Media Verification by Gretchen AI:
Questions and Answers
Media Verification Dashboard
What was the Gretchen AI Dashboard developed for, and what exactly does it do?
Our Gretchen AI Media Verification Dashboard integrates two essential tasks to support users in detecting misinformation:
- AI-powered deepfake detection to verify image authenticity: i.e., analyzing whether an image or video has been manipulated or synthetically generated.
- Context analysis to check whether the image or video is presented within its original, truthful context or has been taken out of context.
Is the effectiveness of the tool scientifically proven?
Yes, we have benchmarked our tool against scientific benchmarks and other state-of-the-art (SOTA) methods. On benchmarks such as Deepfake-Eval-2024, we achieved a performance increase of up to +44% compared to the state-of-the-art.
Why is our tool superior to others?
We analyze images across multiple scales—ranging from minute image sections (patches) for local artifact detection to the full image. Our model leverages state-of-the-art Vision Foundation models and possesses comprehensive knowledge acquired by training on a vast spectrum of deepfake modalities.
Deepfake-Erkennung zur Berechnung des Authentizitäts-Scores
What is the image deepfake detection tool intended for?
This tool helps users identify manipulated or synthetic (i.e., AI-generated) images. It analyzes visual and digital signals within the image to generate a consolidated "authenticity score." This score estimates the likelihood that the image has been artificially generated or altered.
IMPORTANT: It does not replace human judgment but serves as a decision-making aid. Whenever possible, always combine authenticity score results with context analysis.
What types of deepfake detection does Gretchen offer?
Two types of detection mechanisms are integrated:
- Human Faces: The first modality focuses exclusively on faces (up to five per image). It checks whether a face has been altered via face-swapping, face-reenactment (modifying emotions or lip synchronization), or face-editing (altering traits, e.g., aging). These manipulations usually occur locally, while the rest of the image remains unaltered.
- Synthesis Detection: The second mechanism analyzes the entire image to determine if it was fully AI-generated. The system also flags anomalies if only a small portion has been modified by AI tools such as Midjourney, Qwen, or Nano Banana.
Although both methods operate independently, both indicate when an image fails to meet authenticity criteria.
Does Gretchen AI provide a visual marking of the manipulated areas (heatmap)?
Yes, we provide visual indications that highlight which image regions contributed to the model decision in all enterprise plans. What matters to us here is transparency about what these results actually mean: such a marking shows what our detector based its decision on, not forensically proven evidence of manipulation. For localized interventions such as face swaps, it can be informative; for fully AI-generated content, by contrast, there is no single "manipulated spot" that could be meaningfully circled. That is why we always combine the visual indication with a calibrated confidence score and – where required – traceable documentation, so that the result remains robust even in compliance and evidentiary contexts.
What do I see on the dashboard after uploading my content?
Once you have uploaded an image or video to our dashboard, you can choose whether to run a quick deepfake detection or a comprehensive context analysis that includes deepfake detection.
The resulting authenticity scores are displayed directly beneath the respective images. Authentizitäts-Scores werden direkt unter den jeweiligen Bildern angezeigt.
Below this, you will find the results of our web-wide search for reference images, which illustrate the distribution of the discovered contextual information. Date, action, location, person, and source are researched in a single step and compiled into a clear, concise evaluation. This allows you to identify consistencies and contradictions at a glance.
How should the level of the authenticity score be interpreted?
A very low score (below 10%) means that there is strong evidence of synthetic generation or manipulation. The range up to approximately 50% indicates a high probability that the image in question has been altered. Conversely, a verdict of "Certainly authentic" (around 100%) signifies a high degree of confidence regarding the image's authenticity. Because the estimation involves multiple steps that can occasionally be ambiguous, we deliver these results with total transparency.
We believe that disclosing algorithmic certainty—even in borderline cases—is key to building trust.
Why is the result sometimes "inconclusive"?
AI-generated media is evolving rapidly. New models frequently leave behind digital fingerprints that are not yet recognized by the detection model. On the other hand, genuine images can exhibit artifacts caused by heavy editing, filters, or compression that appear suspicious to the detector. Modern smartphone cameras also frequently deploy AI filters (often automatically). In such scenarios, we recommend further investigation using the dashboard's metadata and context analysis features.
What types of deepfake models can the tool detect?
Our detection models are trained on data from over 120 different deepfake creation algorithms and AI models. These include methods that either manipulate real images (face swaps, regional changes) or create entire synthetic images from scratch.Technically covered methods include, among others: FaceShifter, FaceSwap, SimSwap, DeepFaceLab, HeyGen, VQGAN, StyleGAN3, SD-2.1, Midjourney 6, Gemini 2.5 Flash, Nano Banana, Grok-Imagine, and many more."
What are the limitations of the tool?
The tool operates best with images in their original resolution. Screenshots or scaled versions frequently lose vital signals. Noise or heavy JPEG compression increase uncertainty. Currently, the tool is optimized for faces and human-centric imagery; inanimate objects or landscapes are more challenging to detect.
Can the tool make mistakes?
In short: yes. No AI tool is perfect. There are two error scenarios:
- False Positives: A genuine image is flagged as a fake.
- False Negatives: A manipulated image is classified as genuine.
This is why we adhere to the philosophy of "assist, don't replace." Our dashboard supplements forensics with context analysis for a 360-degree evaluation to provide optimal support for fact-checkers.
Do blurring filters (e.g., "beautifiers") affect detection?
If these filters utilize generative AI to a significant extent, the detector will be triggered. However, since the impact of these filters varies widely, we are continuously working to improve robustness against such influences.
The API
What are the prerequisites for the Gretchen AI API?
To use the Gretchen AI API, you require an active Gretchen AI account and an API key generated within your dashboard. Authentication occurs via this key in the request header. Uploaded content must be provided in a supported format. Further details can be found in our documentation..
What are the advantages and disadvantages of integration?
API integration allows you to embed Gretchen AI's deepfake detection directly into your own systems and workflows—such as editorial, verification, or content management tools. This enables automated and scalable verification of large volumes, as well as real-time assessment of live data, such as ongoing streams or broadcasts. Everything integrates seamlessly into your existing processes without requiring manual uploads. Alternatively, our web interface remains available at any time for occasional individual verifications.
Media Context Analysis
What does the context analysis do?
Operating independently of the authenticity score, it searches the web for identical, similar, or event-related images. Results are visualized in clusters. Clicking a reference image reveals which sources are disseminating specific information (date, location, etc.), rendering contradictions instantly visible. Graphic color-coding represents the source types and can be customized.
How up-to-date are the search results?
Our search index is updated regularly. Major websites are indexed on average twice a day. Dashboard results reflect the status on the day of analysis. For the most current web data, please initiate a new search.
Can the timeline provide inaccurate results?
Yes, timestamps can occasionally be misleading. This frequently occurs with archival pages of media outlets when the timestamp displays the date of the archive page rather than the original publication date of the article in which the image appeared.
What happens if no reference image or original is found on the web?
Even if no direct duplicate is located, the system displays similar images. Independent of the search results, the authenticity score for your image is always calculated.
What is the event-based image search?
Event-based image search complements traditional reverse image search by adding an expanded perspective. Rather than just finding identical copies, it helps determine the temporal and thematic context of an event. In doing so, it also identifies different captures of the same occurrence from various angles. This creates a more comprehensive and easily context-driven overview.