How DetectFake analyzes media—and where detection stops

1. The browser creates a JPEG analysis copy, with its longest edge limited to 1,024 pixels. The original file on your device is unchanged. This conversion can discard original EXIF fields and embedded provenance records.
2. For video, the browser samples evenly spaced, downscaled frames; the original video remains on the device.
3. Local image-classification models assess the prepared image and selected crops. Texture, edge, noise and compression measurements help combine the model evidence. Model files may need to download on first use; inference then runs on your device.
4. The combined evidence produces a verdict and confidence score. Video analysis compares sampled still frames; it does not examine every frame, authenticate speech, or perform a dedicated lip-sync test.
5. The interface presents the verdict, uncertainty, summary, and individual signals for human review.
6. If local models cannot load or run, the prepared image or video frames and filename are sent to DetectFake’s server for a fallback assessment. The default fallback checks file, compression and cross-frame statistics and does not use a paid AI gateway. Check the reported model and signals to distinguish this limited fallback from local model inference.

A visual classification is not an identity check, a verified editing history, or proof that an event happened. A result does not establish the date, location, creator, or truthfulness of a caption. An authentic photo can still be shared in a misleading context.

The analysis copy is not a reliable substitute for inspecting the original file’s metadata. DetectFake does not currently validate C2PA signatures in this upload flow. Inspect the original separately with a Content Credentials verification tool when provenance matters. Missing credentials do not by themselves establish that an image is synthetic.

Do not crop, annotate, screenshot, forward, or re-export the only copy you have. Those steps can change pixels, remove metadata, and make it harder for another reviewer to reproduce what you saw. Save the received file separately, then use a working copy for uploads, reverse-image searches, or visual annotation.

1. Record the claim being made, the source URL or account, and when you accessed it.
2. Where possible, request or download the highest-quality original rather than a social-media preview.
3. Keep surrounding context such as the post, caption, timestamp, and replies with the file.
4. Inspect original-file metadata and Content Credentials separately; missing metadata or credentials is not proof of manipulation.
5. Write down the tool, file version, settings, verdict, and limitations for every automated check.

This record makes disagreement easier to resolve. A detector can flag unusual pixels, while provenance or independent reporting may explain where the media came from and what it actually depicts. Use the media verification checklist to keep those observations together.

Confidence is the current scoring workflow’s strength of support for its selected verdict. It is not a measured accuracy rate or a calibrated probability. For example, an 85% score does not establish that 85 out of 100 similar decisions would be correct. Similar visual artifacts may have several causes.

Read the verdict, model and signal list together. A result that leans toward AI or non-AI is weaker than a firm classification. A processed-photo or processed-scan label describes possible file handling, not verified origin. Even a high-confidence result can be wrong and should be checked against provenance, context, and independent evidence.

This page does not report a measured accuracy benchmark. A useful evaluation needs documented camera originals and AI-generated exports, varied subjects and sources, and a held-out set that was not used to adjust detection thresholds.

Test originals alongside compressed, resized and scanned copies. Copies of the same source are not independent examples. Record the model version and whether a server fallback ran, and report those groups separately.

• False positives: known camera images incorrectly labelled AI-generated.
• False negatives: known AI images incorrectly labelled likely authentic.
• Decision coverage: the share of files receiving a firm origin verdict.
• Undecided cases: weak leanings, processing labels and limited-analysis results, reported separately rather than counted as correct.

Report sample counts and coverage alongside error rates. High accuracy on a small subset of decided files can hide weak performance on the rest. Document edits, face swaps and video require separate evaluations; a binary image-origin test does not validate those capabilities.

• Social-media compression, screenshots, and repeated re-encoding can erase forensic traces.
• Filters, denoising, sharpening, unusual lighting, and computational photography can resemble synthetic artifacts.
• Small images and short videos provide less evidence to examine.
• New generation methods may produce patterns that differ from previously observed media.
• A composite may contain both authentic and generated regions, making a single label incomplete.
• A change between sampled video frames may reflect an ordinary scene cut, not manipulation. A short edit between sampled frames may be missed entirely.

When local inference succeeds, the selected media is analyzed on your device. Model and runtime downloads still require network connections. If inference fails, the server fallback receives the prepared image or sampled frames and filename; the original video file is not sent by this flow. Scan history, including a preview, is stored in the browser. See the privacy notice for service data handling.

Start with a file whose origin you can document, save the verdict and model name, and compare an original with a compressed copy. A few examples can reveal a failure case but cannot establish overall accuracy. Our accuracy guide explains how to interpret false positives, false negatives and confidence scores.

Use the media verification checklist to record your evidence. For a correction or reproducible problem, contact support@wiki-services.com. Describe the browser, model name and steps; do not send sensitive media without checking that you can share it.

DetectFake is suitable for preliminary review, moderation triage, scam checks, and deciding whether deeper investigation is warranted. It is not a substitute for a qualified forensic examination, chain-of-custody procedures, or independent verification in legal, employment, financial, medical, or safety-critical settings.

For a repeatable verification process, read our guide on how to tell if an image is AI-generated, and for realistic expectations see are AI image detectors accurate.