Tweeted By @harvardnlp

New tool (https://t.co/ovMLD7jpiD) argues that automatically generated text, eg. gpt2, looks real because it is "too good". Generation constraints leave behind markers that are hard for humans to detect but can be exposed to readers with statistical tests. (with @MITIBMLab) pic.twitter.com/69BkwoV9tt

— harvardnlp (@harvardnlp) March 6, 2019