In late July 2025, people started finding their private-feeling ChatGPT conversations in ordinary Google searches. Not through a hack — through a feature. It is the clearest warning yet about what really happens to the words you type into someone else's AI, and why a growing number of teams now run their own.
The week people found their AI chats on Google
A simple site:chatgpt.com/share search on Google surfaced other people's conversations — nearly 4,500 of them, according to the Fast Company report that broke the story. Among them were exchanges about mental health, relationships, addiction and confidential work problems that the writers clearly never expected a stranger to read.
The mechanism was mundane, which is exactly the lesson. Clicking Share created a public link, and a checkbox on that dialog — “Make this chat discoverable” — let search engines index the page. It was opt-in and off by default, but it sat one careless click away, with nothing warning that “discoverable” meant “anyone on Earth can Google this.”
OpenAI reacted fast. Its Chief Information Security Officer, Dane Stuckey, announced on August 1, 2025 that the company had ended what it called a “short-lived experiment” that “introduced too many opportunities for folks to accidentally share things they didn't intend to.” Search engines don't instantly forget, so the cleanup dragged on for weeks.
Here is the part that should stop you: this was not a one-off at one company.
xAI's Grok had the same class of problem days later — its share button produced public, indexable pages, and press reporting put the number of Grok chats swept into Google in the hundreds of thousands (widely cited as ~370,000, though that is a search-result estimate, not an audited count).
Google's own Bard did this in 2023: shared conversations were indexed by Google Search until Google added a no-index rule to stop it.
Meta's AI app shipped a public “Discover” feed in 2025 that surfaced people's chatbot interactions — some strikingly personal — with many users apparently unaware their prompts were being published at all.
Four companies, one root cause: “Share” quietly meant “publish,” and “public link” quietly meant “search-indexable.” When the biggest names in AI make the identical mistake within two years, it stops being an accident and starts being a property of the model: the moment your data leaves your control, someone else's product decisions govern who sees it.
The leak you can see is the small part
Public search results are the exposure you can see. For a business, the exposure that matters most is the part you can't.
By default, your prompts can become training data
OpenAI's own Help Center states it plainly: “ChatGPT… improves by further training on the conversations people have with it, unless you opt out.” For consumer accounts, training on your inputs is the default — you have to go and switch it off. Business tiers (ChatGPT Business, Enterprise and the API) are not trained on by default — which is precisely the point: the free and personal tools most employees actually reach for are the leaky ones.
“Delete” doesn't always mean deleted
In the New York Times copyright litigation, a U.S. federal magistrate judge issued a preservation order on May 13, 2025 requiring OpenAI to “preserve and segregate all output log data that would otherwise be deleted” — overriding its normal 30-day deletion and sweeping in chats users had explicitly deleted, and even “Temporary Chats.” OpenAI fought it and, by late September 2025, was no longer required to retain conversations originating from the EEA, Switzerland or the UK. Read that sequence again: for months, a U.S. court order dictated the retention of chat logs, and it took a legal battle to carve European data back out. That is the CLOUD Act problem in one story — data held by U.S. providers sits under U.S. legal reach, wherever you are.
Employees paste things they absolutely shouldn't
Security firm Cyberhaven, measuring real usage across roughly 1.6 million workers, found that 8.6% of employees had pasted company data into ChatGPT, and 4.7% had pasted confidential data — and that was mid-2023, before adoption exploded. It is how Samsung leaked internal source code into ChatGPT in 2023 and then banned the tools outright. Every one of those pastes is a copy of your intellectual property on someone else's servers, under someone else's retention policy and someone else's subpoenas.
Why this is specifically a GDPR problem
If you handle data about people in the EU or UK, the incidents above aren't just embarrassing — they map onto obligations you can be fined for:
Art. 5 — data minimisation & purpose limitation. Pasting a customer's personal data into a consumer chatbot that may train on it is hard to square with “collected for specified, explicit purposes.”
Art. 28 — processors. Using an external AI as a processor needs a proper data-processing agreement and control over sub-processors. A free consumer chatbot gives you neither.
Art. 32 — security of processing. A “share = publish” design that lands chats in Google is close to the opposite of “appropriate technical measures.”
Art. 44+ — international transfers. Sending EU personal data to U.S. AI providers reopens the whole Schrems II / Data Privacy Framework question — and, as the retention order showed, U.S. legal process can reach that data.
The uncomfortable summary: the fastest, most convenient AI tools are also the ones that make EU compliance hardest.
The fix isn't “don't use AI.” It's “own the AI.”
You don't have to choose between modern AI and data protection. You choose where the model runs. If it runs on infrastructure you control, your prompts, documents and customer data never leave your building or your contract. There is no share-link to leak, no default training on your inputs, no foreign retention order reaching your logs — because the data was never handed to a third party in the first place.
This used to be theoretical. In 2025 it became ordinary — and, tellingly, the AI labs themselves now endorse it. When OpenAI released its first open-weight models, gpt-oss-120b and gpt-oss-20b, on August 5, 2025 (Apache-2.0 licensed, freely downloadable), it explicitly pitched “hosting these models on-premises for data security.” When the company that started the public-chatbot era tells businesses to run open models on their own hardware for privacy, the argument is essentially settled.
Two ways to run your own
Local / on-device. Tools like Ollama, LM Studio and vLLM run open-weight models — Meta's Llama, Mistral, Qwen, DeepSeek or OpenAI's new gpt-oss — on your own machine. A capable workstation GPU (16–24 GB of memory) comfortably runs models in the 7B–20B range; gpt-oss-20b is designed to run on edge devices with as little as 16 GB. Ideal for individuals and small teams, and the data never touches the internet.
Dedicated / hosted AI server. For real throughput — multiple users, larger models, retrieval over your own documents, always-on availability — you want data-center GPUs. Here is the number worth remembering: gpt-oss-120b, a 117-billion-parameter model, runs on a single 80 GB GPU such as an NVIDIA H100. One modern accelerator now hosts a frontier-class open model for your whole team. Put that server in an EU data center or your own rack and you get serious capability with full data residency and a clean GDPR story.
The honest trade-offs
Self-hosting isn't magic, and we won't pretend it is:
Factor
| Public AI (ChatGPT, Grok…)
| Your own LLM
|
|---|
Where your data goes
| Third-party US servers
| Your hardware / EU data center
|
Trained on your inputs?
| Consumer tiers: yes, unless you opt out
| Never
|
Exposed by a “share” bug?
| Yes — proven in 2025
| No share-link exists
|
GDPR / data residency
| Transfers, retention orders
| You control it end to end
|
Peak capability
| Highest (frontier models)
| Strong, a step behind
|
Cost at scale
| Per-token fees, forever
| Fixed hardware, then near-zero
|
The capability gap is real but shrinking every quarter, and for the vast majority of business tasks — drafting, summarising, classification, retrieval over your documents, code assistance, support — today's open models are already more than enough. What you gain is decisive: your data is never the product, never the training set, and never one careless checkbox away from Google.
Where MillionMiner fits
Running your own AI is a hardware problem before it's a software problem — and hardware is what we do. We supply and host the exact
data-center GPUs (NVIDIA H100, H200 and more) that turn “we should really run our own model” into a rack that's online next week — delivered to you, or hosted in our facilities with the same transparent, verifiable operations we run for mining. One H100 is enough to serve a 120-billion-parameter model to your whole team, privately.
If data protection is on your roadmap for 2026, this is the move: keep the AI, drop the exposure.
Talk to us about a private AI server
Tell us your model, your users and your timeline — an engineer replies with a concrete, GDPR-friendly hardware plan, usually within a couple of hours.
Get a hardware recommendationFree, no obligation, no signup.
Frequently asked questions
Did ChatGPT really leak private conversations to Google?
Not through a hack. Conversations users chose to share via a public link, with the “Make this chat discoverable” option enabled, were indexed by search engines. Fast Company found roughly 4,500 such chats on Google, and OpenAI disabled the feature on August 1, 2025, calling it a short-lived experiment. Grok, and earlier Bard, were exposed the same way.
Does OpenAI train on my ChatGPT conversations?
For consumer ChatGPT, yes by default — its Help Center says the model “improves by further training on the conversations people have with it, unless you opt out.” Business, Enterprise and API usage are not trained on by default. You can disable training in your data controls.
Can a US AI provider be forced to keep my deleted chats?
It happened. A May 2025 court order in the New York Times v. OpenAI case forced OpenAI to preserve output logs that would normally be deleted. OpenAI later won relief for EEA, UK and Swiss data, but the episode shows US legal process can reach data held by US providers.
Is running your own LLM realistic in 2026?
Yes. Open-weight models (Llama, Mistral, Qwen, DeepSeek, OpenAI's gpt-oss) are freely available. Small models run on a good workstation GPU; large ones like gpt-oss-120b run on a single 80 GB data-center GPU such as an H100. Tools like Ollama and vLLM make deployment straightforward.
Is a self-hosted model as good as GPT-5 or Gemini?
At the absolute frontier, the big closed models still lead. But for everyday business work — writing, summarising, extraction, retrieval over your own documents, coding help — current open models are more than capable, and they run entirely under your control.
What hardware do I need to run my own AI?
Personal use: a 16–24 GB GPU. Team or production use: one or more data-center GPUs (H100 or H200) in a server, on-premises or hosted in an EU facility for data residency. MillionMiner can spec, supply and host it.
Sources include Fast Company, The Register, OpenAI's Help Center and policy pages, OpenAI's August 1 and August 5, 2025 statements, court filings in NYT v. OpenAI (preservation order, May 13, 2025) and Cyberhaven workforce data. The ~370,000 Grok figure is a search-index estimate from press reporting, not an audited count. This article is general information, not legal advice — consult your data-protection officer for your specific obligations.