Samsung's 2023 ChatGPT Ban Is the Warning HR Has Not Fully Absorbed

By Tom Whitfield

Samsung banned generative AI across the entire company in 2023, weeks after engineers in its semiconductor division pasted proprietary source code and internal meeting notes into ChatGPT. Nothing malfunctioned. The tool worked exactly as designed, and the material left the building anyway, a failure mode that HR and security teams at firms of every size still struggle to see.

That is the uncomfortable core of a new analysis on AI-assisted workflows. The risk, the author argues, is almost never the model being wrong. It is that nobody decided what should go into the tool in the first place, and nobody is accountable for where the data ends up.

What actually failed at Samsung?

Two structural conditions made the leak easy, and neither involved carelessness. First, most generative tools run in the browser. Data loss prevention software watches files, endpoints and network egress, and it often cannot see a paste action into a web application at all. Much of the activity also runs through personal accounts that sit outside every control the organisation has.

Second, consumer-grade accounts can retain what users give them. OpenAI's own documentation states that content from consumer ChatGPT may be used to improve its models unless the user turns that setting off. Paste a client list from the wrong account, and that material may sit inside a training pipeline the employer neither controls nor can claw back. The leak does not trip an alarm, because nothing was downloaded and nothing was emailed; text was simply typed into a box.

The takeaway in plain words: the point of failure sits one layer above the model, in how work is handed off and who decided what was safe to hand over.

What does GDPR demand of an AI workflow?

Where the data identifies a person — a client, candidate, employee or investor — GDPR treats the tools and people touching it as processors acting on the employer's behalf. Under Article 28, every processor must operate under a binding contract that sets out what is processed, for what purpose and for how long, with confidentiality commitments covering everyone who handles the data.

An AI vendor the employer has never papered is a processor it has not accounted for. A free account used quietly by one engineer is a sub-processor nobody signed off. The obligation does not disappear because the tool was convenient.

Location adds a second constraint that most teams miss. Processing outside the EU counts as a data transfer and requires a safeguard such as the Standard Contractual Clauses on top of the Article 28 contract. The UK position mirrors this: the Information Commissioner's Office sets out identical core terms, including documented instructions, security measures, sub-processor rules and deletion or return of data at the end of an engagement.

A second regulatory layer now applies. The EU AI Act, in force since 2024, adds obligations on how AI systems are deployed and governed. The detail is still settling, the analysis concedes, but the direction is clear: unmanaged, undocumented AI use is getting harder to defend, not easier.

Is the human in the loop a security control?

The phrase usually describes a quality check: a person reviews output so errors do not ship. That framing undersells the point, the author argues. The human is also the only part of the workflow that can judge what should never enter the tool at all.

A model cannot decide a board deck stays out of a consumer chatbot. A browser extension cannot know one folder holds client data governed by a contract that forbids sub-processing without written consent. The person doing the delegating is the policy enforcement point, sitting at the exact moment the decision is made — what gets pasted, what gets redacted first, what runs only inside an enterprise account with training switched off.

The division of labour that holds up gives AI the mechanical layer: summarising threads, drafting first versions, extracting action items, reformatting data. The human keeps the judgment layer: what matters, how it is framed, and above all where the data is allowed to go. Remove the human, and the only control reading the situation is gone.

Six steps the analysis recommends

• Classify before delegating. Decide once which data never touches a general-purpose tool: source code the employer does not own rights to expose, client personal data, anything under NDA, unreleased financials, board material. The author calls this the single highest-leverage step and says it costs an afternoon.
• Delegate access, do not share it. Grant delegated access in Google Workspace or Microsoft 365 rather than handing over passwords; delegates work without knowing the login, and access can be withdrawn in one step.
• Keep processing in your own tenant. Use enterprise accounts with model training disabled; personal accounts sit outside every policy and log.
• Contract every processor, including AI vendors. Personal data processing requires an Article 28 arrangement, plus a transfer safeguard if the vendor operates outside the EU.
• Keep accountability inside a structure. A named owner is good; a company that can be held responsible is better, because accountability tied to one individual disappears the day that person does.
• Plan the exit on day one. If offboarding of access, credentials and files cannot be described in two sentences, the workflow is not yet controlled.

The first four steps take about a week, the author estimates. The last two are harder, because they are not configuration changes — they test whether the workflow survives turnover, and most teams discover they got that part wrong only after someone has already left.

For HR leaders, the message is blunt. The speed of AI is not in question; what it cannot supply is judgment about what enters it and accountability for where the output goes. Employers that treat the human in the loop as a security control keep the speed without the quiet leak, while those that remove the human have, in the author's words, automated only the leak.