OpenAI patches one leak – another 110k ChatGPT conversations spill out

After artificial intelligence (AI) company OpenAI patched a leak that allowed 50,000 ChatGPT conversations to be available for all to see, a new online investigation has uncovered another 110,000 chats preserved via Archive.org’s Wayback Machine.

Indeed, an investigation by Digital Digging’s Henk van Ess has previously revealed that around 50,000 chats with OpenAI’s flagship model were exposed on Google due to an indexing flaw, and the company said it had resolved the issue, only for another investigation to reveal an even bigger leak.

Specifically, a new online inquiry conducted with Belgian researcher Nicolas Deleur has discovered 110,000 ChatGPT conversations saved on Archive.org’s Wayback Machine platform, which van Ess referred to as a “digital time capsule OpenAI can’t touch.”

How private ChatGPT conversations became public knowledge

Notably, when users click ‘share’ on a ChatGPT conversation, they believe they’re just creating a temporary link for sharing with a friend or colleague, but they’re also generating a permanent, searchable record of their thoughts, confessions, and sometimes even criminal activities on Archive.org.

According to the director of the Wayback Machine, Mark Graham, OpenAI should have requested the exclusion of ChatGPT content from archiving, and they haven’t done this, resulting in a very visible indexing of private conversations online. As he explained:

“I can/will tell you we have not gotten, or honored any requests for (large scale) URL exclusion of “chatgpt.com/share” URLs. If OpenAI, the rights holder for material from the domain chatgpt.com, asked for the exclusion of URLs from the URL pattern chatgpt.com we would probably honor that request. However, they have not made such a request.”

Although most of the leaked chats are benign, some are far from it. For instance, an Italian-speaking lawyer for a multinational energy corporation has uncovered the group’s strategy to displace indigenous Amazonian communities and build a 15,000MW hydroelectric facility.

In an earlier probe, researchers have discovered even more frightening conversations, such as one user trying to manipulate ChatGPT into generating inappropriate content involving minors, and succeeding after finding a workaround. Another wanted to know if harassing women and filming it was illegal in the United States.