Is ChatGPT safe for confidential or business data?

On the consumer tiers (Free, Plus, Pro), OpenAI's own documentation states that conversations are used to improve their models by default, unless you turn this off in Settings → Data Controls → "Improve the model for everyone", or use Temporary Chat (which is not used for training and is deleted after 30 days). Inputs are also retained for a period for abuse monitoring. That is the source of the well-known incidents, most famously Samsung in 2023, where employees pasted confidential code into ChatGPT and the firm banned the tool internally.

On ChatGPT Enterprise, ChatGPT Team, and the OpenAI API, OpenAI commits in writing that customer data is not used to train their models by default, SSO/SAML and admin controls are available, and data is retained under a Data Processing Addendum. This is a real, contractual change in posture, and the right baseline for any business use.

• Your data still leaves your perimeter: Every prompt and document is transmitted to OpenAI's infrastructure. The contractual promise is good; the structural exposure is unchanged.
• US jurisdiction applies: OpenAI is a US company. Under the US CLOUD Act, US authorities can compel disclosure of data US providers control, regardless of where it is stored.
• Sub-processors: OpenAI uses Microsoft Azure as a sub-processor. Your data passes through parties you did not directly contract with.
• You do not own the model: The model can change or be deprecated under you. Behaviour that passed your evaluation last quarter may not pass this quarter.
• Prompt logging and human review: Even with training disabled, prompts can be retained for abuse monitoring (typically 30 days) and reviewed by humans in defined circumstances.

The honest test is not "would I email this externally?" but "would I be comfortable if this appeared in a regulator's request or a discovery process tomorrow?". For most regulated organisations, the following categories should not be pasted into a hosted general-purpose LLM, even an enterprise one:

• Personal data at scale: GDPR special-category data, customer records, employee files, anything where you are the data controller for someone else.
• Client-privileged material: Legal advice, M&A drafts, regulatory submissions, audit working papers.
• Source code touching core IP: Algorithms, model weights, security-relevant code, anything that is competitive advantage.
• Patient and clinical data: Even de-identified clinical data carries re-identification risk.
• Government and defence material: Anything classified or covered by export controls.

Under the UK and EU GDPR, sending personal data to ChatGPT makes OpenAI a processor and triggers transfer obligations because the data leaves the UK/EU. Italy's Garante temporarily banned ChatGPT in 2023 over data-protection concerns, and OpenAI has since added DPAs, EU data residency options, and stronger controls. These help, but they do not eliminate two facts: the data still leaves your environment, and you are still accountable as the controller for what happens to it.

The ICO's guidance is straightforward: you must document the basis for processing, ensure proportionality, and assess transfer risk. Pasting personal data into a general-purpose LLM without that work is the unsafe path, regardless of which tier you are on.

• Fine: Public-information research, drafting non-sensitive content, ideation, code that touches no IP or secrets, learning and training.
• Acceptable with care: Internal documents on ChatGPT Enterprise/Team with training disabled, SSO enforced, retention configured, and clear staff policy, for non-regulated material.
• Not safe: Regulated personal data, privileged legal material, core IP, clinical or government-sensitive content, anything you would not want subject to a foreign legal compulsion.

For confidential, regulated, or competitively sensitive data, the answer is not "use ChatGPT more carefully", it is "do not send the data to a third party at all". An owned model deployed inside your perimeter, on-prem, in your private cloud tenant, or air-gapped, removes the structural risk. Your prompts never leave, no foreign provider can be compelled, and you control the model lifecycle.

Locai Labs builds exactly this: a domain-trained model you own, deployed inside your perimeter (Locai One on-prem, your private cloud, or a UK sovereign cloud), under UK jurisdiction and built to fit GDPR and regulated-procurement standards. For most regulated organisations the practical pattern is straightforward: ChatGPT Enterprise for general productivity, an owned Locai model for everything sensitive, and the principle holds, a smaller expert model built on your data routinely outperforms a much larger generalist on the work you actually care about.