How DeepfakePolicy reaches a result

Reported image detection accuracy

DeepfakePolicy reports 98.8% image detection accuracy, with its evaluation based on more than one million checks. This is our product-reported figure, rather than an independent certification or the probability for one file. Results vary by dataset. Test representative, labelled files to measure false positives and missed detections for your use.

How checks become a report

Available checks depend on the content type. Neural signals, file metadata and visual findings are organised in one report, with public-source research when requested. C2PA Content Credentials can be inspected separately with our local checker.

Image and audio

• Neural analysis: checks supported image and audio files for patterns associated with AI generation.
• Visual review for images: OpenAI examines anatomy, contact between objects, geometry, labels, lighting and reflections.
• Ordinary edits: changes such as meme captions are assessed separately from AI generation.
• Conflicting findings: if the checks disagree, the report shows the conflict and keeps concerning findings visible. A model’s confidence in a “real” label is not treated as a guarantee of 0% AI content.

Video

• Separate checks: supported clips are checked for generated visuals, face manipulation and synthetic speech.
• Report headline: the strongest relevant finding contributes to the headline, while the separate results remain visible.
• Missing signals: a missing face or voice does not make a clip genuine.

Files, ZIP archives, links and documents

Photo and video analysis stays the main workflow. A public link or a document is another way to bring media into the same set of checks. Find media from a link or choose a document.

1. Prepare a preview: extract supported images from PDF, DOCX or PPTX, or retrieve accessible photos and videos referenced by a public page. Direct photo, video and PDF links are supported within the import limits. No detector runs at this step.
2. Select the content: review the actual images or play available clips, choose the files and see their individual and total credit cost. Identical copies are shown once. Obvious website logos and icons are not preselected.
3. Run separate checks: each selected image or video follows its regular analysis route. Optional extracted text is sent separately to Pangram for an AI-writing analysis. Text starts unchecked and has its own price and report.
4. Review and export: the set includes a report per checked file, individual scores and a summary of completion and review status. CSV and a ZIP of PDF reports are available. We do not average image, video and text scores into a document-authenticity percentage.

PDF images are extracted as PNG pixel copies. This does not recover the original camera file, its compression history or all metadata. Vector artwork, masks and some embedded image types may not be extracted. Page numbers are retained for PDF images and slide numbers for presentations. Word images refer to the document body, header or footer; fixed pagination cannot be inferred reliably.

Imports accept sources up to 20 MB and PDFs up to 50 pages and presentations up to 50 slides, with a 24 MB extraction budget, an 8 MB limit per selected Office part and a 4-megapixel limit per decoded PDF image. A page preview includes up to 12 media references. Imported videos must be playable MP4, MOV or WebM clips up to 60 seconds and 20 MB. Password-protected documents, login-only pages, embedded platform players and some streams need a direct file upload instead. Scanned text is not automatically transcribed with OCR. The preview identifies omissions and limits.

For long documents, the displayed extracted text is a sample of up to 20,000 characters; only that sample is checked and billed. Original documents and extracted text bodies are not stored with batch metadata. Personal batches need the tab to stay open while checks run. Company photo and video batches continue on the server after upload. Files not yet uploaded must be reselected if you close the tab.

Optional text analysis through Pangram

We check writing patterns in text from 255 to 20,000 characters. Labels such as AI-written, AI-assisted or human-written are model estimates. We have not independently validated them on a representative benchmark. They do not prove whether a particular person used AI, establish factual accuracy or determine whether the document’s images are synthetic.

Detector scores, visual findings and source research

The detector result, visual observations and source research answer different questions. The report conclusion follows the structured detector result. A realistic appearance or a matching public source cannot cancel a detector signal. A genuine disagreement is marked for review; the original findings remain available.

The colour scale shows the strength of AI-related signals: green from 0 to 20, amber above 20 to 65, and red above 65 to 100. These display bands do not change a detector’s own classification thresholds. A missing score stays grey. Confidence in a category such as authentic is labelled separately and never placed on the AI-generation colour scale.

How to interpret a result

• Detected or high probability

  The model found patterns associated with its synthetic examples.

  Action: Seek corroborating evidence.

• No strong signs

  The tested model did not find a strong signal in this submitted copy.

  Note: Compare the result with the original file and its source.

• Inconclusive or not applicable

  The evidence was insufficient or unsuitable.

  Action: Try a clearer original or another source.

Requesting a report review

Every report includes an ID, creation time and deepfakepolicy.report.v1 format version. These identify the export, but do not prove that a screenshot is complete or unchanged. If something looks wrong, send the original export and report ID to deepfakepolicy@proton.me.

What affects a result

• Compression, resizing and transcoding can remove useful signals.
• Background noise and short recordings can limit audio and video analysis.
• New generation methods and content unlike the model’s training data can affect results.
• Screenshots, illustrations and animation can resemble synthetic patterns.
• Translated text and ordinary editing can also resemble synthetic patterns.

A practical verification workflow

1. Preserve the original, highest-quality file.
2. Find the earliest source and compare surrounding context.
3. Check metadata and available Content Credentials, understanding that missing metadata proves little.
4. Reverse-search key images or video frames.
5. Verify identity through a known, independent communication channel.
6. Escalate high-stakes cases to a qualified forensic specialist.

Content Credentials inspection

The local Content Credentials checker uses the official C2PA web SDK to read an embedded manifest, signer details, recorded actions and validation status without uploading the file. A valid claim is provenance evidence, not proof that every depicted event is true; an absent manifest does not prove manipulation.

Testing the service

Small synthetic files may be used to test uploading, analysis and report delivery. A completed result confirms that the tested path worked; a sign-in prompt, credit error or rate limit does not. These tests do not measure detection accuracy, which needs a separate representative benchmark.

Comparing detectors

Our benchmark plan sets out the data, image-quality conditions and measurements needed for a fair comparison. Results are published only after the labelled dataset is fixed and the required checks pass.

How we evaluate model updates

Keeping content for research is your choice.We do not intentionally keep your content for our research unless you separately opt in.

• Background evaluation: We test model updates separately from live analysis. These evaluation runs do not decide your result.
• What these tests record: Evaluation records may contain scores, model version, agreement between models and processing time, but not submitted content or a direct account identifier.
• Before public use: Each model version must pass a separate evaluation on labelled data before it can influence a customer report.

Questions

For technical or methodology questions, contact deepfakepolicy@proton.me. Use is also governed by the acceptable-use policy.