OpenAI is working on a way to add a “watermark” to information generated by artificial intelligence. The idea is to develop a tool to “statistically watermark” the outputs of its AI systems that generate text.
Each time a system like ChatGPT generates text, the tool would include a secret “unnoticeable” signal indicating where the text came from.
Why is a watermark necessary for AI-generated content?
ChatGPT is an example of how AI can generate high-quality text, but it also raises ethical concerns. Like many other text generation systems before it, ChatGPT could be used to write high-quality phishing emails, harmful malware, or to cheat on school or college assignments. The goal of OpenAI’s watermarking tool is to make it more difficult to employ the output of an AI system as if it were written by a human. This could be useful in preventing academic plagiarism, mass generation of propaganda, and framing someone by imitating their writing style.
what happens if the internet is filled with content generated by hundreds or thousands of AIs?
And of course, the main problem is feedback from the AIs themselves. An AI like ChatGPT feeds on tons of data that thousands of people have generated. But what happens if the internet is filled with content generated by hundreds or thousands of AIs? It would be a big problem if those same AIs start using self-generated content to generate more content. So we have a problem, gentlemen. We’ll see how this field develops to avoid collapse.
How does OpenAI’s watermarking tool work?
The tool works as a “wrapper” over existing text generation systems, using a cryptographic function at the server level to “randomly select” the next token. The generated text would still appear random, but anyone with the “key” to the cryptographic function would be able to discover the watermark. OpenAI engineers have already built a working prototype of the tool, and the hope is to incorporate it into future OpenAI-developed systems.
0 Comments