GitHub Copilot Trust Center: What to Check Before Approval
Key takeaways
- The GitHub Copilot trust center answers different questions depending on the plan, so check the tier before you read the answer.
- Business and Enterprise prompts and suggestions are not used to train GitHub's models; individual plans can be, unless the user opts out.
- Opting out of training does not change retention, so check both settings separately.
- The trust center is a starting point. The Data Protection Agreement is the document that binds GitHub, so get it logged.
- Our AI Tool Risk Directory records no BAA for GitHub Copilot, so do not approve it for patient data on that basis.
If a developer on your team has asked whether GitHub Copilot is safe to use on company code, the first place most people look is GitHub's Copilot trust center. It is a useful page, with the certifications, the privacy statement and the data-protection terms in one place. It is also a page that answers different questions for different products, and a quick read can leave you approving the wrong plan. This guide sets out what the trust center covers, where the answers change between tiers, and what to write into your AI usage policy once you have checked it.
What the trust center is, and what it is not
The GitHub Copilot Trust Center brings together GitHub's security and compliance material for Copilot, including its SOC 2 Type 2 and ISO 27001 status and a FAQ that covers how data is handled. GitHub's own write-up on adopting Copilot responsibly explains the same material for organisations. What the page is not is a contract. For enterprise customers, GitHub acts primarily as a processor for personal data in Copilot for Business, with those commitments set out in its Data Protection Agreement. If you are subject to UK or EU GDPR, the DPA is the document to log, and a public page should not be the only evidence you hold.
Start with the plan, not the brand
GitHub sells Copilot in tiers, and the data terms differ between them. Individual plans (Free, Pro and Pro+) are bought by a person with their own account. Copilot Business and Copilot Enterprise are bought by an organisation and governed by its agreement. A team that says 'we use Copilot' may be running a mix of the two, and the trust center answer for one does not transfer to the other. Before you read any answer on the page, confirm which tier the people in your team are actually signed into. Your policy should name the tier, not just the product.
Training: the question that splits the plans
On Copilot Business and Enterprise, GitHub states that it does not use customer prompts or suggestions to train its models, and that includes third-party models offered through Copilot. The commitment sits in the organisation's agreement rather than in a personal preference that each developer has to set. On individual plans the position is different. Prompts and code snippets may be used to improve models unless the user disables that in their Copilot settings. That setting belongs to the individual, so your organisation cannot see it or enforce it, which is the main reason to keep company code off personal accounts. The general mechanics of this are covered in our training opt-out glossary entry.
Retention: opting out is not the same as deleting
Training and retention are separate settings, and a common mistake is to treat a training opt-out as a data-deletion step. Sources summarising GitHub's Business and Enterprise retention FAQ say that IDE prompts and suggestions for chat and code completion are not retained, while engagement data is kept for about two years. For individual plans, our directory records prompts as retained for around 28 days for the code-completion service. We hold that figure at medium confidence, because GitHub's documentation differs across Copilot surfaces. The coding agent and the command-line tool do not always behave like code completion, so check the retention statement for each surface your developers actually use, not only the headline one.
Certifications and the paperwork
The trust center states that Copilot holds SOC 2 Type 2 and ISO 27001 status. Those are worth checking by date rather than by badge, because certificates can lapse quietly, and the audit report itself is a separate document you should request from GitHub rather than assume from a logo. Our directory record for GitHub Copilot also shows a GDPR DPA as available, SAML single sign-on as supported on the business tiers, and no BAA. It has no ISO/IEC 42001 entry at all, which means we have not confirmed one. Treat a null in a record as 'not confirmed', not 'no', and ask GitHub directly for anything your auditor will need.
A worked check from our own directory
Suppose a 30-person software company wants to let its developers use Copilot on the company's private repositories. Its policy team opens the trust center, reads the reassuring summary about not training on business data, and is ready to approve. Before signing off, it checks the same questions against the directory record we keep for the product, which is the record behind our GitHub Copilot profile. The record shows that training on business data is 'no', training on personal data is 'opt-out', an opt-out is available, SOC 2 and ISO 27001 are 'yes', a GDPR DPA is 'yes', SSO is 'yes', and there is no BAA and no EU data-region commitment on file. The outcome is clear. The company approves Copilot Business for company repositories only, requires that individual accounts are not used for company code, and records that it has no BAA, so the tool is not approved for any patient or health data. The policy the company writes says those things in plain words, and the trust center alone would not have produced the last two.
A short checklist for the review
Run the same review every time a team proposes a coding assistant. Confirm the tier and the account type that will be used. Read the training answer for that tier, not the general statement. Read the retention answer for each surface you will allow, including chat, completion, the command-line tool and any agent. Check the DPA and log it in your data processing register. Check the certifications by their issue dates. Ask about data residency if you handle EU or UK personal data. Ask for a BAA in writing if health data is in scope. Record the date you checked and the page you checked it against, so the record is still useful in a year.
What to write into your AI policy
A coding assistant needs its own line in your policy, because the risk is different from a general chat tool. Name the approved product and tier. State that personal accounts are not approved for company code, and say why in one sentence so people understand the rule. Require human review of suggested code in the same way as any other contributor's pull request, since licence and quality questions still apply. If you allow the tool on private repositories, say so explicitly and name who approves that. Our free AI usage policy generator builds these clauses from your answers, so you do not have to draft them from scratch.
Keep the review current
Copilot changes faster than most enterprise tools, with new surfaces, new models and terms that change over time. Set a quarterly reminder to re-read the trust center and the retention statement, and compare them against your own notes. If your team also uses Microsoft's Copilot, keep the two apart in the policy, because the terms are different products with different owners. Our guide to Microsoft Copilot and data privacy covers that distinction, and our vendor vetting guide sets out the wider review we run before approving any tool.
| Question | Individual plans (Free, Pro, Pro+) | Copilot Business and Enterprise |
|---|---|---|
| Are prompts and suggestions used to train models? | Can be, unless the user turns it off in Copilot settings | No. GitHub does not use these to train its models |
| Who controls the training setting? | The individual, in their own account | The organisation, through its agreement |
| How long are prompts and suggestions kept? | About 28 days for code completion, per our directory record (medium confidence) | Not retained, per summaries of GitHub's retention FAQ |
| Is SAML single sign-on available? | Not recorded for individual plans | Yes, per our directory record |
Related guides
- What Is an AI Policy? A Plain-English Guide
- ISO/IEC 23894: AI Risk Guidance vs ISO 42001
- How to Mitigate AI Risk: A Step-by-Step Framework
Check a specific tool in the AI Tool Risk Directory.