Investigations need different things than platforms, starting with a map of what the tools actually do. The types of deepfake detection tools on the market are not one technology with 60 brand names. They are a handful of method families, and which family a tool belongs to decides whether its output can survive a courtroom.
The method families, counted across 60 tools
We grouped the detection methods recorded for each of the 60 tools in our database into families. Liveness checks lead by a wide margin: 28 tools test whether a live person is present, a family built for onboarding and access control rather than evidence. Face analysis follows at 14, then artifact and signal analysis at 12. Document authentication and injection attack detection record 9 each, behavioral and device analysis 8, voice biometrics 6, and watermark and provenance checks just 4. A tool can sit in several families at once, so the counts describe the market's center of gravity, not 60 tidy boxes.
For an investigator, the ranking reads upside down. The biggest family, liveness, answers a question investigations rarely ask, because case media usually arrives after the fact, with no live subject to challenge. The families that matter for casework sit in the middle of the chart: artifact and signal analysis, which documents what is wrong with the file itself, and provenance checks, which can establish where a file came from when the chain is intact. Only 4 tools record watermark and provenance methods, which is why provenance alone cannot carry a case yet.
Evidence is the other gap. 25 of the 60 tools state a benchmark of any kind. A vendor that cannot name its test set is asking you to put an unmeasured opinion in front of a jury.
What to demand from any family
Evidentiary audit trails. Sensity AI is built for this: forensic documentation of how a conclusion was reached, which matters more than raw speed.
Chain of custody. The tool must log who submitted what, when, and what version of the model scored it.
On-prem or air-gapped deployment. Sensitive case media cannot go to a public API. Reality Defender offers on-prem tiers; ask every vendor.
Independent evaluation. NIST's FRVT program evaluates face recognition; for deepfake detection, ask for third-party benchmark results, not vendor whitepapers.
Explainability. "The AI says fake" does not survive cross-examination. Frame-level evidence and artifact documentation do.
Sources: CISA
Two families investigators underuse
Injection attack detection, recorded by 9 of the 60 tools, exists because the modern fake often never touches a camera: synthetic media is piped straight into a video call or an upload pipeline through a virtual camera. When a case involves a live call that turned out to be generated, this family is the one that was built to notice, and its small count explains why so few vendors can speak to that evidence at all.
Audio is the second gap. 22 tools check audio and only 6 record voice biometrics methods, yet voice cloning is the fraud pattern behind the executive payment scams that land on an investigator's desk. If your caseload is phone fraud, weight those two families heavily in procurement, because the market as a whole does not: it is built for onboarding selfies, and the chart above is what that bias looks like in numbers.
Integration decides whether any of this survives contact with a real evidence workflow. 48 of the 60 tools offer an API, which is good news for case management systems, but an API that returns a bare score is not an audit trail. Ask for the logged model version, the input hash, and the frame-level output in the contract, not in the demo.
Where DeepfakeDetect fits
Our comparison database is the triage step before procurement: search the 60 tools by method family, media type, and API availability on this site, then open DeepfakeDetect to run a case file through the checker and see what a verdict looks like before your agency commits to a platform. Accuracy matters, but admissibility matters more. Buy the tool your prosecutor can defend, not the one with the flashiest demo.