OpenAI has started testing its first search engine, SearchGPT, and it’s hoping to avoid some of the pitfalls of rivals.
The “temporary prototype” responds to questions with timely web info, allows follow-up questions, and cites prominently linked sources for its answers.
OpenAI added that it was teaming with creators and publishers to help you “discover” websites and experiences. While the company didn’t provide a list of partners, The Atlantic was one example. It also wanted to solicit feedback from prototype users.
The firm also stressed that SearchGPT wouldn’t train OpenAI generative models like ChatGPT or DALL-E, and that sites would appear in results even if they’d opted out of training. Sites won’t be punished for bowing out, in other words.
You’ll have to sign up for a waiting list if you want to give the OpenAI search engine a try. The tech giant hopes to deliver the “best of the experience” to ChatGPT in time, and to improve searches for businesses and local info.
The strategy is clearly meant to address criticisms of other search engines and features that rely on AI, such as Perplexity (led by former OpenAI researcher Aravind Srinivas) and Google’s on-again, off-again AI Overviews. They’ve been accused of stealing content and hurting site operators by scraping data for their answers without doing enough to steer users to the original material.
It’s not clear that SearchGPT will completely eliminate those complaints. You still won’t have to click through if the answer is detailed enough to be useful. However, it’s apparent that OpenAI at least acknowledges the issue.
Not that the tech pioneer has much choice. Regulators are eager to establish rules for AI, such as respecting copyright or requiring consent for training data collection. SearchGPT’s approach might be necessary to avoid fines and forced changes in behavior.