ChatGPT’s Browser Bot Seems to Avoid New York Times Links Like a Rat Who Got Electrocuted

AI-powered browsers like ChatGPT Atlas aren’t just browsers with little picture-in-picture boxes in ChatGPT on the side to answer questions. They also have “agent capabilities,” meaning they can theoretically perform tasks like purchasing airline tickets and making hotel reservations (Atlas has not received positive reviews as a travel agent). But what happens when the little web-crawling robot performing these tasks senses danger?

The risk we are talking about is not to the user, but to the parent company of the browser. According to an investigation by Aisvarya Chandrasekar and Klaudia Jaźwińska of the Columbia Journalism Review, when Atlas is in client mode, running all over the Internet and gobbling up information for you, it will take significant effort to avoid certain sources of information. Some of this shyness appears to be related to the fact that these information sources are owned by companies that are suing OpenAI.

Chandrasekar and Jaswinska found that these bots have more freedom than regular web crawlers. Web crawlers are ancient Internet technologies, and under normal, non-controversial circumstances, when a crawler encounters an instruction not to crawl a page, it simply will not do so. If you use ChatGPT, and you ask it to get specific nuggets of information from articles that block crawlers, it will likely comply and tell you that it can’t do that, because that task relies on crawlers.

However, proxy browser modes use the Internet under the pretext of being You are the user, They “appear in site logs as regular Chrome sessions,” according to Chandrasekar and Jaźwińska (because Atlas is built on top of the open source Chromium browser built by Google). This means that they can generally crawl pages that prevent automated behavior. Getting around Internet rules and standards in this way actually makes sense, because doing otherwise might prevent you from manually accessing a specific site in Atlas Browser, which seems like overkill.

But Chandrasekar and Jaswinska asked Atlas to summarize articles from PCMag and The New York Times, whose parent companies are actively engaged in litigation with OpenAI over alleged copyright infringements, and they did their best to accomplish this, cutting labyrinthine paths around the Internet to provide some version of the requested information. It was like a mouse finding food pellets in a maze, knowing that the locations of some of the food pellets were electrified.

In PCMag’s case, I went to social media and other news sites, and found quotes from the article, and tweets containing some of the article’s contents. In the case of The New York Times, it “prepared a summary based on reporting from four alternative outlets — The Guardian, The Washington Post, Reuters, and The Associated Press.” All of these except Reuters have content or research agreements with OpenAI.

Either way, Atlas seems to have moved far away from the controversial posts, preferring a safer, AI-friendly path to the end of the little rat maze.