{"id":2001,"date":"2023-02-08T00:56:24","date_gmt":"2023-02-08T00:56:24","guid":{"rendered":"https:\/\/promptmuse.com\/?p=2001"},"modified":"2023-04-07T10:14:56","modified_gmt":"2023-04-07T10:14:56","slug":"openais-watermarking-to-halt-misuse-of-gpt-3-and-chatgpt-outputs","status":"publish","type":"post","link":"https:\/\/promptmuse.com\/openais-watermarking-to-halt-misuse-of-gpt-3-and-chatgpt-outputs\/","title":{"rendered":"OpenAI&#8217;s Watermarking to Halt Misuse of GPT-3 and ChatGPT Outputs"},"content":{"rendered":"<div class=\"flex-1 overflow-hidden\">\n<div class=\"react-scroll-to-bottom--css-nexvi-79elbk h-full dark:bg-gray-800\">\n<div class=\"react-scroll-to-bottom--css-nexvi-1n7m0yu\">\n<div class=\"flex flex-col items-center text-sm h-full dark:bg-gray-800\">\n<div class=\"w-full border-b border-black\/10 dark:border-gray-900\/50 text-gray-800 dark:text-gray-100 group bg-gray-50 dark:bg-[#444654]\">\n<div class=\"text-base gap-4 md:gap-6 m-auto md:max-w-2xl lg:max-w-2xl xl:max-w-3xl p-4 md:py-6 flex lg:px-0\">\n<div class=\"relative flex w-[calc(100%-50px)] flex-col gap-1 md:gap-3 lg:w-[calc(100%-115px)]\">\n<div class=\"flex flex-grow flex-col gap-3\">\n<div class=\"min-h-[20px] flex flex-col items-start gap-4 whitespace-pre-wrap\">\n<div class=\"markdown prose w-full break-words dark:prose-invert dark\">\n<p>OpenAI, the leading artificial intelligence research laboratory, is taking steps to prevent users from misusing the content generated by their AI models. With the increasing popularity of their GPT-3 and ChatGPT models, it has become evident that some users are using them for unethical purposes, such as cheating on homework, creating fake news, and making social media bots. To tackle this issue, OpenAI has decided to implement a watermarking system to detect and track the misuse of <a href=\"https:\/\/promptmuse.com\/seo-master-class-with-a-i-1\/\">AI-generated content<\/a>.<\/p>\n<h2>How Effective is OpenAI&#8217;s Watermarking System?<\/h2>\n<p>According to <a href=\"http:\/\/www.cs.umd.edu\/~tomg\/\" target=\"_blank\" rel=\"noopener\">Tom Goldstein<\/a>, Perotto Associate Professor in the Department of Computer Science at the University of Maryland, OpenAI&#8217;s watermarking system can detect AI-generated content with an accuracy rate of <strong>99.999999999994%<\/strong>. This high level of accuracy is possible because the watermarking system adds a unique signature to the content generated by GPT-3 and ChatGPT.<\/p>\n<blockquote class=\"twitter-tweet\">\n<p dir=\"ltr\" lang=\"en\"><a href=\"https:\/\/twitter.com\/hashtag\/OpenAI?src=hash&amp;ref_src=twsrc%5Etfw\" target=\"_blank\" rel=\"noopener\">#OpenAI<\/a> is planning to stop <a href=\"https:\/\/twitter.com\/hashtag\/ChatGPT?src=hash&amp;ref_src=twsrc%5Etfw\" target=\"_blank\" rel=\"noopener\">#ChatGPT<\/a> users from making social media bots and cheating on homework by &#8220;watermarking&#8221; outputs. How well could this really work? Here&#8217;s just 23 words from a 1.3B parameter watermarked LLM. We detected it with 99.999999999994% confidence. Here&#8217;s how \ud83e\uddf5 <a href=\"https:\/\/t.co\/pVC9M3qPyQ\" target=\"_blank\">pic.twitter.com\/pVC9M3qPyQ<\/a><\/p>\n<p>\u2014 Tom Goldstein (@tomgoldsteincs) <a href=\"https:\/\/twitter.com\/tomgoldsteincs\/status\/1618287665006403585?ref_src=twsrc%5Etfw\" target=\"_blank\" rel=\"noopener\">January 25, 2023<\/a><\/p><\/blockquote>\n<p><script async src=\"https:\/\/platform.twitter.com\/widgets.js\" charset=\"utf-8\"><\/script><br \/>\nHowever, the effectiveness of the watermarking system decreases if the user makes significant changes to the content. Tom Goldstein explains that <strong>to remove the watermark by hand, the user needs to swap around 40% to 75% of the tokens<\/strong> in a long passage of text. This would result in a low-quality output and may not even fool the watermarking system.<\/p>\n<p>More advanced attacks, such as a synonym attack or rewriting the text, are possible, but they would result in a decrease in quality, closer to that achievable by the attack model. Furthermore, if the user performs a synonym attack or hand-editing, they would need to ensure that they do not leave any long spans of unedited text, as even a span of 36 tokens (<em>36 to 72 words<\/em>) can trigger the watermarking system with <strong>high<\/strong> confidence.<\/p>\n<h2>The Futility of Tricking AI Detectors<\/h2>\n<p>The implementation of OpenAI&#8217;s watermarking system highlights the futility of attempting to trick AI detectors. With the increasing sophistication of AI models, it has become nearly impossible to fool them into believing that the content generated is human-written. Attempting to trick AI detectors would result in low-quality content that would be easily recognizable.<\/p>\n<p>As AI continues to play a larger role in our daily lives, it&#8217;s crucial to ensure that it is used for the benefit of society, rather than for malicious purposes. OpenAI&#8217;s watermarking system is a testament to this effort, providing a layer of accountability and security for AI-generated content. The system&#8217;s ability to accurately detect and deter attempts at manipulation serves as a warning to those who seek to exploit AI for their own gain. As AI technology evolves, it is heartening to see companies like OpenAI taking proactive steps to ensure that it is used for the greater good. As the saying goes, &#8216;<strong>with great power comes great responsibility,<\/strong>&#8216; and OpenAI is demonstrating its commitment to this ideal.<\/p>\n<div class=\"w-full h-32 md:h-48 flex-shrink-0\"><\/div>\n<\/div>\n<p>&nbsp;<\/p>\n<\/div>\n<\/div>\n<\/div>\n<div class=\"absolute bottom-0 left-0 w-full border-t md:border-t-0 dark:border-white\/20 md:border-transparent md:dark:border-transparent md:bg-vert-light-gradient bg-white dark:bg-gray-800 md:!bg-transparent dark:md:bg-vert-dark-gradient\">\n<form class=\"stretch mx-2 flex flex-row gap-3 pt-2 last:mb-2 md:last:mb-6 lg:mx-auto lg:max-w-3xl lg:pt-6\">\n<div class=\"relative flex h-full flex-1 md:flex-col\">\n<div class=\"flex ml-1 mt-1.5 md:w-full md:m-auto md:mb-2 gap-0 md:gap-2 justify-center\"><\/div>\n<\/div>\n<\/form>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>OpenAI, the leading artificial intelligence research laboratory, is taking steps to prevent users from misusing the content generated by their AI models. With the increasing popularity of their GPT-3 and ChatGPT models, it has become evident that some users are using them for unethical purposes, such as cheating on homework, creating fake news, and making<\/p>\n","protected":false},"author":1,"featured_media":2005,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_lock_modified_date":false,"footnotes":""},"categories":[6,405,404,21,17],"tags":[],"class_list":{"0":"post-2001","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-a-i","8":"category-gpt3-2","9":"category-gpt4","10":"category-latest-news","11":"category-news"},"featured_image_src":"https:\/\/promptmuse.com\/wp-content\/uploads\/2023\/02\/sirenjpg.jpg","blog_images":{"medium":"https:\/\/promptmuse.com\/wp-content\/uploads\/2023\/02\/sirenjpg-300x200.jpg","large":"https:\/\/promptmuse.com\/wp-content\/uploads\/2023\/02\/sirenjpg.jpg"},"acf":[],"ams_acf":[],"_links":{"self":[{"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/posts\/2001"}],"collection":[{"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/comments?post=2001"}],"version-history":[{"count":0,"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/posts\/2001\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/media\/2005"}],"wp:attachment":[{"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/media?parent=2001"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/categories?post=2001"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/promptmuse.com\/wp-json\/wp\/v2\/tags?post=2001"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}