This invisible watermark will be a part of OpenAI’s efforts to comply with the transparency provisions of the EU AI Act.This invisible watermark is set to be added to OpenAI’s products to meet the transparency provisions of the EU AI Act.
The company said it will introduce the technology gradually to eligible ChatGPT and Codex users in the EU in the plans available. The watermark will be built into the text itself, and won’t be visible to the naked eye, so it can only be detected by specialized systems that are built to recognize text generated by OpenAI models.
Why Is OpenAI Adding Watermarks to AI-Generated Text?
The transfer relates to the EU AI Act’s transparency rules. Providers of some AI systems are obligated to ensure that AI-generated or manipulated content be “technically distinguishable” with rules that can be detected by other systems.
The text watermarking method OpenAI plans to use aims to give a method to identify whether a text was created or edited by one of its AI systems while not offering an overt watermark to readers.
The watermark will be initially rolled out in the European Union “not as a global standard,” says the company.
How Does OpenAI’s Invisible Text Watermark Work?
The watermark does not show as a logo, character or special symbol in the content.
Rather, the system lightly nudges the word selection of the model during the generation of response. The small changes generate a statistically hidden pattern in the text. A special detector can then read the writing, and check to see if the pattern is there or not, using the correct key.
The watermark is one advantage, it is part of the text itself. So, it may be detectable in the target document or site if the user copy/pastes the content generated by this component.
OpenAI has also shared technical details regarding its method, called textGrain, which details the watermarking process and detection method.
Will the Watermark Affect ChatGPT’s Performance?
OpenAI says they didn’t see any real difference in model performance during testing when the watermarking system was turned on.
This is crucial as it can potentially impact on factors like response quality, fluency, or the model’s ability to follow instructions in the generated text.
According to OpenAI, their current test shows that this method can be applied without major alterations to the quality of the model’s output.
Can AI Text Watermarks Be Removed?
You can make the watermark on the paper less clear or get rid of it. The tests conducted by OpenAI revealed that the ability to detect AI-generated content can be complicated if there are some relatively small changes to the text. In one of the company’s tests, for instance, using synonyms instead of some of the actual words decreased the detection rate.
Other pieces of content can be more challenging to spot, too. When the watermark is contained in short passages, maths answers, or translated text, it might not contain enough information for the detector to reliably identify the watermark.
This implies of course that the watermark created by an AI is not a true AI-content detection system.
A Missing Watermark Does Not Prove Human Writing
OpenAI is also stressing a key caveat: When a detector does not discover a watermark, that does not necessarily indicate that the content was completely human-made.
There are a number of possible causes for non-detection of a watermark. The text could be too short, heavily edited, translated, or generated by a different AI system.
Likewise, if it were found that a video had an OpenAI watermark, it would not be a sure sign of the extent to which it was generated by humans or by the AI.
ChatGPT can generate a draft, after which a user can spend a lot of time researching, editing, rewriting and adding original content.
Hence, watermark detection can be viewed more as a sign of an AI System’s involvement than a sign of a human or AI System’s authorship.
OpenAI to Limit Initial Detector Access
Due to these restrictions, OpenAI intends to first roll out its detection technology to trusted researchers and expert organizations.
The aim is to give experts an opportunity to assess the reliability of the technology and explore the potential for the use of AI-generated-content detection.
Such a method may also be useful for OpenAI to identify where watermarking is causing false-positive results in watermark detection before it opens up to wider access.
OpenAI’s Text Watermarking Plans Extend Beyond the EU
While the first release of ChatGPT and Codex is available to European Union users, OpenAI is also offering the technology via its API to developers who choose to use certain models.
This watermarking feature can be enabled for eligible models by developers worldwide but is not enabled by default.
Therefore, at launch OpenAI has decided not to adopt text watermarking as a default feature.
OpenAI Follows a Growing Industry Trend
OpenAI’s announcement follows other leading AI companies attempts to find AI-generated content.
Anthropic has also said that it will add watermarking to text produced by Claude, and multiple big tech firms have agreed to help with the identification and labeling of AI-generated content.
The wider sector trend reflects the increasing demand for transparency from AI companies as generative AI is used more and more across the education, business, publishing, software and online content creation sector.
What Does This Mean for Writers and Businesses?
The implementation of invisible watermarking could raise awareness among writers, marketers, developers, students, and businesses about AI-generated content when using ChatGPT.
The technology is not meant to be a substitute for human review or editorial decisions, however. AI-generated content can be significantly modified, interwoven with human content, or creatively processed using other tools.
The bottom line for businesses is that transparency laws regarding AI are becoming increasingly a critical component of the generative AI ecosystem.
As governments introduce new AI regulations and technology companies come up with the ways to detect generated content, organizations will have to take note of how they are creating, editing, publishing and disclosing AI-assisted content.
Conclusion
The invisible text watermark is another feature to make AI-generated content more recognizable, but not by marking it with obvious labels.Another feature is the invisible text watermark, which is another layer of making AI-generated content more recognizable without obviously labelling it.
The technology consists of creating a hidden pattern within the process of selecting words for the model, which can then be detected by a special detector. But the results of OpenAI’s own testing reveal significant drawbacks, especially in cases of heavily edited, translated, or extremely short text.
The EU rollout is mainly spurred by the AI transparency regulations in that region, but developers outside EU can use the technology for certain API models.
With the growing integration of generative AI into everyday content production, tools like text watermarking may assume a greater significance in pinpointing the origin of AI-generated material for researchers and organizations.
