Remember the story about Samsung employees “leaking” sensitive data through ChatGPT, which some outside third party then found? It's not quite true.

I have heard this story misstated many times at legal conferences and in webinars. Even prominent law firms have published accounts that go beyond the evidence. An Allen & Overy article, still available on A&O Shearman’s website, claims that Samsung’s confidential code was subsequently reproduced for users at other organizations.

The distinction matters. Entering confidential information into an outside provider’s service and having that information reproduced for an unrelated user are two different events. Evidence of the first does not establish the second.

On March 30, 2023, Economist Korea reported that Samsung employees had entered confidential source code and meeting content into ChatGPT. The article described three incidents: an employee submitted source code to obtain help correcting an error; another submitted code for optimization; and a third submitted a transcript of a recorded meeting to generate minutes.

What actually happened, according to that reporting, was that Samsung employees inserted confidential information into ChatGPT. By virtue of that happening, the information reached a third party: OpenAI, the company operating the service.

That was a legitimate confidentiality concern. Information had been transmitted outside Samsung’s systems. It did not have to reach a competitor or another chatbot user for the company to take the disclosure seriously.

But nowhere in the original Korean report is it stated that an unrelated ChatGPT user subsequently discovered Samsung’s private information. The article described a warning about possible further disclosure if the information were incorporated into the model’s learning. It did not identify a person who recovered the code or provide a generated response containing it.

Subsequent coverage widely cited the Korean article. Bloomberg separately reported reviewing a Samsung internal memo, which described concern that information submitted to AI platforms could be disclosed to other users. That, too, described a risk, rather than establishing that such disclosure had occurred.

Yet the story has been retold as a confirmed instance of ChatGPT reproducing one user’s confidential information for someone else.

The article on A&O Shearman’s website states that Samsung’s code was: “subsequently reproduced in response to users from other organizations.” That is unsubstantiated.

I have found no verified public evidence establishing that ChatGPT reproduced Samsung’s submitted code for an unrelated user. More broadly, I know of zero verified public instances establishing the specific sequence often attributed to Samsung: a company or its employee submits confidential information to a chatbot, that submission is used to train the model, and the model later reproduces it for an unrelated user.

This is not to say it has never happened, nor that it could never happen.

Researchers have demonstrated that models can reproduce memorized training data. OpenAI also uses eligible user content for model training, depending on the service, settings, and applicable terms. There have also been documented security failures involving exposure of information across users, but those incidents involve different security vulnerabilities and should be assessed on their own evidence.

Organizations should take confidentiality seriously when deciding which information employees may enter into AI tools.

But a possible consequence is not the same as a documented event. The Samsung incident should be taught for what the reporting establishes: employees submitted confidential company information to an outside AI service. It should not be presented as proof that ChatGPT then gave that information to another user.