Thousands of user chats with Elon Musk’s AI chatbot Grok are now publicly available on Google.
More than 370,000 Grok chats have been indexed by search engines, exposing hundreds of sensitive prompts that include medical and psychological questions, business details, and at least one password.
The chats have been exposed due to a Grok’s “share” feature—which users might use to send a record of a conversation to another person, or even to their own email. The share feature creates a unique URL when for the conversation. Those links were automatically published on Grok’s website and left open to search engines, seemingly without users’ knowledge.
The disclosure of the chats was first reported by Forbes.
Some of the transcripts available on Google Search that were reviewed by Fortune contained chats that went against Grok’s terms of service. One chat showed Grok telling a user how to make a Class A drug, while another offered detailed instructions on how to assassinate Elon Musk. xAI’s terms of service prohibit using Grok for “critically harming human life.”
It’s not the first time users have found conversations with AI chatbots that they thought were private ending up online.
OpenAI briefly experimented with a similar feature that allowed users to share their ChatGPT conversations via a link, which also made those conversations discoverable by search engines like Google. Despite the feature being opt-in and containing a disclaimer stating that chats could end up in search results, over 4,500 shared conversations were indexed by Google, including some that contained highly personal or sensitive information.
OpenAI pulled the feature entirely shortly after the publicly-indexed chats gained media attention. The company called it a short-lived experiment and acknowledged that it “introduced too many opportunities for folks to accidentally share things they didn’t intend to.”
At the time, xAI CEO Musk used the incident to promote Grok as an alternative to ChatGPT, taking to X to post: “Grok FTW.” Unlike OpenAI, Grok’s “Share” function does not include a disclaimer that chats could be shared publicly.
Meta’s AI app also has a similar share feature that publishes chats directly to the app’s Discover feed, which also led them to be indexed by Google. Many users did not realize that their chat were being shred to this feed and the chats often contained highly-personal and sensitive information.
Meta still allows shared chatbot conversations to be indexed by search engines, according to Business Insider. Google also previously permitted chats with its AI chatbot, Bard, to appear in search results, but removed them in 2023.
Privacy experts have warned users that chats with AI bots might not be as private as they appear to be. Oxford Internet Institute’s Luc Rocher told the BBC that AI chatbots are “a privacy disaster in progress.”
Once conversations are online, they are hard to remove completely. Many casual AI users may not fully understand how their data is stored, shared, or used. For example, two users who had their Grok chats indexed by Google were unaware that they had been public when they were identified and contacted by Forbes.
In jurisdictions like the EU, mishandling personal information may violate provisions of the bloc’s strict data privacy law, GDPR, which includes principles like data minimization, informed consent, and the right to be forgotten.
Users increasingly treat chatbots as confidants, feeding them sensitive details like health information, financial details, or relationship issues they likely would not want to be public. Even if chats are anonymized, prompts often contain identifying details that could be traced back to users or mined by malicious actors, data brokers, or hackers for targeted campaigns.
Some people have already identified a business case for Grok’s public chats. According to Forbes, some marketing professionals are discussing ways to exploit shared Grok conversations to boost business visibility. As the Grok conversations are published as web pages with individual URLs, businesses could potentially script chats that mention their products alongside keywords in order to game search results or create backlinks. It’s unclear how successful this would be, however. The tactic might pass on some SEO value and help to manipulate Google’s rankings, but it could also be perceived as spam by Google and ultimately hurt visibility.
Representatives for xAI did not immediately respond to a request for comment from Fortune.
This story was originally featured on Fortune.com