Technicians have discovered entries in Google Search Console for a website that strongly resemble ChatGPT prompts. Apparently, OpenAI is forwarding these to Google Search unprotected. The AI manufacturer did not confirm the problem directly but stated that a corresponding “error” has now been fixed.
The problem was discovered by tech consultants Jason Packer and Slobodan Manić, who last month informed Packer in a blog post about it. Packer was struck by the entries in Google Search Console (GSC). Here, website operators are shown Google search queries for which their own site appeared in the results. In Packer’s case, however, these entries did not consist of a few words as usual, but were small texts. And they were ones tha…
Technicians have discovered entries in Google Search Console for a website that strongly resemble ChatGPT prompts. Apparently, OpenAI is forwarding these to Google Search unprotected. The AI manufacturer did not confirm the problem directly but stated that a corresponding “error” has now been fixed.
The problem was discovered by tech consultants Jason Packer and Slobodan Manić, who last month informed Packer in a blog post about it. Packer was struck by the entries in Google Search Console (GSC). Here, website operators are shown Google search queries for which their own site appeared in the results. In Packer’s case, however, these entries did not consist of a few words as usual, but were small texts. And they were ones that strongly resembled LLM prompts. He counted a total of 200 such entries, some with very personal content. What was striking: all had a ChatGPT URL at the beginning.
OpenAI: “Temporary error”
Furthermore, after closer inspection, they discovered a bug in a ChatGPT prompt input field. This apparently caused the model to perform a web search for every prompt and add a ChatGPT URL at the beginning of the prompt for this purpose. Here, ChatGPT then likely also used Google, but not via an API or a private connection – otherwise the prompts would not appear in the GSC, the author explains. The prompts can appear in the GSC of any website that Google’s search algorithm deems relevant to the ChatGPT URL. Packer points out in his post that ChatGPT prompts processed in this way are definitely not private. They would end up with Google, in the GSCs of various website operators, and possibly also with an entity that handles Google searches for ChatGPT prompts.
To the news agency Ars Technica OpenAI did not want to confirm this, but stated that they are aware of the problem with GSC. OpenAI also spoke of having fixed an error that existed “temporarily” and “affected a petite number of search queries” and had impacted the forwarding of search queries. After OpenAI’s statement, Packer assumes that the error lay in the ChatGPT layer responsible for converting prompts into search queries. Instead of breaking down a prompt into a few relevant search terms as intended, the entire prompt was simply used for the search query.
Packer: Strong indications
Packer supports the thesis that these are indeed ChatGPT prompts partly based on earlier reports that OpenAI is scraping Google Search. He sees the content of the GSC entries themselves, and the fact that there is always a ChatGPT URL at the beginning, as further strong indications. However, he would rather not speak of one hundred percent certainty.
It is not the first time that ChatGPT prompts, believed to be private, have become public. In the summer, a misunderstood function made very private chats publicly searchable – including names, if in doubt.
(nen)
Don’t miss any news – follow us on Facebook, LinkedIn or Mastodon.
This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.