Atlas

The master key to the internet

I was interested to try out ChatGPT Atlas, like many people. I genuinely wasn’t sure of the point of the product, which actually made it more interesting to me, as usually the point of a new product is quite clear.

Upon downloading, installing and opening the macOS app, I was underwhelmed. It’s just a fairly basic browser experience. I don’t see why it would have such a draw that people would switch over, learn new keyboard shortcuts, and move bookmarks and passwords¹. Sure, it’s nice to be able to open the panel and use ChatGPT in context, but it’s not that much harder than copying and pasting text or links into the ChatGPT tab I always have open.

Of course, ChatGPT may become more helpful as it observes your browsing habits, but this also has security implications. Perhaps it could discover new and interesting content based on what you’ve been reading and asking it to summarise or interact with. Perhaps it could even identify more interesting material when it notices that you’ve read it yourself rather than asking for a summary. It might eventually offer entirely synthesised content that exists only for you—tailored to your interests but still anchored in the ‘truth’ of existing content².

It’s like a series of conversion steps of interest: clicked the link, looked at the content, scrolled a bit, asked ChatGPT about it or to summarise it, read it yourself, shared it. The perfect dataset to build a recommender system. Still, I didn’t think it would be that impactful as a launch. I don’t believe the current benefits balance with the drawbacks enough to persuade people to move from their favourite browser. After all, your favourite browser is like your favourite brand of underwear—are you really going to switch if you’re comfy?

Then the other, and perhaps main, purpose of Atlas dawned on me. I wanted ChatGPT to summarise a blog post I couldn’t be bothered to read. When I pasted the link into ChatGPT, it couldn’t open it—probably because of a robots.txt restriction or something else preventing it from consuming the content.

When I pasted the same link into an Atlas tab, it opened immediately because Atlas is a browser. Then, when I used the right-hand panel to ask ChatGPT to summarise it, it was able to do so because it already had the rendered page to look at.

Atlas is a way for OpenAI to gain access to new training material currently unavailable to it—effectively crowdsourcing access to the parts of the internet it isn’t otherwise allowed to use, or which are paywalled. If even one percent of users switch to Atlas, OpenAI will have access to 99 percent of the content it doesn’t have today. From a legal point of view, this doesn’t appear as OpenAI accessing unauthorised content; instead, it’s users feeding content to it, without the system knowing whether it’s allowed to see it or not. It’s equivalent to us copying and pasting the text of an article—whose link ChatGPT wouldn’t be allowed to open—into the chat. OpenAI has simply made this an implicit step: asking the ChatGPT panel anything about an open page effectively copies the page into it. It now holds the URL and its content as a key–value pair, even where it wouldn’t normally be allowed to open the URL directly.

Many have said that OpenAI is running out of high-quality training content for new models. It looks as though they’ve found a solution to that problem. OpenAI gains access to all its users’ paid content subscriptions and any link³ they view—robots.txt or not.

1. I don’t do this, but I know most people save their passwords in their browser!

2. It would need to be able to recognise its own content to avoid circular references.

3. Company Sharepoint or Confluence included.