An existing tool that OpenAI developed and tested can identify whether any written content was produced with ChatGPT. The Wall Street Journal claims that the tool's company is delaying its public release due to multiple concerns.
By adding a pattern to the large language model's (LLM) writing style, the tool enables OpenAI to discern whether or not ChatGPT generated the output. But the pattern is still invisible to humans, therefore it has no effect on the quality of the LLM. Although OpenAI has not yet released the program, internal documentation claims that it is 99.9% successful in identifying ChatGPT's output.
While text watermarking is a very useful tool for identifying ChatGPT-authored information, it is incompatible with output from other LLMs, such as Llama 3 or Gemini AI. Moreover, it is simple to go around this tactic. To effectively remove the watermark, you can, for instance, paste ChatGPT's output into Google Translate, convert it to another language, and then return it to English.
The detection tool can be fooled by simply instructing the AI to insert a special character, emoticon, or brief phrase between words, which can then be deleted later (using Microsoft Word's Find and Replace function, for example) or by asking another LLM to completely reword the output.
Another issue that OpenAI has with the tool's release is that it claims it might be biased against writers who are not native English speakers. A generative AI text identification tool was first made available by the business last year. However, because to its poor detection rate and tendency to produce false positives, OpenAI terminated it after just seven months. It has even gone so far as to cause a professor to fail a whole class since all of the papers that were turned in were mistakenly labeled as being created by artificial intelligence.
Additionally, OpenAI takes into account its clientele; according to consumer surveys, 69% of ChatGPT users think that the tool will lead to unfounded allegations of AI cheating. If OpenAI used the technology, 30% of respondents even stated they would probably move to a competitor LLM. Another problem the firm is experiencing with the widespread use of the ChatGPT detector is that other users could be able to reverse engineer OpenAI's watermarking method and create an app or plug-in that would neutralize it.
Nevertheless, OpenAI has been investigating alternatives to the text watermarking method since it recognizes the risk that artificial intelligence (AI)-generated content poses to society. Additionally, there is a need for an AI detector; according to internal statistics, 80% of respondents worldwide are in favor of its development.
No comments