MIT Rolls out GLTR, a tool to Detect Automatically Generated Text
MIT has teamed with IBM's Watson AI lab to develop Giant Language Model Test Room (GLTR), a machine learning algorithm to fight AI-generated text like that generated by OpenAI's algorithm. The model enables a forensic analysis of how likely an automatic system generated a text.
Language models are growing in popularity, and they are capable of effortlessly processing natural human language. For instance, the Gmail app can try to predict the next word users want to type while composing an email. Now language models have improved dramatically leaving a space for manipulation. Malicious attackers could use text generators to spread propaganda. However, a new forensic machine learning algorithm is recognizing AI-generated text. Being called as GLTR- it accurately predicts automatically generated text. The platform uses language models as OpenAI's text generator, GPT-2 117m.
The model ranks words based according to their use; the most used will be ranked top. The top ten most common words are highlighted in green. Yellow highlights the words which fall into the top hundred, the top thousand words highlighted in red and the rest of the words highlighted in purple. Researchers will get a visual insight into how likely each word was to be used with his method.
In a test on the AI text generator developed by Open AI —carried out by the researchers —the first sentence was the prompt used by the model to generate the text, while OpenAI’s artificial intelligence generated the rest. By running the text through MIT's algorithm, users can see that most of the words used are those the AI would expect to see as highlighted in green, suggests human didn’t write it.
Additions are still needed for the GLTR machine learning algorithm as it has a limited scale. Researchers hope that despite this limitations the model can be vastly improved and used to detect AI- generated text.