Reliable deepfake detection does not exist as a single tool you can point at a file. The workable approach in 2026 is layered: check cryptographic provenance metadata first, check for an invisible generator watermark second, and treat AI detector scores as weak circumstantial evidence rather than proof. This article covers what each layer does, where it breaks, what the EU now legally requires, and what businesses and individuals can verify today.

Why deepfake detection classifiers keep letting you down
Most products marketed as deepfake detection are statistical classifiers trained on synthetic media from a particular set of generators, scoring how closely a new file matches those patterns. That works reasonably well on the generation of fakes they were trained on and degrades quickly on anything newer. Accuracy numbers in vendor marketing come from clean benchmark datasets, not from a screenshot of a screenshot that has been re-compressed four times, which is what actually arrives in your inbox.
Generalisation gap
A detector trained before a new model launches has never seen that model’s output, and artefacts shift with every architecture change.
Compression fragility
Transcoding, cropping and re-encoding destroy the pixel-level and frequency-domain cues classifiers depend on.
Base rate problem
Even a 95 percent accurate detector produces a flood of false accusations when almost all the content it sees is genuine.
No audit trail
A probability score is not evidence, cannot be independently reproduced, and cannot be inspected by the other side.
False positives are the underrated risk. An authentic photograph that has been heavily denoised, upscaled or edited can trip a detector, which matters a great deal if the output is used to accuse someone of fabricating evidence.
C2PA and Content Credentials: the content provenance layer
C2PA, the Coalition for Content Provenance and Authenticity, maintains the open specification behind Content Credentials. Rather than guessing whether something is synthetic, Content Credentials attach a cryptographically signed manifest recording what created the file, which tools edited it, and whether generative AI was involved. Version 2.4 was published in April 2026.
Membership and steering committee participation includes Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI and Sony. The specification is also moving through international standardisation as ISO/DIS 22144, Authenticity of information: Content Credentials, currently at Draft International Standard stage, with a separate draft covering PDF. We could not confirm publication of a final ISO 22144, so treat it as in progress.
Signing at capture is the strongest claim provenance can make, because it binds the image to a device certificate before any editing. Hardware support remains a short list.
| Manufacturer | Support |
|---|---|
| Leica | M11-P shipped C2PA support in late 2023; SL3-S added it in January 2025 |
| Sony | ~9 bodies including Alpha 1 II, Alpha 9 III, FX3, FX30, PXW-Z300; Camera Verify service for news orgs since June 2025 |
| Canon | Firmware for EOS R1 and R5 Mark II (July 2025); Authenticity Imaging System for EMEA news orgs since May 2026 |
| Pixel 10 series (Aug 2025) is the first smartphone to sign captures with Content Credentials by default | |
| Nikon | Z6III added C2PA in firmware 2.00 (Aug 27, 2025); service later suspended after a signing flaw was found |
How Content Credentials travel, and where platforms strip them
This is the weak link. Content Credentials live inside the file, usually in an embedded JUMBF box, and anything that re-encodes the file without deliberately preserving that box destroys the manifest. Platform behaviour in 2026 is inconsistent enough that you cannot assume credentials survive a single upload.
| Platform | Behavior |
|---|---|
| Displays a CR icon and a provenance panel — the most complete public implementation among large social platforms | |
| YouTube | Shows C2PA-derived disclosures such as “captured with a camera,” though its transcoding pipeline removes original file metadata |
| TikTok | Reads Content Credentials on upload and applies automatic AI-generated labels, backed by invisible watermark checks |
| Instagram / Facebook / Threads | Read manifests on upload to drive AI info labels, but served files carry neither EXIF nor C2PA data |
| X | Strips metadata on upload and has shipped no provenance display at all |
Two consequences follow. A missing Content Credential proves nothing, because the likeliest explanation is a platform, a screenshot or a messaging app rather than a forger. And if provenance matters to you, publish the derivative but archive the signed master somewhere you control. Anyone who already thinks about machine-readable markup for search will recognise the pattern from our explainer on what is schema markup.

AI watermarking, hashing and fingerprinting
Provenance metadata is a claim made by the creator, watermarking is a claim embedded by the generator, and hashing is a claim you can make about a file you already hold.
SynthID and generator-side AI watermarking
Google’s SynthID embeds an imperceptible signal directly into generated content: pixel-level patterns for images and video, spectral patterns for audio, and token-selection patterns for text. Google has said SynthID has watermarked more than 100 billion images and videos plus roughly 60,000 years of audio, and that verification inside the Gemini app has been used around 50 million times.
Google extended verification beyond a single app during 2026. You can ask Gemini whether a file was made with AI, Google added C2PA reading to the Gemini app with Search and Chrome to follow, and it published an AI content detection API on Google Cloud. OpenAI, ElevenLabs and Kakao have also adopted SynthID. The limits are structural rather than something a better model fixes: it only finds content from participating generators, robustness weakens against determined adversarial attack (text is the most fragile case), and absence of a watermark is not evidence of authenticity.
Hashing and perceptual fingerprinting
When metadata is gone, hashing is the fallback. A cryptographic hash such as SHA-256 proves two files are byte-identical, which is invaluable for chain of custody and useless after any re-encode. Perceptual hashes and fingerprints, the family of techniques behind content matching on video and audio platforms, survive compression and small edits and let you match a suspect file against a known original. The business workflow is unglamorous and effective: hash and timestamp every asset you publish, keep the originals, and produce the signed master on demand.
The evidence problem: AI-altered forensic files
The most consequential story in this space during 2026 was not a viral celebrity deepfake. In May 2026 researchers disclosed a vulnerability in the software and file formats used to store computerised scans of physical DNA evidence, reported exclusively by the Wall Street Journal.
Be precise about what this was: a research demonstration and a coordinated vulnerability disclosure, not a documented criminal case. The researchers were Laura Gaydosh Combs, a forensic scientist at the University of New Haven, Sarah Chu of the Perlmutter Center for Legal Justice, and Nathan Adams of Forensic Bioinformatics.
Adams used an AI coding assistant to write code that added and removed DNA profiles from evidence scan files without leaving a detectable trace, and reported doing it in roughly 45 minutes, including getting past file encryption using publicly available keys. The affected software lineage has been in use since 1995, putting about three decades of stored files and more than 200 laboratories in scope. Thermo Fisher Scientific privately acknowledged the issue in July 2026 and shipped an update adding digital signatures to the files. There is no public evidence the weakness was ever exploited in a real case.
Combs framed the problem better than any vendor pitch: these are data files that lack the tamper-evident markings the field already requires for a paper evidence bag. That is exactly the gap Content Credentials try to close for photographs, and it is why courts, forensic labs and newsrooms are converging on the same answer. Sign at the source, log every transformation, and make tampering detectable rather than trying to spot it afterwards.
What the EU AI Act now requires
Article 50 of the EU AI Act, its transparency chapter, became applicable on 2 August 2026. It reaches providers and deployers whose systems affect the EU market wherever they are based, and creates four duties.
- Providers of AI that interacts directly with people must make clear the user is dealing with an AI, unless that is already obvious.
- Providers of generative systems must embed machine-readable markings identifying synthetic audio, image, video and text, and provide a means of detecting it, with narrow exceptions for assistive editing that does not substantially alter the input.
- Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them.
- Deployers who publish deepfakes must disclose that the content is artificially generated or manipulated, with carve-outs including material that has had substantive human editorial review.
The European Commission adopted guidelines on these obligations on 20 July 2026, and the AI Office published a voluntary code of practice on transparent generative AI; signatories receive a presumption of conformity and more favourable enforcement treatment. Penalties for Article 50 breaches run up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.
Voice clones, executive impersonation and what businesses should do
Synthetic audio is the highest-yield attack in this category. There is no visual channel to give the fake away, and the target is usually being deliberately rushed. Cloning needs only a short reference sample, and executives publish plenty of usable audio in earnings calls, podcasts and conference videos. The defences that work are procedural, not technical.
No voice-only authorization
Adopt a hard rule that no change to bank details or wire instructions is ever executed on a voice or video instruction, regardless of who appears to be asking.
Callback verification
Require callback verification on a number pulled from your own directory, never a number supplied during the request.
Shared challenge phrase
Agree a shared challenge phrase for high-value approvals, used by finance teams and, separately, by the families of senior staff.
Dual authorisation
Enforce dual authorisation above a monetary threshold inside the payment system itself, so it cannot be waived under pressure.
Tabletop the scenario
Tabletop a cloned-executive wire request the same way you would tabletop ransomware, and measure whether anyone escalated.
Watch unmanaged tools
Know what your staff are already feeding into unmanaged tools — recorded sales and support calls uploaded to a consumer speech app are precisely the training data an impersonator wants.
This is the discipline covered in our post on shadow AI security platforms. The music and entertainment side of synthetic audio has its own licensing rules, covered separately in AI voice cloning and AI music tools.
A verification workflow you can actually run
Whether you are a journalist, a paralegal or someone checking a suspicious video of a public figure, the sequence is the same and needs no specialist software.
- Get the original file rather than a screenshot or a re-shared copy. Ask the source for the camera file or a direct download.
- Inspect provenance. Open the file in the free Content Credentials verify tool, and note whether a manifest exists, who signed it, and whether that certificate is still trusted.
- Check for a generator watermark. Ask Gemini whether the file was made with AI, which reads SynthID and increasingly C2PA data as well.
- Do the unfashionable open-source checks. Reverse image search for earlier copies, compare shadows, reflections, signage and weather against the claimed time and place, and look at whether the posting account existed a month ago.
- Contact the claimed source directly through a channel you already trusted before this file appeared.
- Record what you did. A dated log of the checks you ran is worth more than a detector percentage, because someone else can review and repeat it.
For individuals, treat any unexpected, urgent audio or video as unverified until you have called the person back on a number you already had. For your own material, keep the signed original and publish from a property you control, because your own site is the one place provenance is not stripped on upload. If you are rebuilding anyway, put asset preservation into the website design and development brief rather than retrofitting it.
Frequently Asked Questions
Is there a reliable deepfake detection tool?
No single tool is reliable across generators, compression levels and file types. Detection classifiers are useful as a triage signal, but provenance metadata such as C2PA Content Credentials and generator watermarks such as SynthID give far stronger answers because they are cryptographic or embedded claims rather than statistical guesses.
What are Content Credentials and how do I check them?
Content Credentials are a cryptographically signed record of how a file was created and edited, defined by the C2PA specification, now at version 2.4. You can inspect them using the free Content Credentials verify tool in a browser, or by asking the Gemini app about a file, which reads both C2PA data and SynthID watermarks.
Can SynthID detect AI images made by other companies?
Only partly. SynthID finds its own watermark, so it covers Google models and partners that have adopted it, including OpenAI, ElevenLabs and Kakao. It cannot detect output from generators that never embedded a watermark, so a negative result is not proof that content is authentic.
Does the EU AI Act require watermarking AI-generated content?
Effectively yes for providers of generative AI. Article 50 has applied since 2 August 2026 and requires machine-readable marking of synthetic audio, image, video and text plus a detection mechanism, with deepfake publishers separately required to disclose. Fines reach 15 million euros or 3 percent of worldwide turnover.
Was DNA evidence actually faked using AI?
It was demonstrated by researchers, not documented in a criminal case. In May 2026 a team including Laura Gaydosh Combs, Sarah Chu and Nathan Adams disclosed that AI-assisted code could undetectably alter DNA evidence scan files, and Thermo Fisher acknowledged the flaw in July 2026 and shipped a signing update. No exploitation in a real case has been made public.
Need to protect your organization’s content and communications from synthetic media risk?
WD Morgan Solutions is a veteran-owned digital marketing and development firm in St. Petersburg, Florida, helping businesses across Tampa Bay and nationwide with website design and development and marketing automation, including preserving original assets, metadata and verification trails. Visit wdmorgan.com or get in touch for a free consultation.

























