OpenAI Inc. told a federal judge that news organizations are seeking sanctions against it for alleged discovery misconduct in a sweeping copyright lawsuit because publishers have \u201cno evidence whatsoever\u201d that ChatGPT systematically regurgitates their articles.<\/p>\n
OpenAI said<\/a> Friday that it has followed court orders to preserve users\u2019 chatbot conversations \u201cto the letter\u201d and \u201cgone to herculean efforts\u201d to turn over all of its training data and tens of millions of ChatGPT outputs.<\/p>\n Accusations by the New York Times Co. and other publishers that the AI giant hid its ability to search training data and output logs for their articles for two years is an attempt \u201cto rewrite history\u201d and avoid proving their case, OpenAI told the US District Court for the Southern District of New York.<\/p>\n \u201cPlaintiffs\u2019 real gripe appears to be that these materials show their claims premised on regurgitation are baseless,\u201d OpenAI\u2019s opposition brief said.<\/p>\n The Times and other publishers including The Daily News moved for sanctions<\/a> last month, arguing OpenAI\u2019s willful deception about its technical capabilities has prevented them from establishing the foundation of their case \u2014 the extent to which OpenAI scrapes, trains, and copies its content to respond to user prompts.<\/p>\n OpenAI deleted user chats against a court order, and rendered sample consumer logs unusable due to redactions, according to the motion. The publishers seek to prohibit OpenAI from relying on the produced output logs \u201cfor any purpose,\u201d a finding that the logs include or would have shown substantial regurgitation of their articles, and attorneys\u2019 fees associated with the alleged obstruction.<\/p>\n OpenAI argued that the court\u2019s preservation order didn\u2019t require \u201cwholesale\u201d preservation or for OpenAI to stop all deletions, so users could still delete chats that were more than 30 days old as of the date of the court\u2019s directive.<\/p>\n That resulted in about 5% of the logs the publishers requested being deleted, but OpenAI says they were replaced with other logs that didn\u2019t bias the sample. The sample is \u201ccritically\u201d important to examine the publishers\u2019 claims, and precluding it would \u201csignificantly hamper OpenAI\u2019s ability to mount a defense.\u201d<\/p>\n And it\u2019s \u201ca technical reality\u201d that OpenAI\u2019s existing tools can\u2019t execute searches across all training datasets and output logs for publishers\u2019 articles, OpenAI said. While an internal OpenAI project searched for Times\u2019 content across a \u201crough-and-ready\u201d sample of 78 million conversations, that search \u201cdiffers fundamentally\u201d from the publishers\u2019 request.<\/p>\n OpenAI urged the court against handing the publishers an \u201cunwarranted windfall.\u201d<\/p>\n Though plaintiffs didn\u2019t seek sanctions against co-defendant Microsoft Corp., the tech company also opposed<\/a> the publishers\u2019 bid Thursday, arguing the requested restrictions against OpenAI would unduly prejudice Microsoft.<\/p>\n Susman Godfrey LLP, Rothwell Figg Ernst & Manbeck PC, Loevy & Loevy, Klaris Law PLLC, and Ruttenberg IP Law represent the news plaintiffs.<\/p>\n OpenAI is represented by Williams & Connolly LLP, Keker Van Nest & Peters LLP, Latham & Watkins LLP, and Morrison & Foerster LLP.<\/p>\n Microsoft is represented by Faegre Drinker Biddle & Reath LLP and Orrick, Herrington & Sutcliffe LLP.<\/p>\n