Nonprofit research company, OpenAI, has created a machine learning algorithm that can generate sensible strings of text using human-created writing prompts. Known as Generative Pre-trained Transformer 2 (GPT-2), the system can mimic human writing so well that its developers have decided to scale back its capabilities and hold it from the public.
At a time when concerns are rising over “deepfake” content generation, the research team wants to be cognizant of how GPT-2 could be used in unsafe ways if in the wrong hands. “Due to our concerns about malicious applications of the technology, we are not releasing the trained model,” states a blog post on OpenAI’s site. “As an experiment in responsible disclosure, we are instead releasing a much smaller model for researchers to experiment with, as well as a technical paper.”
The Ultimate Fake News Writer
GPT-2 was built using 40 gigabytes of text sourced from all over the Internet. The system is much more sophisticated than other text generation AI bots that have been released, as it produces higher-quality “writing” all the way through a story. Whereas other models lose focus on a given subject matter, GPT-2 stays on track.
There are several published examples of GPT-2’s outputs, including a story about a unicorn discovery in the Andes, a story about a Miley Cyrus shoplifting attempt, and a report about stolen nuclear materials near Cincinnati. GPT-2 created each of these from simple two-sentence, human-generated prompts. The algorithm’s articles matched the styles of the prompts given and even included quotes from other sources.
Generally, it takes GPT-2 several tries to produce a decent sample of writing. The system’s output is largely dependent on how much it already “knows” about a given topic. GPT-2 creators state that the algorithm can generate reasonable writing half of the time for topics that are well-represented in the training data.
OpenAI Stepping In
Founded by Elon Musk and Sam Altman in 2015, OpenAI is a nonprofit research company focused on “discovering and enacting the path to safe artificial general intelligence.” The organization is supported by contributions from the “PayPal mafia,” a group of technology investors and executives which includes Peter Thiel, Jessica Livingston, Reid Hoffman, Greg Brockman, Elon Musk, and Sam Altman.
OpenAI’s goal is to positively shape the future of AI and guide developers away from potentially harmful applications of the technology. GPT-2’s capabilities are concerning, especially at a time when deepfake content is being used already to generate money, defame individuals, and spread misinformation.
GPT-2’s researchers have stated that the model could be fine-tuned for specific tasks, such as writing reviews on Amazon products, or be used for broader, more damaging reasons. As a result, OpenAI has chosen not to release its dataset, training code, or model weights. Instead, the organization will only publish a lighter, less powerful version of the algorithm.