It’s 11:40 PM and you’re three commits deep into a hotfix for a payment reconciliation job. Copilot autocompletes a function signature you didn’t type, and it’s uncannily close to the internal naming convention your team invented two years ago in a private repo. You pause. You didn’t teach it that. Or did you?
That moment of unease is the whole story. Backend engineers have spent years building careful boundaries around data — VPC peering, IAM policies, secrets managers, audit logs — and then quietly handed the whole codebase to an autocomplete extension because it saved us forty seconds per function.
The opt-out question nobody reads the answer to
Can you opt out of your input or output data being used for training? The answer depends entirely on the vendor, and the vendors have made it deliberately difficult to compare.
GitLab and Mistral both offer an opt-out. Mistral’s documentation is about as clear as this gets: once the opt-out is confirmed, Mistral no longer uses your input or output data for the purpose of training its models. That’s an unambiguous statement of behavior. You can point at it in a security review.
GitHub’s Copilot is a different situation. Its privacy statement update covers interaction data — inputs, outputs, code snippets, and associated context — being used to train and improve their AI models. Using Copilot inside a private repository means that context is coming from your private repository. There’s a deadline attached, and if you don’t opt out by April 24, GitHub will train on data from your private repos.
The practical problem, and this is the part that should annoy every engineer reading this, is that people cannot find the setting. One user on the FAQ thread reported that the Privacy section only shows “Suggestions matching public code” — not the model training toggle they were told to look for. A privacy control you cannot locate is not a privacy control. It’s a compliance artifact.
Atlassian joins the list
Starting August 17, 2026, Atlassian will begin collecting customer metadata and in-app content from Jira, Confluence, and other cloud products to train its AI offerings, including Rovo and Rovo Dev. There’s an opt-out path here too.
Think about what actually lives in Jira and Confluence at a backend shop. Incident postmortems with exact failure modes. Architecture decision records. Runbooks with internal hostnames. Capacity planning docs with real traffic numbers. Ticket comments where someone pasted a stack trace containing a connection string they forgot to redact.
Your code repository is not your most sensitive system of record. Your wiki is. Code describes what the system does; the wiki describes why, where it breaks, and how to get around the guardrails at 3 AM.
Why this is an infrastructure problem, not a legal one
The instinct is to send this to legal and move on. Resist it. Opt-out defaults are an operational concern for the same reason default-open security groups are.
Every one of these programs shares the same shape:
- The default is participation, not abstention
- The opt-out is per-account or per-org, buried in a settings tree you visit twice a year
- The deadline is announced through a channel engineers don’t monitor
- Nothing in your CI pipeline fails if you miss it
That last point is the one that matters. We have automated checks for expiring TLS certs, drifting Terraform state, and unpinned dependencies. We have exactly zero automated checks for “did a vendor change the default on whether our source code trains their model.” The failure mode is silent and permanent — you cannot un-train a model on data it already absorbed.
So treat it like any other config drift problem. Maintain a list of every SaaS vendor with access to code, tickets, docs, or logs. Record the current training-data setting for each, with a link to the exact settings page and the date you last verified it. Put a recurring calendar item on it. When someone adds a new tool to the stack, the training-data question goes in the same checklist as SSO and data residency.
The regulatory picture is still forming
Europe is moving in a direction that may formalize this. Under the EU AI Act’s opt-out trend, various organizations — including Dutch collecting society Pictoright for images and French society Sacem — are pushing on data use for training. What’s not settled is how a legally valid opt-out request would actually be made. There’s no agreed mechanism yet.
Which means for now the burden sits with you, in a settings panel, before a date you probably didn’t have on your radar. Not a satisfying answer. But it’s the accurate one, and pretending otherwise doesn’t get your code back.
Go check your settings. Then write down what you found, because you will need to check again.
🕒 Published: