OpenAI claims to have developed a watermarking method that accurately detects text written with ChatGPT, but there continues to be debate about whether or not this method will probably be released.
Detecting AI-generated text is becoming increasingly difficult as LLMs turn out to be higher at writing content. Educators are faced with the difficult task of determining whether their students accomplished the assignments themselves or just had ChatGPT write them.
In a updated blog postOpenAI announced that it had “developed a technique for inserting text watermarks that we proceed to contemplate as we search for alternatives.”
The company says the strategy is extremely accurate and proof against local manipulation akin to paraphrasing. However, it admits that it just isn’t foolproof. Running the text through a translation system or using one other LLM to rephrase the text will circumvent the effectiveness of the watermark.
Using an AI model to insert a special character, akin to an emoji, between each word after which deleting that character may also render the watermark useless. However, these technical limitations aren’t the one reason the feature has not been released.
ChatGPT is a very useful writing tool for non-native speakers. OpenAI says its research has shown that releasing the watermarking tool could disproportionately affect such groups and stigmatize their use of AI as a useful writing tool.
Alternatives for text metadata
OpenAI engineers are working on ways to make use of metadata as a substitute of watermarking as a technique of text provenance. Images generated by DALL-E 3 have already got C2PA metadata.
According to OpenAI, it's too early to say how effective adding metadata to AI-generated text could be, but it surely has some potential advantages. For one, metadata is cryptographically signed, so there's no risk of false positives.
The problem with using metadata is that it may possibly be easily removed. OpenAI hasn't explained the best way to apply metadata to text, but removing C2PA metadata from AI-generated images is amazingly easy.
Some social media platforms remove metadata when images are uploaded, and taking a screenshot of the image bypasses C2PA. Will similar workarounds be effective on AI-generated text with metadata added?
Once ChatGPT has generated text and added metadata, you may take a screenshot of the text, upload it to ChatGPT, and have the image converted to text. Metadata is a thing of the past.
Bad for business
The other reason OpenAI could also be hesitant to release the tool is that it only recognizes text generated by ChatGPT. If users know that their AI-generated content is definitely recognized, they’ll quickly switch from ChatGPT to a different platform.
The Wall Street Journal reported that OpenAI's tool has been ready for release for a yr and is 99% effective. The report states: “In deciding what to do, OpenAI staff vacillated between the startup's stated commitment to transparency and their desire to draw and retain users.”
A world survey commissioned by OpenAI found that the concept of an AI recognition tool is supported by a 4:1 majority, but an internal survey found that just about a 3rd of ChatGPT users could be delay by an AI text detector.
Users want AI-generated content to be easy to discover, so long as it just isn’t content they generated.
