How to Read Your Scan Results
Reality Defender returns a decision-ready signal with enough transparency to stand up in high-trust environments.
Most scans return three things:
Overall Results Score — a number produced by Reality Defender's models indicating the likelihood that the content has been manipulated or AI-generated
Final Conclusion — the actionable label (Authentic, Suspicious, or Manipulated) determined by Reality Defender's models. When enabled, Context and Metadata analysis may also influence this conclusion.
Model indicators — visual signals showing which detection models were triggered
The Overall Results Score reflects the likelihood of manipulation — not the amount of manipulation present, and not whether every frame, second, or pixel is synthetic.
NOTE: When Metadata analysis drives the Final Conclusion, no Overall Results Score appears. The conclusion is set directly by that analysis. Model indicators remain accessible in the RD Models tab.
Each detection model specializes in a different type of manipulation or synthesis method. Some models look at artifacts, others at context, content patterns, temporal dynamics, or linguistic signals. The Final Conclusion reflects a weighted consensus across all models that were triggered.
Final Conclusions
Reality Defender normalizes results into three Final Conclusions across audio, image, and video. Each conclusion corresponds to a score range on a 0–100 scale.
Authentic (0-44) — low likelihood of manipulation or AI generation; continue normal workflows
Suspicious (45-55) — moderate likelihood of manipulation; worth a closer look or human review
Manipulated (56-100) — high likelihood that the content has been manipulated or AI-generated; trigger escalation or secondary validation
A higher Overall Results Score means a higher likelihood of manipulation — not that the entire file is synthetic.
When RealScan cannot return a standard result
In some cases, RealScan may not be able to return a standard Authentic or Manipulated result. This can happen when media is skipped during preprocessing, when the file does not contain enough usable signal for analysis, or when a processing issue prevents Reality Defender from completing the scan.
When this happens, the scan result may include a plain-language reason, such as limited usable audio, a model-specific evaluation issue, or a preprocessing error.
NOTE: Detailed error explanations appear only for new imports created after July 16, 2026. Older scans may still show a generic inconclusive or unable-to-evaluate result because they do not include the fields required to display detailed explanations.
🔈 Understanding Audio Results
When analyzing audio, Reality Defender evaluates indicators of voice synthesis, voice cloning, or audio splicing. The RD Models tab displays the Overall Results Score alongside model indicators showing which models were triggered.
Models used
Advanced — detects AI-generated audio using a larger foundation model trained on highly diverse synthesized-speech datasets. Looks for neural vocoder artifacts, frequency-domain anomalies, and model-specific generation fingerprints.
Generalizable — detects AI-synthesized audio using style and linguistic patterns unique to real human speech. Looks for unnatural prosody, overly consistent tone or pacing, and style mismatches common to audio LLMs.
How the Final Conclusion is determined
Both models produce independent signals, which are combined into a single Overall Results Score. Model indicators show which models were triggered.
📸 Understanding Image Results
Image detection evaluates synthetic or manipulated visual signals across GAN, diffusion, and traditional image-editing workflows. Context and Metadata analysis are available as additional layers and can be enabled or disabled by admins via settings.
Results are organized into three tabs: RD Models, Context, and Metadata.
RD Models
The RD Models tab displays the Overall Results Score alongside model indicators showing which models were triggered. Models analyze both facial regions and the full image frame.
Face-focused models:
GANs — detects faces generated or manipulated using GAN-based methods. Looks for StyleGAN fingerprints, resolution-frequency mismatches, and geometric distortions.
Diffusion — detects images created by diffusion models (e.g. Midjourney, SDXL). Looks for diffusion sampling artifacts, uniform noise fields, and overly smooth textures.
Faceswaps — detects traditional and modern faceswap-based manipulations. Looks for boundary inconsistencies, compositing artifacts, and identity mismatches.
Visual Noise Analysis — detects fake images by analyzing texture and distribution of visual noise. Looks for diffusion-grid artifacts, upsampler inconsistencies, and GAN-style frequency patterns.
Full-frame
Full-frame models extend the same detection capabilities — GANs, Diffusion, Content Swap, and Universal — across the entire image frame rather than just facial regions. They extend detection to images where faces are small, blurry, or absent, and can catch manipulation signals in the body, background, or surrounding context.
Note: models are optimized for images containing people.
Context
Context model analysis uses a vision-language model to evaluate an image for signs of AI generation, informed by contextual signals such as Reverse Image Search results, visible AI-generation watermarks, filename indicators, source context, and concrete visual artifacts. It relies on stronger evidence, so it may miss subtle or clean AI-generated images and is not meant to detect every kind of image manipulation.
The Context tab appears whenever Context analysis is enabled for a scan. Results are shown as two distinct signals: General Context, which provides a plain-language explanation when contextual signals suggest the image may have been manipulated. In addition, Reverse Image Search may also appear under a separate header with a clickable link to scrollable details.
When Context analysis is enabled but no signals are found, the tab displays: “No context signals to suggest this media has been manipulated.”
Note: Because Context Aware requires additional processing, enabling it may increase scan latency compared with scans where Context Aware is disabled. Admins can enable or disable Context Aware in Settings → RealScan → Enable context models.
When context models are enabled, Reality Defender may also send relevant media data to optional third-party sub-processors for contextual signal analysis. See our current list of sub-processors for details.
Metadata
Metadata analysis examines provenance signals embedded in the image file for indicators of AI generation or manipulation. This includes software tags embedded by common generation tools such as Stable Diffusion, Midjourney, DALL-E, and GPT Image, as well as C2PA provenance records indicating AI generation.
When markers are found, they appear as AI Markers in this tab and the Final Conclusion is set to Manipulated. When no markers are found, an explicit empty state is shown — this means no provenance signals were detected, not that the image is confirmed authentic.
Metadata analysis is available for images only and can be enabled or disabled by admins via settings.
When Metadata drives the Final Conclusion, no Overall Results Score appears.
How the Final Conclusion is determined
Each model contributes a signal, weighted by relevance. Face-focused and full-frame models are combined into a single Overall Results Score.
When Metadata analysis detects AI generation markers, it overrides the model ensemble and sets the Final Conclusion to Manipulated. When no markers are found, models run as usual and the Overall Results Score reflects the ensemble output.
Context analysis runs alongside the models — if Context returns Manipulated but the models return Authentic, the Final Conclusion is set to Manipulated with no Overall Results Score. Model indicators remain accessible in the RD Models tab.
🎥 Understanding Video Results
Video analysis evaluates frame-level, temporal, and contextual indicators of deepfake generation. The Overall Results Score is displayed alongside model indicators showing which models were triggered.
Models used
Context Aware — uses a vision-language model to evaluate video frames for signs of manipulation. Particularly useful for video where faces are absent, small, or unclear.
Dynamics — detects deepfake faces generated with a variety of methods by analyzing temporal information. Looks for frame-to-frame inconsistencies, motion artifacts, and lip-sync irregularities.
Guided — focuses on specific facial features known to differ between real and generated faces. Looks for eye-region anomalies, facial microexpression inconsistencies, and local generation fingerprints.
Universal — detects deepfake faces generated across many methods, including GANs, diffusion, and hybrid approaches. Looks for global artifact patterns and multimethod synthesis signals.
Context
When context models are enabled, Reality Defender may use a vision-language model to analyze video frames directly, alongside the core video detection models. This can provide additional signal when faces are absent, small, unclear, or when the relevant evidence appears outside a detected face region.
When Context analysis is enabled for a video scan, results appear in the Context tab as General Context. Video Context results do not include Reverse Image Search, since it does not apply to video.
Note: Because Context Aware requires additional processing, enabling it may increase scan latency compared with scans where Context Aware is disabled. Admins can enable or disable Context Aware in Settings → RealScan → Enable context models.
When context models are enabled, Reality Defender may also send relevant media data to optional third-party sub-processors for contextual signal analysis. See our current list of sub-processors for details.
How the final score is determined
Temporal models (Dynamics and Context Aware) take heavier weight, with Guided and Universal refining the final Overall Results Score. Model indicators show which models were triggered.
🔡 Understanding Text Results
Text detection evaluates whether text has been generated or heavily edited by a large language model (LLM).
Models used
Text Detector – Generative
Detects linguistic patterns characteristic of LLM-generated text.
Looks for:
Over-optimized phrasing
Statistical smoothness
Predictable structural patterns
Low-variance word choice
How the final score is determined
Since text uses a single model type, the Overall Results Score reflects the model's likelihood assessment directly.
If Your Result Seems Unexpected
Content may test as manipulated due to:
Heavy compression or filtering
AI-assisted editing
Image upscaling or enhancement
Non-human voices (e.g., IVR, TTS, synthetic accents)
AI-generated copy mixed with human text
Blend of real and generated segments
If you have questions about a specific file, use the "Report an Issue" button at the top right corner of your Scan Results page, or contact support@realitydefender.com.
Downloading Reports
Every completed scan can be exported as a PDF or CSV report containing: upload timestamp, detected modality and file metadata, triggered model indicators, Overall Results Score, and Final Conclusion with explanation.
Reports are timestamped and can be shared internally for record-keeping or audits.