You can’t reliably identify every deepfake by looking for visual glitches. Treat a detector’s result as one piece of evidence: check where the media came from, examine its context, look for provenance information, and seek independent confirmation. The right method depends on whether you’re checking an image, video, audio clip, or a specific kind of manipulation.
What does “deepfake detection” actually measure?
There is no single test that answers whether any image, video, or audio clip is real. Detection systems are evaluated on particular media and manipulation tasks, and results from one task do not automatically apply to another. For example, NIST’s published performance figures discussed below concern face morphs—altered facial photos that combine features from multiple people—not deepfakes in general. NIST’s OpenMFC program distinguishes media-manipulation detection and localization from deepfake detection, and describes separate image and video tasks.
As an Amazon Associate I earn from qualifying purchases.
A detector may identify signs associated with manipulation in a file; it does not, by itself, establish who made the file, how it circulated, or whether the event it depicts happened. A result is most useful when its scope, test conditions, and error rates are known.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Why do detector results vary?
A system’s performance can change with the type of manipulation, the media available for comparison, the software used to create the content, and later changes such as compression or blur. NIST’s Guardians of Forensic Evidence program highlights whether systems generalize to newer generation methods and withstand post-processing as practical evaluation concerns. Testing should resemble the evidence and conditions in which a tool will actually be used.
#1 Best Overall
That distinction matters when reading accuracy claims. In an August 2025 account of NISTIR 8584, NIST reported the following results for face-morph detection—not for general-purpose deepfake detection:
| Task | Evidence available | Reported NIST result and qualification |
|---|---|---|
| Single-image morph detection | One questionable photo; no known-genuine comparison photo. | In best-case conditions, detection reached up to 100% at a 1% false-detection rate when the detector was trained on examples from the software that generated the morph. On morphs made with unfamiliar software, accuracy could fall well below 40%. |
| Differential morph detection | The questionable photo and a second image known to be genuine. | Best-case accuracy ranged from 72% to 90% across morphs made with the tested open- and closed-source software. The genuine comparison photo is required. |
These figures are reported by NIST for the face-morph use case. They are not a universal detector score, and the best-case results should not be read as a promise of performance on other media, manipulation types, or conditions. NIST computer scientist Mei Ngan, a co-author of the guidance, said some modern morph detectors “could be useful in detecting morphs in real-world operational situations.” Her statement is about face morphs and identity credentials, not synthetic media generally.
Rank #2
How should you check a suspicious image, video, or claim?
Use a detector as one check among several. The steps below can help you assess a questionable file without mistaking a plausible appearance—or a single automated result—for proof.
- Preserve the original file. If possible, keep the file as received rather than relying only on a screenshot, repost, or re-encoded copy. Changes to a file may remove information useful for assessment.
- Trace its source and context. Check who first shared it, when and where it appeared, and whether the account or publication is in a position to know. Look for independent reporting or other confirmation of the depicted event or statement.
- Check for provenance information. A file may carry information about its origin or history. Its presence can add context; its absence does not establish that the media is genuine or manipulated. Technical provenance approaches do not have universal adoption or complete coverage.
- Use an appropriate forensic check. If you use a detector, find out what media type and manipulation it evaluates, what conditions it was tested on, and what false-positive and false-negative rates apply. A face-morph tool’s results, for instance, do not establish whether an audio clip was generated or a video was altered.
- Escalate consequential cases. When a decision could affect someone’s identity, safety, reputation, or access to a service, seek a trained reviewer and a defined process for investigating suspicious results.
NIST’s 2024 overview of synthetic-content transparency treats authentication and provenance tracking, labeling such as watermarking, synthetic-content detection, and testing as distinct technical approaches. They can complement one another: provenance can provide information about a file’s history, while forensic detection evaluates indications in the media. Neither is a standalone guarantee of authenticity.
What should organizations require from a detector?
For remote identity proofing, NIST SP 800-63A calls for submitted media to be analyzed for signs of modification, manipulation, tampering, or forgery. The standard’s guidance is more specific than simply adding a detector to a workflow:
- Test analysis algorithms against both available attack artifacts and genuine media.
- Document expected false-positive and false-negative rates.
- Use manual review to augment algorithmic analysis and automated decisions.
- Use technical measures to increase confidence that media comes from a genuine sensor.
- For attended remote collection, train staff and use random human-in-the-loop cues.
Those requirements come from NIST SP 800-63A, Identity Proofing Requirements. NIST’s face-morph guidance likewise describes a process that combines human review, automated tools, and investigation of images flagged as suspicious. A system’s test results matter, but so does what an organization does when the system is uncertain or wrong.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What questions help you compare detection approaches?
Before relying on a tool or comparing performance claims, ask:
Recommended Free Tools
Quick Recap
Best Value
- What input does it need? Is one image or video enough, or does the method require a second, known-genuine image?
- What was evaluated? Check the media type and manipulation class. Results for face morphs, still images, video, and audio are not interchangeable.
- How similar was the test material to the suspected media? Ask whether the generator was represented in the training or test data, and whether newer generation methods were considered.
- What happens after ordinary processing? Look for testing with transformations such as blur or video compression.
- What are the error rates at the operating threshold? A detection rate without its false-positive and false-negative context is not enough to judge a consequential decision.
- Who reviews uncertain results? For high-stakes use, identify the human reviewer and the escalation or investigation process before relying on automated decisions.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
