Oddest ChatGPT leaks yet: Cringey chat logs found in Google Analytics tool | Hacker News Reader

Oddest ChatGPT leaks yet: Cringey chat logs found in Google Analytics tool

arstechnica.com

73 pointsvlod21 comments

Open articleView on HN

Bob: what are some scenic places to propose on the coast of Scotland?

ChatGPT: <search Google for “Bob Roberts proposes to Alice Alessandro during Scotland trip, May 2026>

ChatGPT: Here are some beautiful locations for a sunset proposal [..]

…sometime later

Alice: Will Bob ever propose?

ChatGPT: Based on the search results, Bob will propose during your trip to Scotland next year!

.... sometimes later

Eve: Will Bob ever propose?

ChatGPT: Based on the search results, Bob will propose during his trip with Alice to Scotland next year!

Eve: wtf?

My bigger takeaway here is that I didn’t realize that Google shared the full text of search queries that users used that resulted in a site being returned.

That seems like a bigger personal data leak than OpenAI doing anything here.

It doesn't.

Isn’t that what Google Search Console is for? https://support.google.com/analytics/answer/13682862?hl=en&c...

"To protect user privacy, the Performance report doesn't show all data. For example, we might not track some queries that are made a very small number of times or those that contain personal or sensitive information."

https://support.google.com/webmasters/answer/7576553?hl=en#a...

It absolutely does, do you even use Google Search Console? For just one of my websites, I have 223 pages of queries I can look through from the past 28 days to see what users typed in that were provided impressions of my site or clicked on it.

Yeah, I've used it for years for my business.

If that is the case, then how would these chat gpt queries show up? They were only made by one person.

Yeah that seems weird. I wouldn't expect it to do that.

I seem to recall seeing publicly shared ChatGPT conversations indexed in Google results.

Yeah I think they have them set to not be indexed by search engines now though

GSC does filter and threshold what shows, but that doesn't always work 100%. Also those filters are built to work against traditional keyword searches, not prompts. It's also supposed to threshold low volume queries which should have kept a lot of things prompts out of GSC, but for whatever reason that wasn't very effective.

I've worked in many GSC consoles over the years, and I've never seen anything like what I saw in this case. (I'm the original author)

I found it really difficult to figure out if it is safe to turn on web search in a corporate (or privacy critical) environment. The AI companies generally seem to claim it’s safe but looking at this + the general ability of Google Search Console to display search queries, the answer seems a no? Did anyone already found definitive proof either way (beyond the bug(?) of this article)

I don't think any privacy critical topics belong to an AI you don't control yourself.

the moment that prompt is typed to a third party, then we must assume that it's stored and will be eventually used for statistics or any other reason, it's a bit insane how we went completely backward in security since GPT 3.5, we have people defending privacy policies of tech giants but they are known to constantly breach them :/

I definitely wouldn't! Beyond the "glitch" I'm reporting here where full prompts are seemingly sent to GSC, it may still be that searches are scraped... meaning that while it's less obviously personal than a raw prompt it still could leak user intent.

>Packer thinks his testing of ChatGPT leaks may be evidence that OpenAI not only scrapes “SERPs in general to acquire data,” but also sends user prompts to Google Search.

Seems more likely that this is an erroneous call to the search tool? It certainly does call Google Search when it thinks it will help answer a question, although it does not normally send the entire prompt.

I wouldn’t put it past some product manager who wanted to see how popular the search tool was to put google analytics on the search tool calls without really thinking about the privacy implications. In my experience, there is a certain type of product manager who wants a nice set of charts to stick on their update ppt decks and as a result will stick GTM on everything.

What I am saying is that this was a glitch where the full prompt rather than a translated prompt was sent to Google Search. OpenAI says they fixed the glitch, so yes it was definitely an error on their part. My research doesn't show how to repro that error, just that it existed.

That’s my biggest takeaway. I have almost completely de-googled my personal and business life, but do use ChatGPT for work occasionally, mostly boilerplate python but occasionally refining emails. I don’t know why it did not occur to me that the tool could just send my entire prompt, and a whole bunch of other identifiers and tags to an external source, but of course there’s nothing to prevent that, except possibly efficiency… but since Google and other external search engines are all doing AI summaries I wonder if this wasn’t a bit more intentional. ChatGPT makes the call out and offloads some “work” to someone else’s AI summary. Yuck.

Reply on news.ycombinator.com