\n<\/p>\n
OpenAI has had a system for watermarking ChatGPT-created text and a tool to detect the watermark ready for about a year, reports The Wall Street Journal<\/em><\/a>. But the company is divided internally over whether to release it. On one hand, it seems like the responsible thing to do; on the other, it could hurt its bottom line.<\/p>\n<\/div>\n OpenAI\u2019s watermarking is described as adjusting how the model predicts the most likely words and phrases that will follow previous ones, creating a detectable pattern. (That\u2019s a simplification, but you can check out Google\u2019s more in-depth explanation<\/a> for Gemini\u2019s text watermarking<\/a> for more). <\/p>\n<\/div>\n Offering any way to detect AI-written material is a potential boon for teachers trying to deter students from turning over writing assignments to AI. The Journal<\/em> reports that the company found watermarking didn\u2019t affect the quality of its chatbot\u2019s text output. In a survey the company commissioned, \u201cpeople worldwide supported the idea of an AI detection tool by a margin of four to one,\u201d the Journal <\/em>writes. <\/p>\n<\/div>\n After the Journal<\/em> published its story, OpenAI confirmed it\u2019s worked on watermarking text in a blog post update today<\/a> that was spotted by TechCrunch<\/em><\/a>. <\/em>In it, the company says its method is very accurate (\u201c99.9% effective,\u201d according to documents the Journal<\/em> saw) and resistant to \u201ctampering, such as paraphrasing.\u201d But it says techniques like rewording with another model make it \u201ctrivial to circumvention by bad actors.\u201d The company also says it\u2019s concerned about the stigmatization AI tools\u2019 usefulness for non-native speakers.<\/p>\n<\/div>\n But it seems OpenAI is also worried that using watermarking could turn off surveyed ChatGPT users, almost 30 percent of whom evidently told the company that they\u2019d use the software less if watermarking was implemented. <\/p>\n<\/div>\n Despite that, some employees still reportedly feel that watermarking is effective. In light of nagging user sentiments, though, the Journal <\/em>says some suggested trying methods that are \u201cpotentially less controversial among users but unproven.\u201d In its blog post update today, the company said it\u2019s \u201cin the early stages\u201d of exploring embedding metadata. It says it\u2019s still \u201ctoo early\u201d to know how well it will work, but that because it\u2019s cryptographically signed, there would be no false positives. <\/p>\n<\/div>\n