What actually happens when your employee pastes client data into Claude or ChatGPT? The answer depends on which plan you're on - and most people using free or personal accounts are operating under terms they haven't read.
Free vs Paid vs Enterprise: The Data Handling Gap
For free ChatGPT users, OpenAI's terms allow the company to use conversations to improve its models unless you opt out. The toggle lives at Settings > Data Controls > "Improve the model for everyone" and defaults to on. Anthropic follows similar logic: free-tier Claude conversations can be used for training, paid Pro users can opt out, and Teams or Enterprise accounts have training data exclusions written into their contracts by default.
This doesn't mean engineers are reading your messages. Training data pipelines are automated, with human review reserved for flagged samples. But if you're handling client contracts, patient records, or proprietary financial projections, the possibility that text could theoretically inform a future model version is a legitimate compliance issue - not paranoia.
The actual risk profile breaks into four distinct concerns:
Training data inclusion. Paid plans give you control; enterprise plans give you contractual guarantees.
Data in transit and at rest. Both Anthropic and OpenAI use standard encryption. Enterprise tiers come with SOC 2 Type II certification. Healthcare organizations can negotiate HIPAA Business Associate Agreements on enterprise plans.
Employee behavior. The clearest example is Samsung's 2023 incident: engineers pasted proprietary source code and internal meeting notes into ChatGPT before the company had any AI usage policy. That data was processed under ChatGPT's standard consumer terms. Samsung banned the tool internally for months afterward. This is the actual vector that damages companies - not the AI spontaneously disclosing data, but humans handling sensitive information without a policy that covers AI tools.
Account compromise. Your full chat history is accessible to anyone who gets into your account. Two-factor authentication is not optional if you use these tools for anything sensitive.
For freelancers and small businesses on Pro plans: the risk is comparable to any cloud productivity tool. Check your data retention settings, enable opt-out where available, and decide which categories of client information simply don't go into AI tools at all.
For businesses in regulated industries - healthcare, finance, legal - you need the enterprise tier with explicit data processing agreements before anything sensitive goes in. "Probably fine" is not a compliance posture, and the free-tier defaults were not designed with regulatory requirements in mind.