OpenAI has released a free online tool AI Text Classifier, designed to identify AI-generated texts.
We’re developing a new tool to help distinguish between AI-written and human-written text. We’re releasing an initial version to collect feedback and hope to share improved methods in the future. https://t.co/4dQE3dX6vX
— OpenAI (@OpenAI) January 31, 2023
The program is based on a language model and assesses the probability that a fragment of text was generated by AI. The five-point scale ranges from “very unlikely” to “likely”.
The developers acknowledged that the classifier is not sufficiently reliable.
“In our assessments on an English-language test set, the classifier correctly identified 26% of AI-written text (true positives) as “likely AI-written”. At the same time, human-written text was incorrectly flagged as AI-written in 9% of cases (false positives),” the blog said.
OpenAI says the tool may help detect machine-written text, but it should not be the sole basis for decisions.
“The model is trained on records from various sources, which may not reflect all forms of human-written text,” the organisations added.
AI Text Classifier is intended to detect AI-generated works from various sources. It was trained on 34 generators created by five organisations. It was also trained on texts written by humans.
To test the programme, a sample of 1,000 characters must be provided. It performs poorly for languages other than English.
The classifier will also be of little use to teachers seeking to assess the source of students’ work. The tool is not sensitive enough to changes made by humans to AI texts. Such works may evade detection.
“We warn that the model has not been thoroughly evaluated for many of the expected core goals, including student essays, disinformation campaigns or chat transcripts. It is known that neural-network based classifiers are poorly calibrated beyond their training data. For inputs that differ greatly from the text in our dataset, the tool can sometimes be very confident in incorrect predictions,” the company warned.
OpenAI is also exploring other approaches to detecting AI-generated text, including a watermarking method.
OpenAI unveiled ChatGPT in December 2022. Within 40 days the chatbot reached 10 million daily active users.
Since then the tool has repeatedly drawn concern from educators. Schools and universities in the United States, Australia, France and India have banned the use of ChatGPT by students.
Experts also fear that bad actors could use text-generation models to spread disinformation, phishing emails and vast volumes of meaningless text to flood the internet with spam.
In January, New York state schools banned the use of ChatGPT on campus. Later, the largest universities in Australia joined them.
In the same month OpenAI released a report on the possible use of text generators to create disinformation and propaganda.
At the end of January, for the first time in history on the floor of the U.S. Congress, a speech written by AI was delivered.
Subscribe to ForkLog’s Telegram updates: ForkLog AI — all the news from the AI world!