投稿者: BUNKERZ

Although technology for identifying AI-generated content

Although technology for identifying AI-generated content is evolving, it remains difficult to make a definitive determination using a single tool. Currently, the following methods and approaches are the most widely used.

1. Digital Watermarking Technology

SynthID: A technology developed by Google DeepMind that embeds digital “watermarks”—imperceptible to the human eye or ear—into images, audio, video, and text. This makes it possible to reliably identify content generated by AI.

C2PA (Content Authentication Information): An international standard that records the history of content (creator, edit history, AI usage, etc.) as metadata. An increasing number of platforms and tools now support this standard.

2. Probabilistic Detection (AI Detectors)

There are many tools, such as ChatGPT and Gemini, that calculate the “probability that a text was written by AI.” These tools are trained on large volumes of both human-written and AI-generated text and analyze statistical patterns (such as vocabulary choices and structural tendencies).

However, these tools may mistakenly identify “human-written text” as AI-generated (false positives), so they are not 100% reliable.

3. Metadata and Contextual Verification

In addition to technical detection, “fact-checking”—which involves verifying the credibility of the source, the publication date, and ensuring there are no inconsistencies with seasonal elements, lighting, or historical context within the content—is extremely important. For particularly critical information, it is recommended to cross-reference it with official verification tools and trustworthy sources.

Current Status and Conclusion

As AI-generation technology advances, detection technology finds itself in a constant “cat-and-mouse game.” Therefore, rather than relying on a single technology, a “multi-layered approach” that combines verification of watermark data, cross-referencing of metadata, and logical, objective substantiation of information has become the latest recommended standard.

To put it simply, while it is technically extremely difficult to completely circumvent digital watermarks, methods for evading detection under specific conditions are currently being researched and discussed.

We will now explain the possibilities and limitations of “circumvention” with current technology.

1. Common Approaches to Circumvention

The invisible digital watermarks embedded by generative AI often add identifying information by making minute changes to parts of the data. In response, attempts are being made to destroy these watermarks by introducing “noise” through the following operations:

For text: Methods have been proposed to disrupt the statistical patterns of watermarks through large-scale human rewriting, translation, or regeneration (rewriting) using a different model.

For images and videos: Examples include changing resolution, applying filters, adding noise, compression (such as JPEG compression), and cropping.

2. Why Is “Complete Evasion” Difficult?

High Robustness: Recent watermarking technologies (such as Google’s “SynthID”) are designed to withstand common manipulations like cropping, compression, and color correction, and are engineered so that the information does not disappear with minor edits.

A Cat-and-Mouse Game: While watermark removal technology evolves, the detection capabilities of AI models are also improving, and more robust embedding methods continue to be developed.

Fragmentation of Information: Watermarks are often distributed “holographically” throughout the entire dataset, so even if you crop a portion of an image or a few lines of text, there is still a high probability that traces indicating the data was AI-generated will remain.

3. Risks and Precautions

Forcing data processing in an attempt to circumvent watermarks can lead to the following adverse effects:

Quality Degradation: The quality of the output itself—such as image noise or unnatural phrasing in text—deteriorates significantly.

Detection Reliability: The more one attempts to completely remove the watermark, the more the original data structure is destroyed; conversely, there is a risk that the content will be identified as AI-generated based on other indicators, such as “artificial manipulation.”

In conclusion, while watermarks are not “absolutely impossible to remove,” the current reality is that it is extremely difficult to reliably evade detection while maintaining content quality at a practical level.