With the increasing use of AI to generate images, video and text, detecting content created by AI models has become one of the major challenges in the digital space. The removal of visible watermarks has made this task harder and raised the question of how users, journalists and online platforms can verify the origin of content when the obvious sign of AI generation is gone.
The answer to this question is not summed up in a single tool or signal. New technologies for identifying synthetic content rely on a set of methods such as invisible watermarks, provenance metadata, cryptographic standards, digital fingerprinting and analysis of the content itself.
An invisible watermark
One of the new approaches by tech companies is to abandon visible watermarks and place an invisible signal within the content itself; a method that is harder to remove.
One of the best-known examples is Google’s SynthID. This technology can embed an invisible watermark in AI-generated content, including images, video, audio and text.
Google DeepMind says the SynthID watermark in images and videos is designed to remain detectable against some common alterations, such as cropping, applying filters, changing frame rate and compression.
For text the mechanism is different. The system creates a statistical pattern by making controlled changes to the probability of selecting words when generating content; a pattern that is not visible to the reader, but a system equipped with the detection key can identify.
Google has also stated that videos generated by the Gemini Omni model have the invisible SynthID watermark and that checking it is possible through some Google services.
C2PA; a digital passport for content
Another layer of content verification technologies is the C2PA standard. This standard can be considered a kind of "digital passport" for media files.
Instead of placing a visible logo on an image, C2PA records information related to the origin of the content and changes made to it as cryptographically signed data.
This information can include details about how the content was created, edits performed, the device or software used, and the relationships between different versions of a file. The data are digitally signed to allow verification of their authenticity and detection of possible tampering.
OpenAI also uses this approach for images generated with its tools and employs C2PA data alongside the SynthID watermark.
However, C2PA is not immune to removal. For example, editing a file or even taking a screenshot can remove provenance information. In such cases, the embedded watermark may still be able to provide information about the origin of the content.
What if all signals are removed?
Here one of the most important limitations of AI content detection technologies becomes apparent: the absence of a watermark or provenance data alone does not prove that a piece of content was produced by a human.
If a system can identify a SynthID watermark, it can be considered strong evidence that the content passed through a tool equipped with that technology. But if no watermark is found, one cannot conclude that the content is necessarily human-made.
Provenance data may have been removed, the watermark may have been damaged during editing, or the content may have been produced by a tool that does not support such technologies.
For this reason, AI detection tools should not be used as definitive proof; rather they are best used as part of a multilayered process to verify content authenticity.
An invisible watermark for text is also on the way
The challenge of detecting synthetic content is not limited to images and video. AI-generated text has also become a major area of competition among tech companies and researchers.
Anthropic has announced that future Claude models will use a kind of imperceptible watermark in text.
This method works by influencing the model’s word choices. When several words are semantically suitable continuations for a sentence, the model can steer the choice among them so that a particular statistical pattern emerges in the text.
This pattern is not visible to the ordinary reader, but a system that holds the detection key can assess the likelihood that the text was produced by the model.
However, Anthropic itself acknowledges the limitations of this method. Completely rewriting a text can remove the watermark, while minor edits may not necessarily eliminate its effect entirely.
On the other hand, identifying such a pattern can only increase the probability that Claude was involved in producing or editing the text and does not necessarily indicate who created the content or who owns it.
When the technology turns to the content itself
If watermarks and provenance data are not available, another method remains: analysis of the content itself.
Traditional AI-detection tools try to identify statistical patterns and features that may be more common in content produced by models.
In text, this method can include examining linguistic patterns, sentence structure, word choice and stylistic features. In images, video and audio, algorithms can look for statistical patterns or specific technical features.
But this method differs fundamentally from watermarking. In watermarking, the model creator intentionally embeds a detectable signal into the content from the start; in content analysis, the system attempts to infer the origin of the content from its inherent features.
For this reason, the error rate in this method is higher and its results should be combined with other evidence.
A continuous competition between producers and detectors
As detection technologies advance, methods to circumvent them are also being developed.
Various studies have shown that under certain conditions it is possible, with specific modifications, to weaken or remove watermarks from AI-generated content without noticeably changing the overall appearance of the content.
This situation has created a technical arms race: on one side companies and researchers try to design more robust watermarks, and on the other, investigators seek methods to detect or bypass these signals.
As a result, it is unlikely that relying on a single technology will provide a permanent reliable way to identify all AI-generated content.
The future of content detection; a combination of layers
The most realistic solution is to use multiple verification layers simultaneously. An invisible watermark can be one layer of this system, while C2PA data, file provenance records, edit histories, content analysis and searches for original versions provide complementary information.
In such a model, the main question is no longer just "Does this image have a watermark or not?" but rather where the content came from and what changes have been made to it since its creation.
The C2PA standard is designed precisely for this purpose: to record the creation and modification history of content as verifiable information. Technologies like SynthID also create an additional layer that can leave a signal in the content itself even if some external data are lost.
Ultimately, detecting synthetic content will likely become more of a digital investigation than finding a single "mark"; a process in which a set of technical evidence is assembled to determine the origin of an image, video or text.
As AI models become better at producing content indistinguishable from real examples, verifying provenance and using digital fingerprinting will become important components of information validation on the internet.