SecurityNov 3, 2026, 1:46 AM UTC·2 min read

Shadow AI: Deleting a ChatGPT Chat Doesn't Erase Company Data

Gartner warns that deleting a ChatGPT conversation does not remove confidential data from servers, as shadow AI becomes a growing source of corporate data leaks.

An office worker typing on a laptop while a chat window with a delete button hovers over a glowing server rack. — Image: 미래를 보는 창 - 전자신문

SUMO

Dialo

ediviewer

Deleting a ChatGPT conversation after pasting in company documents does not mean the information is gone. Gartner warns in a new report, titled in English roughly as "Respond quickly to shadow AI data leaks", that a chat disappearing from the screen is a separate matter from data being erased on the AI provider's servers.

The risk centers on "shadow AI": employees using AI services through personal accounts without company approval. Even on a work computer, logging in to ChatGPT with a personal account may mean corporate security policies do not apply. Using a company device alone does not protect confidential input.

The cost is rising

Gartner cites IBM's 2026 data breach cost report. Between March 2025 and February 2026, shadow AI was involved in 43% of all data breaches, more than double the previous year's share. The average cost of such an incident was $5.39 million, or about 7.2 billion won.

Where the data can linger

AI services may store information beyond the chat itself, including request and response logs, safety review systems, temporary storage and backup servers. In the general retention examples Gartner gives, deleted conversations and server logs can remain for about 30 days. Data flagged for safety review may be kept for 2 to 7 years, and data collected for AI training for up to 5 years. Actual periods depend on the service, account type and data-handling policy.

Information already used to train a model is harder still to remove. Because inputs are statistically reflected in a model's many parameters, deleting the original file does not fully strip out its influence. Machine unlearning, which aims to remove specific information selectively, still has practical limits.

What companies should do

If a leak is confirmed, Gartner recommends first checking AI-training settings and blocking further data input. Companies should then change exposed passwords and credentials, ask the provider to delete the data, and verify whether server logs and backups are retained or deleted.

Banning AI outright is not the answer. Organizations need to know which AI services employees use and what they enter, then set approved services and data-management standards.