Deepfake detection in 2026 works best as a combination of three signals: provenance data such as C2PA Content Credentials, invisible watermarks such as Google's SynthID, and classifier-based AI image detectors such as Hive or Reality Defender. Provenance and watermarks can confirm that a participating AI tool made an image, but say nothing when they are absent. Classifiers can flag unlabeled fakes, but research shows their accuracy drops sharply on real-world content. No single tool is proof either way.
This guide explains how each approach works, what independent research says about accuracy, which tools to use for images, video and audio, and a practical checklist for checking a suspicious file.
Three ways to detect deepfakes
It helps to separate the question "was this made with AI?" into three different techniques, because they fail in different ways.
- Provenance metadata (C2PA Content Credentials). A cryptographically signed record attached to a file that documents where it came from and how it was edited. Cameras, editing apps and AI generators can all write it. It is strong evidence when present, but metadata can be stripped by re-saving, screenshots or platforms that discard it.
- Invisible watermarks (SynthID and others). A signal embedded in the pixels, audio or text itself by the generator. It survives many edits that strip metadata, but it only exists if the generator chose to add it, and only that ecosystem's detector can read it.
- Classifiers (AI image and deepfake detectors). Models trained to spot statistical traces of generation or manipulation. They work on any file, labeled or not, but they produce a probability, and their accuracy depends heavily on whether they have seen similar fakes before.
The practical rule: a positive provenance or watermark result is strong evidence; a negative one tells you almost nothing; a classifier score is a lead to investigate, not a verdict.
How accurate are deepfake detectors?
The most useful independent evidence comes from Deepfake-Eval-2024, a benchmark built from deepfakes actually circulating on social media and submitted by users of a detection platform in 2024. It contains 45 hours of video, 56.5 hours of audio and 1,975 images from 88 websites in 52 languages. When open-source state-of-the-art detectors were tested on it, their performance, measured as AUC, fell by 50 percent for video, 48 percent for audio and 45 percent for images compared with older academic benchmarks. Commercial detectors and models fine-tuned on the new data did better, but still did not reach the accuracy of human deepfake forensic analysts (paper).
The lesson matches what we found for text in our guide on whether AI detectors are accurate: detectors look impressive on the data they were built for and degrade on new generators, compression and real-world editing. Treat any single score with caution.
Deepfake detection tools compared
| Tool | Approach | Media | Who can use it | Key limitation |
|---|---|---|---|---|
| SynthID (Gemini app check) | Watermark | Images, video, audio | Signed-in Gemini users, with daily limits | Only recognizes Google AI content |
| OpenAI verification tool | Watermark plus C2PA | Images, audio | Public preview, plus API | Only recognizes OpenAI content |
| Content Credentials Verify | C2PA provenance | Images, video, audio, documents | Anyone, free | Says nothing if metadata was stripped |
| Hive AI-generated content detection | Classifier | Images, video, audio, text | API, demo and free Chrome extension | Probability score; can miss new generators |
| Reality Defender | Classifier ensemble | Audio, video, images, live calls | Enterprises and platforms via API | Aimed at business deployments |
In plain terms: watermark checkers from Google and OpenAI can confirm content from their own tools; the Content Credentials verifier reads signed provenance from any participating app; and classifiers from Hive and Reality Defender are the only option for unlabeled content, at the cost of uncertainty.
SynthID
SynthID is Google DeepMind's invisible watermark, embedded across Google's generative AI consumer products for images, video, audio and text (Google DeepMind). The simplest way to check a file is to upload it to the Gemini app while signed in and ask whether it was created or edited by Google AI. According to Google's help page, a detected watermark means all or part of the content was made or edited by Google's models, while no watermark means it was not made by Google AI but could still come from another AI system. There is a usage quota of roughly ten image checks, ten video checks and ten audio checks in any rolling 24-hour window (Gemini help).
SynthID matters beyond Google. In May 2026 OpenAI began adding SynthID watermarks, alongside C2PA metadata, to images generated through ChatGPT, Codex and its API, and previewed a public tool that checks uploads for its provenance signals. In July 2026 it extended this to audio and added API access for verification (OpenAI). Note that each company's checker currently recognizes only its own content.
OpenAI's documentation is candid about the limits: watermarks are designed to survive light edits, copying, cropping and screenshots, but heavy cropping, compression or other extensive changes can make them undetectable, and a positive result does not confirm that content is accurate, unedited or shown in the right context (OpenAI help).
Content Credentials (C2PA)
Content Credentials are the consumer-facing name for the C2PA open standard, which records a signed history of a file's origin and edits. They are used for both AI and non-AI content, so a credential can just as easily show that a photo came from a real camera. The free Content Credentials Verify site lets you drop in a file and read any credential it carries. The weakness is simple: if a platform or a screenshot removed the metadata, there is nothing to read, which is why generators increasingly pair C2PA with watermarks.
Hive and Reality Defender
Hive offers AI-generated and deepfake content detection APIs that scan images, video and audio and return confidence scores, plus a free Chrome extension for checking images as you browse (Hive). It is a practical choice for journalists, moderators and anyone who needs a quick classifier opinion on unlabeled content.
Reality Defender focuses on enterprises, detecting synthetic audio, video and images across live calls, meetings and uploaded files, with results fed into existing security workflows. Voice-clone fraud against call centers and executives is one of its main use cases. See the vendor's site for access and pricing.
How to check a suspicious image or video
- Check the source first. Who posted it, when, and where did it first appear? A reverse image search often finds the original or an earlier, unedited version.
- Read the provenance. Drop the file into Content Credentials Verify. A valid credential from a camera or a named AI generator is strong evidence.
- Check for watermarks. If you suspect a Google or OpenAI tool, use the Gemini app or OpenAI's verification tool. Remember a negative result rules out only that vendor.
- Run a classifier. Use a detector such as Hive for a second opinion, and treat the score as a probability.
- Look for physical inconsistencies. Lighting, reflections, text, hands and background details can reveal fakes, but modern generators fix many classic tells, so absence of artifacts proves nothing.
- Verify through another channel. For audio or video of a person asking for money or access, call them back on a known number. This beats any detector for voice-clone scams.
- Record what you found. Note each signal and its result, so others can judge the strength of the evidence.
Pros and cons of each approach
Provenance and watermarks. Pros: when present, they give near-certain answers and explain where content came from. Cons: they only cover participating tools, metadata is easily stripped, and absence of a signal means little.
Classifiers. Pros: they work on any file, including content from tools that add no labels. Cons: probability scores, false positives on heavily edited real media, and large accuracy drops on new generators.
Who should use which tools
Everyday users checking a viral image should start with the source, Content Credentials Verify and a Gemini check. Journalists and fact-checkers should combine all three signals and keep notes. Trust and safety teams moderating at scale need classifier APIs such as Hive's, ideally evaluated on a sample of their own content first, using the methods in our guide on how to evaluate LLM apps. Enterprises worried about voice and video impersonation should look at Reality Defender alongside call-back procedures. For text rather than media, see our guide to whether AI detectors are accurate, and for security testing of AI systems themselves, see LLM red teaming tools.
FAQ
Can deepfakes be detected?
Often, but not reliably by any single method. Provenance metadata and watermarks can confirm content from participating AI tools, and classifiers can flag many unlabeled fakes, but research shows classifier accuracy drops sharply on new, real-world deepfakes.
What is the best free AI image detector?
There is no single best one. For free checks, combine Content Credentials Verify for provenance, the Gemini app for Google's SynthID watermark, OpenAI's verification tool for OpenAI images, and a classifier such as Hive's free Chrome extension for unlabeled images.
How do I check if an image was made with Google AI?
Upload it to the Gemini app while signed in and ask whether it was created or edited by Google AI. Gemini checks for a SynthID watermark. If none is found, the image was not made by Google AI, but it could still come from another tool.
What are C2PA Content Credentials?
They are signed provenance records, based on the open C2PA standard, that document where a file came from and how it was edited, including whether AI tools were used. They can be read with free tools such as Content Credentials Verify, but disappear if the metadata is stripped.
Can watermarks like SynthID be removed?
They are designed to survive common edits such as cropping and screenshots, but heavy cropping, compression or other extensive changes can make them undetectable. That is why a missing watermark is not evidence that content is real.
How can I tell if a voice is a deepfake?
Listen for unnatural pacing or emotion, but do not rely on your ears. The most reliable defense is to hang up and call the person back on a number you already know, and organizations can add audio deepfake detection to their call workflows.