AI for Verifying Media Integrity: What Actually Works in 2026

Verifying media integrity means answering two questions about an image, a video or an audio file. Where did it come from, and has it been changed since? AI helps with both, in three different ways. Some tools read signed provenance data attached when the file was made. Others look for invisible watermarks that AI generators embed in their output. A third group tries to spot manipulation from the content alone. These methods answer different questions, and reading a detector's score as a final verdict is a common and costly mistake.

The question has also become a compliance one. Since 2 August 2026, Article 50 of the EU AI Act requires deployers to disclose deepfakes and AI-generated text published on matters of public interest, unless the text has gone through human review or editorial control. Providers of generative systems must also mark their outputs in a machine-readable way. Systems already on the market before that date have until 2 December 2026 for the marking part, under the Digital Omnibus on AI. The European Commission has published a voluntary Code of Practice on marking and labelling AI-generated content to show how this can work in practice.

The three layers of verification

1. Provenance with Content Credentials

C2PA (the Coalition for Content Provenance and Authenticity) defines Content Credentials, a cryptographically signed record attached to a file when it is captured or edited. The record says which device or app produced the file and what was done to it. If someone alters the file after signing, the signature no longer matches.

Hardware support is growing. Google's Pixel 10 attaches Content Credentials to every JPEG taken with Pixel Camera, and its camera app reached Assurance Level 2, the highest level then defined by the C2PA conformance programme. Google frames the result as media that comes with verifiable proof of how it was made, versus media that does not. It deliberately avoids an "AI or not AI" label.

The weak point is that metadata gets stripped. Some platforms remove it on upload, and the stripped file gives no sign that credentials ever existed. Durable Content Credentials address this by pairing the signed record with an invisible watermark and a content fingerprint, so the original manifest can be found again in an online repository. The practical consequence matters more than the mechanism. A file without credentials proves nothing, because most authentic photos in circulation have none.

2. Invisible watermarks

AI generators can embed an imperceptible signal in the pixels of an image or in the frequencies of an audio track. Google's SynthID is the largest system of this kind. According to Google, it has watermarked over 180 billion images and videos and 240,000 years of audio since 2023. On 7 October 2026 the SynthID Detector portal opened to everyone, globally in English. It recognises content made with Google's tools and with those of partners including OpenAI, NVIDIA and Kakao, with Apple announced as coming.

A watermark detector only recognises its own marks. A negative result means "no SynthID found". It does not mean the file is authentic, because content from generators that do not use SynthID will not be flagged.

3. Deepfake detection

Detectors analyse the content itself (lighting, edges, compression traces, audio spectra) and return a probability that it was manipulated or generated. They are the only option when a file has neither credentials nor a watermark, and also the least reliable.

The best evidence comes from Deepfake-Eval-2024, a benchmark of deepfakes that circulated online in 2024. It includes 45 hours of video, 56.5 hours of audio and 1,975 images from 88 websites in 52 languages. On this material, open-source detectors lost 50% of their AUC on video, 48% on audio and 45% on images compared with the academic benchmarks they were built on. Commercial models and models fine-tuned on the new data did better, and still fell short of human forensic analysts.

What each method can and cannot tell you

A verification workflow for newsrooms and comms teams

1. Find the earliest copy. Run a reverse image search, or extract keyframes from a video and search those. Many supposed deepfakes are real images from a different event, and this step settles them in minutes.

2. Check provenance. Open the file in a Content Credentials inspector. Valid credentials are strong evidence. Missing credentials are not evidence of anything.

3. Check for watermarks. Upload the highest-quality version you can find to the SynthID Detector, and to any other detector run by the generators you suspect.

4. Run a deepfake detector last. Treat its score as a lead to investigate. Never publish "the image is fake" on the strength of a detector alone.

5. Verify the claim around the media. Who is said to appear in it, and where and when was it supposedly taken? These are fact-checking questions, covered in our guide to AI fact-checking techniques.

6. Record what you checked. If you publish or respond, keep a note of each step and its result. If you publish AI-generated imagery yourself in the EU, label it.

Where Scova AI fits

Scova AI works upstream of the file-level checks. It monitors news sources and blogs, so a team finds out quickly when a manipulated image or a false story about its organisation starts appearing in the press. The text that comes with it can be fact-checked in one click. The checks above then happen on the specific asset. For the wider picture, see our guide to detecting AI disinformation in corporate communications.

FAQ

What does media integrity mean?

Media integrity means that an image, video or audio file is what it claims to be. Its origin is known, and any changes since capture are recorded or absent. Verifying it combines provenance checks, watermark detection and manipulation analysis.

Can AI reliably detect deepfakes?

Not on its own. On real-world deepfakes collected in 2024, open-source detectors lost between 45% and 50% of their AUC compared with academic benchmarks, and the best commercial tools still trailed human analysts. Detection works best as one signal among several.

If an image has no watermark or Content Credentials, is it real?

No. Most authentic images carry neither, and some platforms strip metadata on upload. A missing signal tells you nothing in either direction. Only a positive, valid signal is informative.

What is the difference between C2PA and SynthID?

C2PA Content Credentials are signed metadata that describe how a file was made and edited, and they can apply to camera photos as well as AI output. SynthID is an invisible watermark embedded in AI-generated content by Google and partner generators. The two are complementary, and some providers use both.

What does the EU AI Act require for deepfakes?

Under Article 50, applicable since 2 August 2026, deployers must disclose that image, audio or video content is a deepfake. Providers of generative AI systems must mark outputs in a machine-readable way, with a deadline of 2 December 2026 for systems already on the market before August. For evidently artistic, satirical or fictional work the disclosure can be lighter, so that it does not spoil the work.