Fact Checker is an AI-assisted tool that helps you verify images and text claims. It is not a replacement for professional fact-checkers, journalists, or primary source verification. Our goal is to give you a fast, transparent first assessment so you can decide whether to trust, share, or dig deeper.
Every verdict includes the sources we found, a confidence score, and any red flags we identified. You can always click through to the original sources and judge for yourself.
2. The Fact-Checking Pipeline
When you right-click an image or text and select "Fact Check," we run it through a multi-step pipeline:
Input analysis — For images, we use Google Vision API to extract labels, detect faces, and identify text (OCR). For text, we extract the claim directly.
Reverse image search — TinEye finds where the image has appeared online before, helping detect manipulation or out-of-context reuse.
Fact-check database lookup — Google Fact Check API searches verdicts from professional fact-checkers (Snopes, PolitiFact, AP, Reuters, and others).
News event search — GDELT and Brave Search find news articles and reports related to the claim or image.
AI synthesis — Gemini Flash (or GPT-4o-mini) analyzes all collected evidence and produces a verdict, confidence score, summary, and source list.
3. Verdict Types
TRUE
Multiple credible sources confirm the claim. No significant contradicting evidence found.
FALSE
Credible sources directly contradict the claim, or the claim is a known debunked hoax.
MISLEADING
The claim contains a kernel of truth but omits critical context, uses cherry-picked data, or distorts the original meaning.
INSUFFICIENT EVIDENCE
We could not find enough reliable sources to verify or refute the claim. This does not mean the claim is false — it means we lack evidence.
4. Source Selection and Credibility
We use a curated seed list of trusted sources, categorized by credibility:
High credibility — IFCN signatory fact-checkers (Snopes, PolitiFact, AP, Reuters, AFP), peer-reviewed journals, government statistics agencies.
Medium credibility — Established news outlets with editorial standards (BBC, NYT, Guardian, WSJ).
Low credibility — User-generated content, social media posts, blogs without editorial oversight. Used only as leads, never as sole evidence.
Each source in the verdict card displays a credibility badge (High, Medium, or Low) so you can judge the evidence quality at a glance.
5. AI's Role and Limitations
The AI model (Gemini Flash or GPT-4o-mini) synthesizes evidence from the adapters above. It does not have access to the internet in real-time — it only sees the evidence we collected. This means:
If our adapters miss a source, the verdict may be incomplete.
The AI can be wrong, especially on nuanced or politically sensitive claims.
Confidence scores reflect the AI's assessment of evidence quality, not absolute truth.
Verdicts on breaking news (less than 24 hours old) are less reliable.
We use a low temperature (0.3) for consistency and require the model to cite sources for every claim in its summary.
6. Confidence Score
The confidence score (0–100%) reflects how strongly the evidence supports the verdict. It is not a probability of correctness. A score of 90% means the AI found strong, consistent evidence — not that there is a 90% chance the verdict is correct.
High (75–100%) — Multiple credible sources agree, evidence is direct and unambiguous.
Medium (50–74%) — Some evidence found, but sources may be less authoritative or partially relevant.
Low (below 50%) — Weak or conflicting evidence. Treat the verdict as a starting point, not a conclusion.
7. Caching and Fresh Checks
To keep the service fast and affordable, we cache verdicts. If you check the same image or claim that someone else checked recently, you will receive the cached result at half the credit cost. Cached verdicts may be outdated — if the claim is about a developing story, consider running a fresh check.
8. Corrections Policy
If you believe a verdict is wrong, use the "Report" button in the extension or email alex@bukhalov.com. We review all reports. If a verdict is found to be incorrect, we:
Update the cached verdict so future checks return the corrected result.
Add a correction note visible on the permalink page.
Investigate the root cause (missing source, AI error, adapter failure) to prevent recurrence.
9. What Fact Checker Is Not
Not a definitive source of truth. Verdicts are AI-assisted assessments, not authoritative rulings.
Not a replacement for professional fact-checkers. We aggregate and synthesize their work — we do not replace it.
Not a legal, medical, or financial advisor. Do not make decisions based solely on our verdicts.
Not infallible. AI can be wrong. Always verify critical claims through primary sources.