According to Techcrunch, this approach aims to create high-quality training data for AI, but also raises concerns about legal and security risks.
According to Wired's report, OpenAI and the training data company Handshake AI are asking contractors to upload actual projects they have done in the past and present. These data are used to train and evaluate new artificial intelligence models, aiming to automate many office jobs.
Internal documents show that OpenAI requires contractors to specifically describe the tasks they have undertaken in previous jobs, and provide examples of "the actual work they actually did". The files requested to be uploaded are not summaries but complete products, which may include Word, PDF, PowerPoint, Excel, image or data warehouse documents.
OpenAI guides contractors to delete proprietary information and personal identification data before sending, and suggests using the "Superstar Scrubbing" tool integrated in ChatGPT to clean data.
Intellectual property lawyer Evan Brown said that letting contractors decide for themselves what confidential information contains poses a "very high risk" for AI laboratories.
According to Techcrunch, before these concerns, OpenAI's spokesperson declined to comment.
