With artificial intelligence chatbots all the rage on the internet, it isn’t surprising that a new concern has arisen: how does one accurately identify passages that have been written by robots and not humans?
Since the start of the year, OpenAI has teased a solution to the problem, claiming it was working on a tool to “fingerprint” and watermark texts generated by its algorithm to avoid plagiarism or cheating claims.
The organization willmake use of cryptographic signals to watermark letters, which will still read as natural text to a layperson but, in fact, hides a code that can be deciphered by those who know the key to unlock it.
Now, OpenAI is finally introducing the anti-plagiarism system, which essentially is a box that can analyze at least 1,000 characters at a time. It then tells you if the text ranges from “likely generated” by AI to “very unlikely,” with five possible outcomes based on its calculations.
However, the company itself has admitted the checker is “imperfect.” Policy Research Director Lama Ahmad tells CNN the feature should be taken “with a grain of salt,” considering that it has shown “it can be wrong and will be wrong at times.”
During a demonstration with the publication, the tool correctly identified the book Peter Pan as “unlikely” to be AI-generated, though the firm cautioned once again that the feature incorrectly labeled human-written text as AI-written at least 5% of the time.
While there’s no surefire way to know if an article or document was penned by an actual human or a robot, there are other checkers on the market that may help. Nonprofits Quill and CommonLit recently revealed an open-source system that’s completely free to use, with a claim of 80% to 90% accuracy against ChatGPT scripts. Similar to the OpenAI’s feature, it works by having you copy and paste a chunk of text into a box, and then quickly determining if the passage has indeed been generated by AI.