Correction (19 October 2026): This article has been corrected on how C2PA works (a signed metadata manifest that can be stripped and is often paired with invisible watermarks), the exceptions and fines under Article 50 of the EU AI Act, the US position (Executive Order 14110 was revoked in January 2025 and California's AI Transparency Act now applies), China's labelling rules, and South Korea's 2024 criminal-law amendments and AI Basic Act labelling duty.
It now takes seconds to put words a political leader never said into their own voice and play them to millions, to fabricate footage of a CEO confessing to corruption, or to produce hyper-realistic synthetic news of bombs exploding in a peaceful region. Truth can no longer stand on its own; it needs a means of verification.
This is where law and technology have converged on a single remedy: the mandatory watermark. But is a watermark just a transparent logo in the corner of an image? How are the European Union, the US, and China approaching the problem? More importantly, how should Turkey, where political polarization is at its peak, and South Korea, caught in a spiral of deepfake sexual violence, respond? This analysis examines the legal effort to protect the truth.
1. How It Works: A Watermark Is More Than a Logo
Most people hear "watermark" and think of the transparent logos TV channels put in the corner of the screen (a visible watermark). That is not the watermark AI law is concerned with. A malicious user can remove a visible logo in two seconds with a simple crop. What the law requires is cryptographic, invisible (metadata) watermarking.
The C2PA (Coalition for Content Provenance and Authenticity) standard, backed by the major technology companies, attaches a cryptographically signed provenance manifest ("Content Credentials") to a file when AI generates an image or video, recording that it is an AI production. The manifest is tamper-evident: edits are either recorded in it or break the signature. But because it is metadata, it can be stripped, for example by re-encoding the file or taking a screenshot. That is why it is increasingly paired with an invisible watermark embedded in the pixels or the audio signal itself (Google's SynthID is one example), which can help recover the provenance information when the metadata is lost. Platforms that read these signals can label the content; Meta, for example, shows an "AI info" label.
2. The European Union Model: The Gold Standard (EU AI Act)
The European Union Artificial Intelligence Act (EU AI Act) sets out the clearest and most binding watermarking mandate anywhere in the world. It imposes a strict "transparency obligation" on providers of generative AI (Sora, Midjourney, DALL-E).
EU AI Act - Article 50 (Transparency Obligations)
The Act is explicit: providers of AI systems that generate synthetic audio, image, video or text content must "ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."
A technology company (US-based OpenAI, for example) that offers a system on the European market without a watermark, or with a watermark that is easily removed, faces fines of up to EUR 15 million or 3% of its worldwide annual turnover, whichever is higher. The obligation does not apply where the system only performs an assistive function for standard editing, does not substantially alter the input data, or is authorised by law to detect, prevent, investigate or prosecute criminal offences. Satire and parody fall under a different rule: deployers who publish deepfakes must disclose them (Article 50(4)), but for evidently artistic, creative, satirical or fictional works the disclosure only has to be made in a way that does not hamper the display or enjoyment of the work.
3. The Other Major Players: The US and China
The rest of the world is less strict than Europe. The US and China have each taken an approach that reflects its own political and economic system.
Global Approaches to Watermarking Regulations
In the US, the Biden administration's Executive Order 14110 encouraged companies, through voluntary commitments, to develop technology that proves content provenance (watermarking), but it was revoked in January 2025 and there is no binding federal mandate. The binding rules come from the states: California's AI Transparency Act requires large generative AI providers to embed latent provenance disclosures in the images, video and audio they generate from 2 August 2026. In China, the Measures for Labeling AI-Generated Synthetic Content, issued by the CAC (Cyberspace Administration of China) with other authorities and in force since 1 September 2025, require both explicit (visible) labels and implicit labels embedded in file metadata on AI-generated content, enforced by the CAC alongside its algorithm-filing system.
4. The Impending Crisis: The Cases of Turkey and South Korea
The two countries most exposed to the destructive effects of AI are Turkey and South Korea, for different reasons.
- South Korea's K-Pop and Nth Room Syndrome: South Korea is dealing with a large-scale deepfake pornography crisis centered on Telegram. Since the scandal broke in 2024, women ranging from schoolgirls and teachers to K-Pop stars have had their faces grafted onto explicit videos made with readily available AI tools. Korea has responded on two fronts: in 2024 it made possessing, purchasing, storing or viewing sexual deepfakes a crime, and its AI Basic Act, in force since 22 January 2026, requires generative AI outputs to be labelled and deepfakes to be clearly indicated. But neither labels nor criminal law undo the harm once content has spread, and public alarm and distrust remain acute.
- Turkey's Election and Polarization Dynamics: Smartphone and social media use in Turkey is at record levels. In a deeply polarized country, a fake voice recording that spreads in seconds over WhatsApp and X (Twitter) can change the minds of millions of voters overnight. Opaque algorithms now threaten the democratic will of Turkish voters.
Burhan Doğuş Ayparlar's View
In countries such as Turkey and South Korea, where society is divided along fault lines and technology adoption is very high while media literacy is relatively low, leaving AI watermarking to 'the initiative of companies', as the US does, or to 'slow-moving bureaucratic penal processes', as the EU does, would be self-defeating.
We need a 'digital notary' (cryptographic provenance) system. A transparent logo in the corner of an image is no obstacle to a malicious actor. What we need are cryptographic provenance standards embedded in the metadata of the image or audio, so that any alteration breaks the signature and becomes visible. The deepfake sexual violence cases in South Korea and the risk of political disinformation in Turkey show that we need a firm, cross-border 'upload filter' law.
My proposal: the new artificial intelligence and cybersecurity laws that the Turkish and South Korean parliaments will consider must make the 'distribution of unwatermarked synthetic content' a standalone offense carrying a severe penalty. Platforms such as X, Telegram, Meta, or TikTok that operate in Turkey or South Korea must be required to build a C2PA-standard 'digital notary' verification mechanism into their systems. When a user uploads an image or audio file, the system must check its origin (provenance) in the background; if it finds a cryptographic AI watermark and the file impersonates a real person, it must 'quarantine' the content before it goes live on the platform. Looking for the truth in a courtroom after a lie has reached millions in seconds does not deliver justice. The truth has to be protected by an algorithmic barrier in the very first millisecond, when the 'enter' key is pressed.
