How GPTAnon Checks News Bias: Our Sources, Evidence and Methodology

A bias label is only useful if you can understand how it was reached. An article can contain supported facts while emphasizing one perspective. Another may sound neutral but offer little evidence for its central claim.

GPTAnon Check separates those questions. We examine the article, seek relevant public evidence, and explain what supports the findings. Publisher background provides context; it does not decide the verdict.

This article describes the behavior currently implemented in GPTAnon Check. It does not claim human review, scientific validation, proprietary training, or continuous monitoring.

The short version

GPTAnon Check is a temporary, evidence-bounded article check. It accepts a public article URL or pasted article text, extracts readable content, searches for relevant public sources, and asks the selected active Check model to return structured findings: what holds up, what needs context, and what remains unverified. The server validates that every cited passage and source came from the supplied inventory before showing it.

The result is not a truth score, a political label, an AI-authorship detector, or a verdict about every sentence. It is a compact explanation of the claims and passages that were actually checked.

Check an article with GPTAnon Check.

What we assess

Evidence support asks whether the key factual claims checked are supported, contradicted, or still unresolved by the available evidence. Framing examines wording, emphasis, selection, and context in the readable article. Source context describes available publisher background separately from article-level findings.

Political leaning does not establish accuracy. An unfamiliar publisher is not automatically unreliable. We do not compress these distinctions into a universal truth percentage, outlet score, or consensus number.

The checking process

This process is bounded. It is not a claim that every relevant source was found or every article statement was verified.

Establish what we can read

A public URL is not the same as readable evidence. GPTAnon first fetches the public page, follows a bounded number of redirects, and extracts article content from structured article data or readable page text. If the page is inaccessible, excerpt-only, malformed, or too thin to establish the article's scope, the UI asks for pasted text rather than pretending the missing material was checked.

Headline-only material cannot support a confident claim about everything the full article includes or omits. For incomplete text, the result keeps evidence support and framing at Not assessed. This is a scope limit, not a judgment about the publisher.

Check important claims with specific evidence

The checker searches for original documents, public data, studies, direct statements, and independently reported accounts when they are available. A source must support the specific finding linked to it, not merely discuss the same subject.

For example, a headline might say that a policy caused a change when the available evidence only shows that the two events occurred together. A useful check explains that distinction with the relevant passage instead of stamping the entire story true or false.

The model is constrained to cite only exact article passage IDs and source IDs supplied by the server. It cannot invent a citation, URL, publisher, date, or primary-source designation. If search fails or no independent reporting is available, the result says that evidence is insufficient instead of treating the failure as corroboration.

Count reporting origins, not links

Five websites repeating one report are not five independent confirmations. GPTAnon normalizes source families by hostname and, when fetched source-page metadata identifies a known reporting origin, groups syndicated coverage under that origin. Republishing is labeled as other reporting rather than independent reporting.

This is provenance handling, not a vote. The implementation does not calculate an arithmetic average, consensus score, outlet score, or comprehensive verification percentage. A repeated link cannot create artificial corroboration, and a single source cannot become strong evidence merely because it is displayed prominently.

Inspect framing in the article

Loaded wording, selective quotations, unsupported certainty, and headlines that extend beyond the reporting can shape interpretation. Missing context is harder to judge: enough article text and relevant external evidence must be available before declaring an omission.

A framing finding must point back to article passages. In the qualitative summary, Strong framing requires full article text, at least two concrete framing signals in the model's explanation, at least two article passages, and no explicit uncertainty language. Some framing requires full text, at least one recognized signal, at least one passage, and no explicit uncertainty language. Otherwise the summary is Little detected. Excerpt-only or inaccessible content is Not assessed.

The recognized signals include language about loaded or sensational wording, emotional or selective presentation, omission, foregrounding, emphasis, downplaying, headlines, wording, absence, and selection. These are bounded implementation signals, not a scientific bias scale. Disagreeing with a conclusion does not itself establish bias.

Understand the visual summary

Evidence support and Framing are qualitative summaries of the claims and passages checked, not probabilities and not assessments of every sentence.

| Summary | When the current implementation uses it |

| --- | --- |

| Supported | Full article text, usable search, at least one primary or independent-reporting source attached to the central “holds up” finding, and that finding is marked supported. This does not verify every assertion. |

| Contradicted | Full article text and usable search where a credible primary or independent source contradicts the central checked claim, with no credible supporting source attached to that finding. |

| Mixed | Credible supporting and contradicting sources are both attached to the central checked finding. |

| Insufficient evidence | Search is not usable, no primary or independent source is available, or the checked findings do not justify a broader category. |

| Not assessed | The available article text is not full text, such as an inaccessible or excerpt-only page. |

Mixed means genuinely conflicting checked material, not merely missing information. Strong framing does not mean false reporting. The underlying findings remain visible so a reader can inspect why the summary was produced.

How publisher information helps

Reusable public publisher information is kept separate from article-level findings. It can provide available identity, country, language, ownership/background, freshness, provenance, and licensing context. It does not decide evidence support, framing, reliability, or political leaning. Missing data remains unknown; it does not silently become “safe.”

Enabled source table

| Source | Fields used | Role | Provenance and licence | Limitations |

| --- | --- | --- | --- | --- |

| Wikidata structured data | Publisher name, available countries, languages, ownership/background, source Q-IDs, refresh and expiry times | Shared publisher-directory context attached to evidence sources when an exact public domain has a cached record | Wikidata structured data; CC0 attribution is retained in the product metadata | It is not a fact-check, political-leaning scale, or reliability score. Records can be missing, disputed, stale, or incomplete. Refresh is bounded and explicit rather than performed during every article check. |

Not enabled: CRED-1 is being evaluated only as possible domain-risk context under its stated CC BY 4.0 terms and upstream-attribution review; no CRED-1 fields or labels are used in GPTAnon Check. AllSides and MBFC data are not imported or used. None of these sources produces an outlet score.

Choose a useful alternative

A next read should add original evidence, independent reporting, or a relevant overlooked perspective on the same story. The result explains what the selected source contributes. Weak evidence does not become valuable simply because an outlet is perceived to lean another way. Often another substantive angle is more useful than another political label.

Source ordering can use available metadata to look for a contrasting perspective, but that ordering does not change relevance, source-family independence, evidence status, factual grounding, or publisher context. A second model's opinion is also not independent factual corroboration.

Know which AI checked it

News models come from the active, validated News catalog. The server determines the selected model, quote, any configured fallback, actual model attribution, and settlement. The result shows the model that actually produced it, including fallback attribution when applicable. Another model may interpret the same evidence differently; changing models is a new explicit analysis, not a second factual source.

The live catalog can change, so the result's own model attribution and quoted cost are authoritative for that check. GPTAnon does not claim human review or scientific validation of a model's judgment.

The model request is a separate privacy boundary. GPTAnon routes the article text and required context through OpenRouter to the selected provider; those services apply their own retention and training terms. GPTAnon's current privacy and provider disclosures explain what is sent, what identity is not included, and why the selected provider's current terms still matter. Do not submit names or other identifying details that are not needed for the check.

What stays and what disappears

Shared publisher background can be reused without retaining article checks. This avoids repeated public publisher lookups while keeping reading history out of the shared directory.

Submitted article text, URLs, findings, source previews, search queries, thumbnails, recommendations, and session evidence stay separate from the shared publisher cache. The current News session is browser-owned and temporary:

• Clear removes the visible result and asks the server to remove the session cache.
• Replace starts a new check and removes the previous result from the page; the server only reuses research when the session and article fingerprint both match.
• Close aborts an active request, clears the session, and returns to Chat.
• Expiry clears the session after 15 minutes and shows an expired-session state.
• Tab closure fallback aborts the active request through the page-hide handler; a server-side retry copy expires after the same 15-minute window if the tab disappears unexpectedly.

These controls clear GPTAnon's browser and session copy. They do not promise erasure from OpenRouter, the selected model provider, browser caches, network logs, or every intermediate system. See the privacy page for provider-processing details. Caching publisher facts is not building a history of which articles people read, and News has no history endpoint.

Limits and corrections

AI-assisted checks can miss context, misread a source, or struggle with developing events. Publisher information can be incomplete or outdated. Search results can be unavailable or can fail to establish a useful independent source. Showing passages, source links, provenance, and uncertainty lets you inspect the reasoning and make your own judgment.

Use the existing support form for corrections. GPTAnon does not automatically attach private article or session content to a support request; choose what to share. This methodology does not claim a named human reviewer, scientific validation, proprietary training, or continuous monitoring.

Try the check

Check an article with GPTAnon Check.

If you want the shorter, practical walkthrough, read How to Check News for Bias and Missing Context. It is an introductory guide, not a duplicate of this implementation-focused methodology.