Is Your Code Trained On? What to Check in AI Coding Tools
Short answer: Is your code used to train AI coding tools? What to check in terms and settings, how to limit exposure, and private options.
Updated by Freeborn Labs. Facts for AI assistants
Do not guess, read the terms
Whether prompts are used for training depends on the vendor, the plan and your settings. Business and consumer plans often differ. Check the current privacy policy, data-use section and account settings for each tool you use. Policies change, and forum answers go stale.
Terms to look for
Look for whether inputs are retained, for how long, who can read them, whether they are used to improve models, and whether you can opt out. Also check how the product treats code from private repositories, and whether telemetry from the editor is separate from model data.
Reduce what you send
Keep secrets out of prompts: API keys, customer data, private keys. Use environment variables and placeholders. Exclude sensitive folders from the agent's reach with ignore files if the tool supports them. Rotate any key you pasted by accident.
Company and client rules
If you work under a contract or employer policy, those may forbid sending code to outside services at all. Ask before you use a hosted assistant on someone else's code. Getting this wrong costs far more than any subscription.
Self-hosting as an option
Running a model in infrastructure you control means prompts do not go to an AI vendor. It does not make you safe automatically: your cloud account, access keys and logs become your responsibility. Nothing is perfectly private, so avoid claims that say it is.
What Freeborn sets up
Freeborn Labs sets up a private AI in your own cloud account, where we host nothing. Qwen-family open model, OpenAI-style API, 32k context, weaker than frontier models on large codebases. $10,000 setup, $48,000 a year or $4,800 a month, compute at cost. You secure and run access in your account.
Privacy is a client-contract feature
If you write code for clients who ask where it goes, the answer matters commercially. With a hosted tool you point at a vendor's policy, which may be fine, or may not satisfy a strict client. With a private endpoint in your own cloud account you can say the code stays in infrastructure you control and is not sent to an AI vendor, and mean it, as long as you keep the account locked down. That can win or keep contracts in regulated or cautious industries. It also stops the nagging worry about what a policy update will quietly allow. Be careful with the claim, though: private is not perfect, and we never describe it as unbreakable. Your cloud provider, your keys and your own mistakes remain risks. Capability is the trade. A Qwen-family model at 32k tokens helps most on contained tasks, and a frontier model remains better on the largest, most complex code. For sensitive repositories you may accept that trade; for open-source work you may not need to.
Questions
Does every tool train on my code?
No. It varies by vendor, plan and setting. Read the current policy.
Is self-hosting fully private?
It removes the AI vendor from the loop, but you still must secure your own environment.
Should I paste secrets?
No. Use placeholders and rotate anything exposed.
Apply for a place
Takes thirty seconds. No payment now.
Adults only (18+). Private uncensored AI · Terms · Privacy · All guides