An OpenAI Tool Can Spot Text Made by ChatGPT, but You Can't Use It

Ever since generative AI started generating content--text, images, videos--knowing what is and isn't made by artificial intelligence has been a central question and concern. In an election year when deepfake problems and AI misinformation are in the spotlight, identifying AI-made materials assumes even greater importance. That makes it especially surprising that OpenAI, the market-leading AI company, has long had an incredibly reliable tool for identifying texts made by its systems available--and it won't release it. AI text-spotting The news comes via the Wall Street Journal, which learned from OpenAI insiders that the project has been "mired in internal debate" for about two years and was judged ready for a public launch for around a year. What's been keeping this critical tech hidden away? Apparently, internal concerns about what releasing it would do to the world--and to OpenAI's profits. On one level, the merits of letting the detection tool out matter seem obvious. If OpenAI released a tool that was incredibly accurate (99.9 percent effective, the Journal reports) at detecting its own text, it would help educators spot when students were using the AI tool to cheat, it would help job recruiters spot when applicants were using AI to answer interview questions, and it would have a thousand other uses when it's important or merely useful to know about the presence of inappropriate or maybe even illegal use of AI-generated text. But that would very likely deter people from using ChatGPT, pushing them onto other AI systems, and hurt OpenAI's bottom line. Both sides of this debate are supported by user data that show people across the world support an AI-detecting tool--by a four-to-one margin--according to the Journal. Data from about 30 percent of surveyed ChatGPT users indicated they'd use the software less if they thought the text was somehow watermarked so it could be detected as AI-made. People simply don't want others to know they're getting an AI boost. OpenAI is also concerned that any such move to make it easy to spot ChatGPT-created text may disproportionately impact non-English speakers, presumably because they're using the app to simply translate text from one language to another. TechCrunch reports the company said it also worries it's easy to circumvent the tool, which works only on AI-generated text from ChatGPT, using relatively low-tech tricks like passing ChatGPT-made content through another AI system. That deception means someone using ChatGPT for malicious reasons could get away with it if any monitor assumed its ChatGPT-spotting tool was a completely effective safeguard. Nevertheless, TechCrunch spotted an OpenAI blog post update that says that it's still pursuing a number of different tools to allow easy checking for any AI provenance, including using metadata, the behind-the-scenes labels added to computer files--for example, photo metadata might say where the shot was taken, and AI metadata could reveal which chatbot made the content. Metadata is "cryptographically signed, which means that there are no false positives," OpenAI notes, meaning that in the future it might be harder to hide the fact you've used AI to make content. Tell me, ChatGPT, what do people ask you? Meanwhile, other researchers are assessing exactly what content people want AIs to produce. Make a guess at the top three before you read on. We suspect that you'll get one or two dead right. The Washington Post summarizes the results neatly: it's all about sex and homework. The data, gathered from a research dataset from AI chatbots using the same core tech as ChatGPT, show that 21 percent of the first query of the day chat requests from a random sample were about "creative writing and role play." This is the biggest category of AI prompts in the data, and it sounds like simple, silly uses like "tell me a story about a fish named Steve." Meanwhile, 18 percent of queries were about seeking help with homework and 15 percent were on work or business topics--showing average users really are trying to get AIs to do their dirty work for them. Only 5 percent of queries were about health and advice, and only 2 percent addressed job seeking. Reporting on the dataset, the Washington Post also points out that while AIs have filters in place to prevent them talking dirty, people do seem keen to try to get around the blocks for "emotional connection and sexy talk." This shows how far AIs have already reached into our society, and that as well as being used for their smarts, they're being used for their softer side--despite the fact that they very clearly have no, er, softer parts. By Kit Eaton @kiteaton