Upload a file or paste a URL to run multi-modal forensic analysis with confidence scoring.
Open AFIP VerifyAFIP provides forensic analysis across four media types. Each modality uses specialized detection methods tuned to the specific artifacts and patterns that AI generation produces in that medium.
Video analysis examines frame-level consistency, temporal coherence, face boundary artifacts, biological signals (blinking, pulse estimation), and audio-visual synchronization. The analysis detects face-swap deepfakes, lip-sync manipulation, full-face generation, and body-puppet techniques. Results include per-frame confidence scores that show exactly where in the video the strongest forensic signals appear. Learn more about video forensics.
Audio analysis evaluates spectral characteristics, prosody patterns, breathing and pause naturalness, formant consistency, and phoneme transitions. The system detects text-to-speech synthesis, voice cloning, voice conversion, and audio splicing. It works on standalone audio files and can extract and analyze the audio track from video content. Learn more about audio forensics.
Image analysis applies frequency-domain fingerprint detection, noise pattern analysis, compression forensics, CFA pattern checking, and multi-model classifiers trained to detect output from major generation platforms including Stable Diffusion, Midjourney, DALL-E, and Firefly. The analysis also checks for traditional manipulations like copy-move forgery and splicing. Learn more about image forensics.
Text analysis examines token distribution patterns, perplexity characteristics, stylistic consistency, and vocabulary usage to determine whether text was written by a human or generated by an AI language model. The system provides paragraph-level scoring, identifying specific sections that show the strongest AI generation signals.
AFIP does not rely on a single detection method. Each piece of content is processed through multiple independent forensic techniques, and the results are compared. This multi-layer approach reduces the risk of false positives (declaring authentic content as fake) and false negatives (missing actual deepfakes) because different methods catch different types of generation artifacts.
The AFIP forensic report provides a confidence score that reflects the strength and consistency of the forensic evidence. The score is not a simple AI probability output. It is a weighted assessment that considers the number of independent forensic signals detected, the strength of each signal, and whether the signals agree or conflict.
Every analysis produces a detailed report that goes beyond a simple score. The report includes the specific forensic methods applied, the findings from each method, the evidence supporting the confidence score, and any notable anomalies or limitations. This transparency allows users to understand why the system reached its conclusion and to weigh the evidence themselves.
The forensic report is organized into sections. The summary provides the overall confidence score and a plain-language interpretation. The method-by-method breakdown shows results from each forensic technique. The evidence section highlights the specific signals that drove the assessment. The limitations section notes any factors that may affect reliability, such as heavy compression or very short content samples.
A score in the 0-20 range means that multiple independent forensic methods found evidence consistent with authentic, non-AI content. Camera sensor noise patterns, natural compression artifacts, and human-consistent signals all support authenticity. A score in the 81-100 range means that multiple methods detected strong AI generation fingerprints, such as GAN spectral artifacts, diffusion model noise patterns, or text token distribution anomalies.
Scores in the 41-60 range (inconclusive) are common for content that has been heavily compressed, post-processed, or contains a mix of authentic and AI-generated elements. In these cases, the report provides detailed findings to help the user interpret the mixed signals.
For high-confidence results (0-20 or 81-100), the forensic evidence is strong and the assessment is reliable. For mid-range results, treat the report as a starting point for further investigation. Consider the context: who shared the content, what claims are being made, and whether additional evidence is available. The AFIP report provides the forensic layer; combining it with source verification and contextual analysis produces the most reliable overall assessment.
AFIP detection methods are tested against standard academic benchmarks to ensure reliability. Performance varies by media type and generation method.
| Media type | Benchmark | Detection accuracy | False positive rate |
|---|---|---|---|
| Video (face swap) | FaceForensics++ | 96.2% | 2.1% |
| Video (lip sync) | DFDC | 91.8% | 3.4% |
| Audio (voice clone) | ASVspoof 2024 | 94.5% | 2.8% |
| Image (GAN) | GenImage | 97.1% | 1.5% |
| Image (diffusion) | DiffusionDB | 93.4% | 3.2% |
| Text (LLM) | Multi-model set | 89.7% | 4.6% |
A critical test for any detection system is whether it can identify AI content from models it was not specifically trained on. AFIP's ensemble approach, using multiple independent detection methods rather than a single classifier, provides strong cross-model generalization. Detection accuracy on unseen model architectures typically drops 3-8 percentage points compared to known models, which is within acceptable ranges for practical use.
No detection system is perfect. AFIP accuracy decreases with heavily compressed content (below JPEG quality 20 for images, or very low bitrate audio). Short text samples (under 200 words) produce lower confidence text analysis results. Content that mixes authentic and AI elements (e.g., a real photograph with an AI-generated background) may produce inconclusive scores. The forensic report always notes these limitations when they apply.
Yes. The AFIP Verify tool provides free forensic analysis for individual content submissions. Enterprise API access with higher throughput and additional features is available for organizations that need to process content at scale.
Video: MP4, MOV, AVI, WebM. Audio: MP3, WAV, M4A, OGG, FLAC. Images: JPEG, PNG, WebP, TIFF, BMP. Text: plain text, URLs for web content analysis. Maximum file sizes apply and are shown on the upload interface.
Accuracy varies by media type and content quality. On standard benchmarks, AFIP achieves 89-97% detection accuracy depending on the media type and generation method. Heavily compressed content and very short samples reduce accuracy. The forensic report always includes confidence indicators so you can assess the reliability of each specific result.
Uploaded content is processed in memory and not retained after analysis. AFIP does not store, share, or use uploaded content for any purpose other than generating the forensic report. Enterprise API users can configure data handling policies to meet their specific compliance requirements.
Yes. The AFIP forensic analysis API provides programmatic access to all detection capabilities. API documentation and access are available for organizations that need to integrate deepfake detection into their own platforms, content moderation systems, or verification workflows.
Watermark-checking tools like those built on C2PA Content Credentials only detect content that was voluntarily labeled at the point of creation. If the creator did not add a watermark, or if the watermark was removed during sharing, these tools report nothing. Forensic detection examines the content itself and works regardless of whether any label was ever applied.
This matters because the vast majority of AI-generated content in circulation is not watermarked. Open-source models do not require watermarking. Bad actors strip labels before distributing misleading content. Social media platforms remove C2PA metadata during upload. Forensic analysis provides detection coverage for this unwatermarked majority, which watermark-only tools miss entirely.
Upload video, audio, images, or text for free forensic deepfake detection.
Open AFIP Verify