What a deepfake detector looks for
A deepfake is an image, video, or audio creation that uses machine learning to imitate a person or manipulate a likeness. In a still image, the manipulation may replace a face, change an expression, alter age or identity, or blend one person’s features with another source. The result can be convincing at a glance, especially after it has been compressed and shared on social media.
FauxGuard reads available metadata in the processed still image, then requests separate AI-generation and face-manipulation scores from a third-party visual model. Software or generator names are unverified clues and can be copied or changed. C2PA-related text is not a verified edit history; the service does not validate Content Credential signatures or SynthID watermarks.
The output contains model scores, score-based confidence labels, and plain-language reasons. These are not calibrated probabilities of origin or a validated accuracy rate. The detector does not identify the person in the image or establish who created a manipulation.
Why deepfake checks are difficult
A deepfake model can be trained to imitate a particular face, lighting condition, or expression. As the output quality improves, obvious artifacts become less common. Blur, noise, and compression can hide the remaining differences, while a low-resolution screenshot may remove most of the useful file data.
A real photograph can also look artificial after beauty filters, heavy noise reduction, face retouching, or repeated re-encoding. A model may flag an unusual face even when no likeness manipulation occurred. This is why the confidence band and the explanation matter more than a dramatic single score.
The most reliable review combines file evidence with context. Does the image appear in a video or live event? Is there an original source? Does the face match other authenticated images under similar lighting? Has the publisher provided a raw file or verification? A detector can prioritize those questions, but it cannot answer them alone.
- Face replacement or likeness blending
- Expression and identity manipulation
- Synthetic portrait or profile-photo signals
- Editing provenance and generator fingerprints
Still images versus video and audio
FauxGuard currently accepts JPG, PNG, and WebP still images. It does not analyze video, audio, or a live camera feed. A single frame from a video can be checked as an image, but the result describes that frame, not the full recording. Temporal inconsistencies, lip synchronization, frame-level flicker, and audio mismatch require separate video analysis.
This limit should be stated clearly because different deepfake tasks need different evidence. A still portrait, a face-swapped video, a cloned voice, and a generated news clip are related problems but not the same inspection problem. FauxGuard focuses on the file the user can upload today: a still image.
For a video, extract a few representative frames, preserve the original file, and review the full recording with a tool designed for video. Use a still-image result as an initial clue rather than a conclusion about the entire clip.
Signals that can strengthen a review
The first useful signal is source context. Keep the original file, publisher link, and any independently verified edit history. FauxGuard only reads available file clues; it does not authenticate Content Credentials. Browser resizing and conversion may remove metadata before inspection. A camera field, generator name, or editor name cannot prove the history because it can be copied or changed.
The second signal is visual. The model can measure patterns associated with synthetic faces and generative content. A high score is meaningful when it agrees with the source context or provenance. A low score is useful when it agrees with an intact camera workflow, but it still does not prove that the image is untouched.
The third signal is human review. Look at the source, the account history, the timing of the post, the availability of an original, and whether other credible sources contain the same image. A deepfake accusation can cause serious harm, so the evidence threshold should rise with the consequences of the decision.
How to use the deepfake result responsibly
Read the result as a set of clues. If the score is high, ask what supports the finding: a known generator, a manipulation-related model signal, an edit trail, or several of those together. If the score is low, ask whether the image has enough data left for a useful check. A screenshot or low-quality repost may be unable to support a confident answer in either direction.
Keep a record of the original URL, upload time, file name, and any source information. Do not alter the file before checking it. If the image involves a real person, avoid publishing a private or sensitive image while investigating. Use the smallest amount of information needed to resolve the question.
If the stakes are high, involve a qualified human reviewer and give the person affected by the decision a chance to respond. Automated detection should support a fair process, not replace it.
- Check the original file and the source account.
- Compare multiple frames or versions when a video is involved.
- Document what was checked and when.
- Use a human review path for reputational or legal consequences.
What the result does not prove
A deepfake signal does not identify the creator, the victim, the motive, or the time of manipulation. It does not prove that a public statement is false. It also does not prove that a real person consented to the image or that the file was obtained lawfully.
The result is informational. It is not a legal finding, a forensic report, or a substitute for advice from a qualified professional. Use it to decide what to investigate next, not to make a final claim about a person.
FauxGuard’s goal is to make likeness manipulation easier to detect in a first-pass workflow while keeping the limitations visible. A transparent partial answer is more useful than a false sense of certainty.
Checking a profile photo for face manipulation
A suspicious profile portrait raises several separate questions: was the whole picture generated, was a real face altered, or is a genuine photo being used by someone else? The AI-generation and face-manipulation scores address different model signals. Neither score verifies account ownership or consent. A stolen camera photo can receive low scores and still be used by an impersonator.
Before making an accusation, compare the account history and search for earlier uses of the portrait. Avoid uploading intimate or highly sensitive photos just to test a suspicion. Request proportionate verification through an appropriate channel when identity matters, and use qualified human review rather than a still-image score to decide whether a person is genuine.
Reading a face result alongside a source search
A high face-manipulation score and a low AI-generation score are not necessarily contradictory. A camera-origin scene can contain a modified face. The reverse combination may arise for a synthetic landscape or another image without a clear face. Do not average the two numbers or interpret either as an identity-match score. If the model is unavailable, preserve the partial notice rather than reporting a zero.
After reading the result, inspect the earliest available source and compare the claimed event with independent coverage. A matching photograph on an older page can reveal a misleading caption, but it cannot establish which edits occurred between versions. Keep observations, model inferences, and unresolved questions separate in your notes.
Detector
Check an image with FauxGuard
Upload a JPG, PNG, or WebP image and review separate model scores, score-based confidence, readable file clues, and reasons. No account is required. A score is evidence to review, not proof of origin.
Related image checks
Frequently asked questions
Can FauxGuard detect deepfakes in videos?
The current detector accepts still images only. You can check an extracted frame as an image, but video, audio, and live-stream manipulation require a separate analysis workflow.
What types of deepfake images can it check?
It can evaluate face swaps, synthetic portraits, likeness blending, and other still-image manipulation signals when the file contains enough visual or provenance evidence.
Does a high deepfake score identify the person in the image?
No. The result describes the file, not the identity or intent of the person depicted. Identity and attribution require separate verification.
Why can a real face receive a high manipulation score?
Filters, retouching, compression, unusual lighting, or a low-resolution source can make a real face resemble a manipulated one. Review the confidence and reasons instead of relying on the score alone.
Do I need an account to check an image?
No account or login is required. Your image-check balance is accessed through a private recovery code and a browser cookie. Save the code before paying and keep it private: anyone with it can spend your checks. You can restore the balance from the detector or pricing page on another device.
How much does an image check pack cost?
FauxGuard sells 20 image checks for $4.99 USD as a one-time purchase. There is no subscription or daily free allowance. One completed model analysis uses one check. Failed model requests and metadata-only fallback results do not use a check. Review pricing and the terms before paying.
How should I read the detection result?
Read the AI-generation score and the separate face-manipulation score alongside the reasons and file clues. Confidence labels describe score ranges, not a validated accuracy rate. A high score is a reason for further review; a low score does not certify authenticity. An unavailable score is not a low score.
What are the limits of this image check?
False positives and false negatives are possible, especially after resizing, compression, retouching, or changes in image generators. FauxGuard does not verify C2PA signatures or SynthID watermarks, identify a person, or provide a local edit heat map. It accepts JPG, PNG, and WebP still images, not video or audio. Results are informational, not forensic or legal findings.