OpenAI releases a tool to see through ``text written by AI''
On January 31, 2023, OpenAI announced `` AI Text Classifier '' that distinguishes sentences generated by AI.
AI Text Classifier - OpenAI API
New AI classifier for indicating AI-written text
OpenAI releases tool to detect AI-written text
When you access the “ AI Text Classifier ” released by OpenAI, you will see the following.
Since a sample is prepared, I will try it immediately. First, when you select 'Human-Written' on the left and click 'Submit', 'Classifier judges that the possibility that AI generated this text is extremely low' was displayed. This sample text is the beginning of 'Peter Pan' written by James Matthew Barry, a fairy tale writer. In other words, it is correct because it is a sentence written by a human being.
Similarly, when I tried to identify a sample that was one of the documents output by ChatGPT, I was able to determine that it was created by AI.
About GIGAZINE ', it was judged that 'the possibility that AI generated it is low'.
It also supports Japanese. As a test, when I entered the sentence '
Next, we let ChatGPT generate sentences to test with AI-generated sentences.
In order to satisfy the condition of 1000 characters or more, I sent the output in two parts and connected them. Then, it was judged that 'AI is likely to have generated'.
According to OpenAI, an evaluation by a test conducted in English correctly identified 26% of the text written by AI as ``highly likely written by AI,'' while incorrectly identifying text written by humans. There were 9% of cases where it was discriminated as 'written by AI'. In addition, it has been confirmed that it is possible to avoid detection by modifying the sentences output by AI, and there is also a limitation that accuracy drops significantly when used in languages other than English.
Therefore, OpenAI is particularly useful for educators as a single piece of evidence in determining whether students have used AI for academic misconduct, such as cheating with AI or having them write papers on their behalf. It warns that Classifier should not be used.
BleepingComputer, an IT news site that covered AI Text Classifier, said, ``The detection rate of this tool will likely improve over time as it is trained with more data. It's not a reliable tool for discovering content you create.'