gemini and Grok agreed in 29 of 29 cases, even on deliberately contested claims. Because of this correlation, saying a claim is confirmed by two AIs carries little information.

For deeper analysis, the Full Spectrum feature introduces a Triple-AI reading. This uses three models and one evidence set. The third voice does not search. It receives exactly the sources the main model received and reads them independently. A disagreement here is a different reading of the same material, not a different search result. When we applied a third model of different origin to 26 claims, it disagreed in 3 cases, every time with a machine-verified quote. This independent reading is the strongest verification method we have.

Our source reach measurements confirm the value of avoiding mainstream filters. In a test of 24 claims where the root was named in advance, the system reached 63 percent primary sources, 14 percent legacy media, and only 1 percent fact-checkers. All six non-English roots were successfully reached. In our quote checking test of 633 real citations re-fetched, 73 percent were found verbatim on the page. Another 19 percent of pages were unreachable for automated readers due to paywalls, PDFs, or 403 errors, while 6.4 percent were not findable at all.

If a reader disagrees with a verification outcome, they can log it in our public dispute register. This register is append-only and hash-chained. Entries cannot be silently edited or removed. This ensures that every challenge to a reading remains visible and permanently recorded as a public measurement.

Video verification requires a different technical approach. The wyper extension uses subtitles when reachable. When subtitles are unavailable, an AI model actually watches the video within a specific time window. This covers up to 10 minutes on the free tier and 30 minutes on Pro for YouTube. For X videos, the limit is up to 2 minutes free and 7 minutes Pro. Longer videos on X are refused outright, not partially checked. The web app cannot see your logged-in YouTube session, so subtitles are often unavailable there. For X, both tools take the same route.

We also offer the AI Spotter in the web app for images and video. It provides a provenance report, never a detector. It never outputs a percentage of how artificial something looks, because that number is the harm we exist to prevent. For video, it takes still frames and searches each one separately. A hit on any frame is a hit on the video. The report checks content credentials against 55 official trust anchors. A content credential sits in its own box beside the picture data. Any repackaging drops it, including a lossless repackaging that leaves every pixel untouched. Platforms repackage on upload. On a video taken off a feed, that layer will almost always find nothing. That is the normal case, not a suspicion. We advise checking the original file instead. What carries the weight for video is the earliest appearance plus the event-versus-post date gap. This catches old footage relabelled as new.

There is an exception to the free tier structure for this specific feature. The AI Spotter is not in the free tier for images or video. It costs 1 credit per image and 4 per video, available on paid plans only. A provenance report calls three paid services on every single run, so there is no cheap first look to give away. The text fact-check remains completely free with five checks a day and no account required.

Who needs which: The manual route is best for users who want zero data leaving their browser and have the time to trace primary documents themselves. The wyper web app is built for users who want to check a specific article or claim without installing any software. The wyper extension is designed for social media users who want to verify claims, read evidence chains, and check videos directly in their timeline without switching tabs.

The short version

The person reading the evidence is the only one who decides what is true. By providing an evidence chain of primary sources and a public dispute register, verifiable tools remove the editorial filter. You get the raw data and the original context to make the final judgment yourself.

Keep reading