Watermark in ChatGPT: how to tell if a photo was generated by AI

What SynthID is and how it works

SynthID is a watermarking technology designed to identify machine-generated content without compromising the user experience.
Unlike traditional watermarks, it is not a visible label, logo, or overlaid text that affects the visual quality of an image. Instead, it is an invisible signal embedded directly into the file structure.

The technical mechanism behind SynthID is highly sophisticated: it subtly modifies the distribution of the data that makes up image pixels.
These changes follow a precise and controlled mathematical pattern that does not alter what the human eye perceives, but changes how the image is encoded at a numerical level.

This “hidden signature” has key characteristics:

Statistical analysis: to detect SynthID, tools do not “look” at the image like a human would, but analyze its internal structure, searching for micro-statistical traces left during generation.

Resilience: because the signal is deeply embedded in the pixels, SynthID is designed in theory to withstand manipulations that would normally remove standard metadata, such as screenshots, resizing, compression, or format changes. However, tests show that these operations do not always preserve its integrity: in some cases, the watermark may become difficult to detect after such transformations.

Versatility: although originally developed for images, Google has already applied this technology to over 100 billion images and videos, as well as 60,000 years of audio content.

ChatGPT has integrated SynthID

Since May 2026, major AI platforms have begun converging toward a new transparency standard for generated content: SynthID.
OpenAI has integrated the technology developed by Google DeepMind into images produced by ChatGPT and its API systems, as part of a broader approach that also combines the C2PA system for content traceability.

The announcement was confirmed at Google I/O 2026, where Sundar Pichai emphasized the goal of building a shared ecosystem for authenticating AI-generated content, adopted across multiple companies in the sector.

Why this technology matters

a) Fighting disinformation
AI-generated images are becoming indistinguishable from real photographs. Invisible watermarking like SynthID introduces a structural layer of verification, even if it does not fully eliminate manipulation risks.

b) Ecosystem transparency
The goal is to enable verification directly within services (search engines, browsers, AI assistants), making it easier to distinguish synthetic from photographic content.

c) Non-intrusive approach
SynthID does not alter image quality: traceability is embedded invisibly, without aesthetic watermarks.

d) Multi-layer system
SynthID works alongside C2PA: a standard that adds structured metadata to digital files describing their origin and history.
It can include who created the content, with which tool, and when, as well as any subsequent modifications. The aim is to make digital provenance more transparent and reduce ambiguity about authenticity.

Limitations and challenges

The first limitation is lack of universality: technologies like SynthID only work within ecosystems that implement or support them, so global coverage remains partial.

The second is vulnerability to modifications: even when content is marked, it can be cropped, recompressed, remixed, or combined with other elements, which may reduce detection reliability.

Another issue concerns content types: watermarking systems work better on images or structured data, while for short texts or fragmented content their effectiveness decreases due to limited space for embedding robust signals.

Finally, there is dependency on adoption: if only some companies implement these standards, the result is a fragmented ecosystem without a unified verification framework.

The spread of systems like SynthID and C2PA represents an important step toward greater transparency in digital content, but it does not definitively solve the problem of online authenticity.

These technologies introduce more advanced verification tools than in the past, but they still depend on external factors such as large-scale adoption, cross-platform compatibility, and resistance to post-processing manipulation.

Ultimately, the issue is not purely technological but systemic: trust in digital content cannot rely on a single standard, but rather on a combination of tools, practices, and shared infrastructures.

At the same time, the role of the user remains essential: no watermarking system can fully replace critical thinking in evaluating sources, context, and informational coherence.

The direction is therefore toward a hybrid ecosystem in which technology and human awareness must evolve together to reduce the growing ambiguity between real and synthetic content.