METHODOLOGY
How ImageSignals works
Different image-analysis tasks require different models, signals and levels of certainty.
ImageSignals does not treat every result as the same kind of evidence. Visual predictions, metadata, provenance and generated descriptions each answer different questions and should be interpreted differently.
Different tools use different methods
ImageSignals uses specialised models for different tasks rather than relying on one system for everything.
- Real or AI
- Analyses visual patterns associated with real and AI-generated images.
- Extract Text
- Uses optical character recognition to detect and recover text from images.
- Find Objects
- Uses object detection to identify supported object categories and locate them with bounding boxes.
- Search the Image
- Uses open-vocabulary visual grounding to match a text query with relevant regions in an image.
- Describe Image
- Uses a vision-language model to generate a natural-language description of visible content.
Each model has its own strengths, limitations and confidence behaviour.
AI-image verification uses multiple evidence layers
For Real or AI, ImageSignals keeps different types of evidence separate.
Visual analysis
The visual detector analyses image pixels for patterns associated with real or AI-generated imagery. This is probabilistic evidence.
Metadata
The original file is checked for available information such as EXIF, XMP, editing software and AI-generation metadata. Metadata can provide useful evidence, but its absence does not prove that an image is real.
Provenance
When available, C2PA and related provenance information can provide stronger evidence about how an image was created or edited. Cryptographically verified provenance is fundamentally different from a visual probability score.
Why we keep evidence separate
A visual score, camera metadata and a verified provenance record do not mean the same thing. For example:
- A visual detector may estimate that an image looks AI-generated.
- EXIF may indicate that a file passed through a camera or editor.
- C2PA may provide signed information about creation or editing history.
Combining these into one artificial percentage would hide important differences in the evidence. ImageSignals therefore presents them separately.
Confidence is not certainty
Model confidence reflects how strongly a model favours a prediction based on what it has learned.
It is not a guarantee that the result is correct.
High-confidence predictions can still be wrong, and low-confidence predictions can still contain useful information.
For important cases, results should be interpreted alongside context and other available evidence.
Provenance checks
When local provenance information is available, ImageSignals analyses it separately from visual detection.
For OpenAI and Google provenance signals that are not validated locally, ImageSignals may direct users to external verification tools.
- No image is automatically sent to OpenAI or Google.
- External verification requires the user to upload the file manually.
- Absence of C2PA or SynthID does not prove that an image is real.
Files and privacy
Uploaded images are processed locally on the ImageSignals server and are not retained after processing is completed.
Some tools use temporary image and result files while analysis runs. These temporary files are removed when the job finishes, including when a worker error or timeout occurs.
Reports and image previews remain in your browser for the current session; they are not stored as permanent public result pages.
ImageSignals does not automatically send uploaded images to OpenAI, Google or external provenance services. Local C2PA checks do not fetch remote manifests or online revocation information.
Technical records for usage limits and server operation are separate from image storage.
Known limitations
Automated models can make mistakes
False positives and false negatives are possible.
Image quality matters
Compression, low resolution, blur and other transformations can reduce performance.
Small details can be missed
Tiny objects, small synthetic regions or subtle text may be harder to detect.
New AI systems evolve quickly
Recently released generators or processing pipelines may behave differently from systems represented in training data.
Descriptions are interpretations
Image descriptions are generated by a model and should not be treated as factual verification.
Provenance may be absent
Many real and AI-generated images contain no usable provenance metadata.
Trust & transparency
No hidden certainty
ImageSignals avoids presenting probabilistic model output as definitive proof.
Limitations are visible
Known limitations are explained rather than hidden.
Metrics are labelled correctly
Performance figures distinguish between metrics such as balanced accuracy and ROC AUC instead of presenting everything as “accuracy”.
Evidence types stay separate
Visual predictions, metadata and provenance are not collapsed into one misleading score.
External checks are optional and explicit
Images are not automatically transmitted to external provenance services.
When automated analysis is not enough
Some cases require more context, multiple images or several complementary methods.
For complex or inconclusive cases, ImageSignals offers a separate Advanced Analysis service that can combine multiple signals, larger image sets and custom reporting.
Explore Advanced Analysis →