Employees Are Pasting Company Data Into AI Tools
In 2023, three Samsung semiconductor engineers pasted confidential source code, equipment yield data, and internal meeting recordings into ChatGPT to get help debugging and summarizing. It made international news and Samsung banned employee use of consumer AI tools within weeks. Content was rephrased for compliance with licensing restrictions. At the time it read as a cautionary anecdote about a few careless engineers at one company.
The anecdote is now the median behavior
Industry survey data from 2025 and 2026 no longer describes an edge case. One report puts the share of employees who have pasted company information into an AI tool at 77%, and the share of those who did it through a personal rather than company-managed account at 82%. Content was rephrased for compliance with licensing restrictions. A separate vendor telemetry report counts hundreds of millions of data-loss-prevention policy violations attributable to a single consumer chatbot across 2025 alone.
The cost data confirms this isn't a low-stakes habit. IBM's Cost of a Data Breach research attributes a meaningfully higher average breach cost to incidents involving unsanctioned AI tool use, on top of the standard breach cost, and the same body of research finds that a large majority of organizations experiencing an AI-related security incident had no access controls governing AI use in place at all. Content was rephrased for compliance with licensing restrictions.
Why the Samsung framing undersold the problem
Treating the 2023 incident as a story about careless individuals missed the structural point: pasting code or data into a chat interface for help is a completely reasonable thing for an engineer to do, given the tools available to them and the absence of a sanctioned alternative that's as fast. The behavior isn't a discipline failure. It's a predictable response to a productivity gap that most organizations still haven't closed with a governed alternative.
This is the same shape as the "shadow IT" problem from a decade earlier — employees adopting unsanctioned SaaS tools because the sanctioned option was slower or didn't exist — except the artifact being exfiltrated this time is frequently the exact kind of information the data governance and privacy discipline is built to protect: source code, customer records, and internal documents whose downstream reach an organization can no longer trace once it's inside a third party's training or logging pipeline.
What actually closes the gap
Banning consumer AI tools outright, Samsung's original response, addresses the visible symptom and pushes the underlying demand somewhere less visible — personal devices, personal accounts, no logging at all. The more durable fix pairs a genuinely usable sanctioned tool with clear, specific guidance about what may and may not go into any AI system, chatbot or otherwise: source code, credentials, customer PII, and anything covered by a confidentiality agreement need an explicit, named boundary, not a general "be careful" policy that nobody can operationalize.
The measurable version of that guidance is a data classification scheme enforced at the point of use — a browser extension, a proxy, or an API gateway that can distinguish "this text contains what looks like a customer record" from ordinary prose — rather than a policy document nobody consults before pasting. Absent that enforcement layer, the 77% figure is not a training problem. It's the predictable outcome of asking humans to self-enforce a rule with no feedback loop.