OpenAI released the GPT-2 algorithm in February of 2019. At the time, only a small parameter model was released due to concerns about the abuse of the model.

Due to our concerns about malicious applications of the technology, we are not releasing the trained model…we are instead releasing a much smaller model for researchers to experiment with, as well as a technical paper.

Over the year, OpenAI has worked with academia and industry to investigate the potential for abuse of the model. Ultimately, they decided to release larger and larger versions of GPT-2. The main takehomes of their research indicate that:

1. Coordination is difficult, but possible.
2. Humans can be convinced by synthetic text.
3. Detection isn’t simple.

That last piece, “detection” precipitated the creation of a browser extension to do just that! @karpathy requested the extension at the beginning of November, @madavidj offered a job to anyone that managed it and several people delivered. The extension ‘GPTrue-or-False‘ was created by Giulio Starace (@thesofakillers) for Chrome and Firefox. The extension “displays the likelihood that selected portions of text were generated by GPT-2”. Take a look at the example above run on a sample of AI-generated text from Robin Sloan’s MageFrame project.

If you would like to learn more about AI-generated text detection check out the blog post or GitHub repo by OpenAI.

Written by Rebecca Minich, Product Analyst in Data Science at Google. Opinions expressed are solely my own and do not express the views or opinions of my employer.