Practical guide · Evidence first
Practical verification guide
Reliable verification is not a search for one “AI-looking” trait. It preserves the original material, checks provenance, compares independent evidence, and records uncertainty. This workflow puts detector output inside an evidence chain instead of treating a score as a verdict.
01 · Define the question
“Was AI used?”, “Was this edited?”, and “Is the claim true?” are different questions. Define the conclusion first so later evidence is not mixed together.
AI may have been used only for rewriting, frame interpolation, or noise removal, while entirely human-made content can still be false. State whether you are testing origin, editing history, or factual accuracy before collecting clues.
Obtain the complete text, original file, original URL, publication time, author statement, and version history when possible. Screenshots, copied passages, and platform transcodes lose context and technical information, so keep them without treating them as the best evidence.
Personal curiosity may justify a quick screen; journalism, academic discipline, hiring, and legal decisions require a fuller evidence chain. A single detector score can suggest what to review but cannot support a high-impact conclusion by itself.
02 · Check each medium
Check provenance and history before medium-specific technical clues. Every clue has counterexamples, and independent signals are more useful than repeatedly running the same kind of detector.
Compare known writing samples, drafts, and version history; verify that citations exist, timelines agree, and terminology fits the context. Fluent prose, consistent structure, or repeated phrasing may also result from human editing and cannot attribute authorship alone.
Open the AI text detectorRun reverse-image search on both the complete frame and meaningful crops to find earlier versions, original resolution, and a credible publisher. Then inspect content credentials, lighting relationships, and crop boundaries; one malformed detail may come from compression or retouching.
Open the AI image detectorExtract keyframes to trace earlier sources, check audiovisual continuity across speech and ambient sound, and compare the original file, upload time, and transcoded versions. Metadata anomalies are supporting clues and cannot independently prove generation.
Open the AI video detector03 · Weigh the evidence
Ranking evidence by verifiability and independence keeps a vivid visual artifact or a high score from outweighing stronger provenance records.
Content credentials signed by a credible party, original files retained by a camera or editing workflow, continuous drafts, and checkable production records usually address origin more directly than appearance. Confirm that the records are complete and the source is trustworthy.
Earlier published versions, reverse-search results, confirmation from the source, contextual contradictions, and reproducible technical anomalies can strengthen a conclusion. Evidence strength comes from independence, not from scoring the same file with many similar detectors.
Claims such as “too polished,” “the fingers look wrong,” or “the frame is too clean,” plus one detection percentage, are easily affected by templates, retouching, compression, and sample length. They can generate questions but should not generate accusations.
04 · Record the conclusion
A useful verification record states what was examined, what was unavailable, which explanation each item supports, and what new evidence would change the conclusion.
Keep filenames, source URLs, acquisition times, necessary file hashes, tools used, and key results. Do not preserve only a final screenshot; repeatable steps let another reviewer detect a changed file, a configuration difference, or an operational mistake.
Use graded language such as “the evidence supports,” “cannot yet exclude,” or “insufficient information,” and list alternative explanations such as compression, translation, or collaborative editing. A low score does not prove human authorship, and a high score does not prove AI generation.
When discipline, employment, reputation, or legal consequences are involved, a qualified person should inspect the original material and evidence chain. The person being assessed should be able to provide files, explain the production process, and appeal an automated result.