Hi - Pranav from the ChatGPT Atlas team here. Appreciate the discussion and feedback here. Just wanted to correct a few specifics so there’s no confusion:
- We have tried to make Atlas a great way to browse the open web, and invested heavily in the search experience as we covered in the livestream. When I test typing “taylor swift" into Atlas I see links to the website in the autocomplete carousel, at the top of the results page, in inline citations within chat, and in the Search tab. We’re also working on making these faster and more prominent when we’re sure you want a link, for ex. you’ll see a large vertical list of links at the top of the chat response when you type in something like “gmail”.
- Browser memories are opt-in, and by default we don’t train on the contents of pages you browse—even when memories are enabled. Keeping the Ask ChatGPT sidebar open has no bearing on what gets sent to us. Webpages are clearly displayed as attachments you can disable in the chat, and only sent to us when you submit the prompts.
- Ask ChatGPT and Agent mode are not workarounds to user or publisher training settings. Even if a user has training enabled, we will still respect publisher preferences to block GPTBot in their robots.txt and do not train on that content.
Haha, wait, the training directly on websites is off by default, but if you use browser memories, training on that is on by default. Browser memories contain arbitrary amounts of information from websites...
Anil's point is that Chatgpt can't, on its own, train on data that requires a login. So Atlas is in fact automating training on login-walled content, and doing so by default, but also adding decoy settings that make you think it's not.
- We have tried to make Atlas a great way to browse the open web, and invested heavily in the search experience as we covered in the livestream. When I test typing “taylor swift" into Atlas I see links to the website in the autocomplete carousel, at the top of the results page, in inline citations within chat, and in the Search tab. We’re also working on making these faster and more prominent when we’re sure you want a link, for ex. you’ll see a large vertical list of links at the top of the chat response when you type in something like “gmail”. - Browser memories are opt-in, and by default we don’t train on the contents of pages you browse—even when memories are enabled. Keeping the Ask ChatGPT sidebar open has no bearing on what gets sent to us. Webpages are clearly displayed as attachments you can disable in the chat, and only sent to us when you submit the prompts. - Ask ChatGPT and Agent mode are not workarounds to user or publisher training settings. Even if a user has training enabled, we will still respect publisher preferences to block GPTBot in their robots.txt and do not train on that content.