What types of images can I scan with Winston AI?
Overview
Winston AI's image detection engine analyzes photos and images to determine whether they were created by a human or generated by an AI tool. It also identifies deepfakes and images that have been manipulated using tools like Photoshop or other editing software. Two scan types are available depending on how much detail you need.
Basic Scan vs. Advanced Scan
Basic Scan
Cost: 200 credits
The Basic Scan runs an AI image classifier and returns:
- AI generation probability score — the likelihood the image was AI-generated vs. human-created
- Image metadata — any available Exif, C2PA, and IPTC data embedded in the image
Results are displayed inline in the app and can also be downloaded as a PDF report.
This is a fast, lightweight check suitable for quickly screening images when you don't need forensic detail.
After completing a Basic Scan, you can upgrade to an Advanced Scan for an additional 300 credits (500 credits total).
Advanced Scan
Cost: 500 credits
The Advanced Scan runs a full forensic analysis using multiple detection methods and produces a written conclusion. It includes:
- AI Image Classifier — same probability score as the Basic Scan
- Metadata Extractor — reads C2PA manifests, IPTC, and Exif data; C2PA manifests from tools like ChatGPT/OpenAI provide conclusive provenance evidence when present
- ELA (Error Level Analysis) — detects post-generation pixel-level manipulation or splicing
- Residual Noise Maps — identifies noise patterns typical of synthetic texture generation
- Edge Anomaly Heat Map — highlights edge inconsistencies that may indicate compositing
- CFA Pattern Analysis — checks for the Bayer/CFA sensor pattern present in real camera images; absence is a strong indicator of AI generation
The Advanced Scan also provides a written Conclusion synthesizing all findings — including cases where individual tools disagree (e.g. the classifier returns a false negative but metadata confirms AI origin).
Results are displayed inline in the app. You can also download and share the results as a PDF report.
Use the Advanced Scan when you need to document your findings, investigate ambiguous results, or present evidence to others.
Plan Availability
| Plan | Basic Scan | Advanced Scan |
|---|---|---|
| Trial | 5 scans included | 1 scan included |
| Essential | Yes | No |
| Advanced | Yes | Yes |
| Elite | Yes | Yes |
How to Submit an Image
You can submit images in two ways:
- Upload a file — drag and drop or browse to select an image from your device
- Paste a URL — submit a direct link to an image hosted online
Image Type Reference
| Image Type | Works Well? | Notes |
|---|---|---|
| Photographs (people, places, objects) | Best results | Realistic photos are ideal. The Advanced Scan's forensic tools are especially effective here. |
| AI-generated artwork and illustrations | Limited (Basic) / Good (Advanced) | The AI classifier is optimized for photorealistic images, so the probability score may be less reliable for stylized or illustrative content. The Advanced Scan's forensic tools (CFA pattern analysis, metadata) can still detect AI-generated artwork reliably. |
| Deepfakes and manipulated images | Best results | Detects face swaps, synthetic faces, and images that have been manipulated using tools like Photoshop or other editing software. |
| Social media profile photos | Good | Works well for detecting AI-generated or synthetic profile pictures. |
| Stock photo-style images | Good | Detects AI-generated stock imagery reliably. |
| Screenshots | Limited | Screenshots of interfaces, apps, or text are not useful for AI image detection. The scanner analyzes image content, not UI. |
| Heavily edited or filtered images | Limited | Significant post-processing can affect detection signals. The Advanced Scan's ELA and noise analysis can help identify whether editing occurred post-generation. |
| Images with large text overlays | Limited | Watermarks and graphic overlays obscure image content and may reduce classifier reliability. |
| Low-resolution or heavily compressed images | Limited | Very small or heavily compressed images lose visual detail. Higher-resolution versions produce more reliable results. |
| Diagrams, charts, and infographics | Not suitable | Graphic design outputs lack the photographic characteristics the detection engine relies on. |
| Illustrations and hand-drawn art | Not suitable | Traditional illustration and graphic art are not the intended target for this scanner. |
Factors That Affect Accuracy
C2PA Metadata
When an image contains active C2PA manifests — provenance data embedded by AI generation tools like ChatGPT, OpenAI/Sora, or Adobe Firefly — this provides conclusive evidence of AI origin, even if the visual classifier returns a different result. The Advanced Scan surfaces this in its conclusion. C2PA data is often preserved in original files but stripped when images are shared on social media platforms.
CFA / Bayer Pattern
Real camera images contain a characteristic sensor pattern (CFA/Bayer pattern). AI-generated images do not. The CFA Pattern Analysis in the Advanced Scan uses this as a strong signal — a missing CFA pattern is a reliable indicator that the image was not captured by a physical camera.
Image Resolution and Quality
Higher resolution images give the model more detail to analyze. Heavily compressed images may produce lower-confidence classifier results, though the Advanced Scan's forensic tools can still provide useful signal even in those cases.
Post-Processing and Editing
Significant cropping, filtering, or compositing can affect classifier accuracy. The ELA and edge anomaly tools in the Advanced Scan are specifically designed to detect whether post-generation manipulation occurred.
Composite and Mixed-Source Images
Images that combine real photography with AI-generated elements will produce mixed signals. The written conclusion in the Advanced Scan is the best guide for interpreting these cases.
Tips for Best Results
- Use the Advanced Scan when you need a defensible conclusion or are investigating an ambiguous result
- Submit the highest-resolution version of the image available
- Use a direct image URL when possible — avoid URLs that redirect to a web page
- If C2PA metadata is present, it is the most authoritative signal; check the metadata section of your results
- For social media images, download the original file when you can — platforms often strip metadata
- Cross-reference with a reverse image search to check whether the image appears elsewhere online