Concerns around abuse of AI text generation have been widely discussed. In the original GPT-2 blog post from OpenAI, the team wrote:
These concerns about mass generation of plausible-looking text are valid. However, there have been fewer conversations around the GPT-2 data sets themselves. Google searches such as "GPT-2 privacy" and "GPT-2 copyright" consist substantially of spurious results. Believing that these topics are poorly explored, and need further exploration, I relate some concerns here.
Inspired by this delightful post about TalkTalk's Untitled Goose Game, I used Adam Daniel King's Talk to Transformer web site to run queries against the GPT-2 774M data set. I was distracted from my mission of levity (pasting in snippets of notoriously awful Harry Potter fan fiction and like ephemera) when I ran into a link to a real Twitter post. It soon became obvious that the model contained more than just abstract data about the relationship of words to each other. Training data, rather, comes from a variety of sources, and with a sufficiently generic prompt, fragments consisting substantially of text from these sources can be extracted.
A few starting points I used to troll the dataset for reconstructions of the training material:
RAW PASTE DATA
About the Author
I soon realized that there was surprisingly specific data in here. After catching a specific timestamp in output, I queried the data for it, and was able to locate a conversation which I presume appeared in the training data. In the interest of privacy, I have anonymized the usernames and Twitter links in the below output, because GPT-2 did not.