AI and Your Safety Data: Privacy, Permissions and Data Handling

Safety managers ask the same question before they connect anything to an AI assistant: is it safe to put incident data in there? Here is the short version. Pasting records into a public chatbot is not safe. Connecting through the Safety Space MCP connector keeps your data inside the same permission model as the platform, and this page explains exactly what that means. It covers who can see what, what the assistant can change, what gets sent, how training is handled, and how to turn it off.

Is it safe to connect incident data to ChatGPT, Claude or Copilot?

It depends on how you connect. There are two very different ways to use an AI assistant with safety data.

• Pasting data into a chatbot. Someone copies an incident report or an export of the incident register into ChatGPT and asks for a summary. The data has now left your systems. You do not control where it is stored, who at the vendor can see it, or whether it is used for training under that person's account settings. Incident records often contain injury and health details, so this is a privacy problem, not just a security one.
• Connecting through the Safety Space MCP connector. The assistant asks Safety Space for the records it needs to answer a question, under the user's own login. Nothing is exported. Permissions, logging and revocation all stay with Safety Space.

Everything below describes the second way. What you can do with the connector is covered on its own page.

Who can see what?

The assistant has the same permissions as the person using it. It can only see the sites, modules and records that the user's Safety Space login can already see.

• A site supervisor asking Copilot about overdue actions gets the actions for their site.
• A group safety manager asking the same question gets the whole group.
• A contractor with access to one project cannot ask about another.

There is no separate AI permission level to configure and no way for the assistant to reach past a user's role. If you want to limit what AI can see, you limit what the user can see.

What can the assistant change?

The Safety Space MCP connector can read and write. That is useful, because a supervisor can log a near miss or raise an action from the assistant without opening the app. It also means you should know the boundaries.

• The assistant can only create or change what that user could create or change in Safety Space.
• Every change is recorded in the audit trail against the user's name, the same as a change made in the app.
• The assistant acts on an instruction from the user. It does not change records on its own.

Is my data used to train AI models?

No. Customer safety data is not used to train AI models, by Safety Space or by the connected assistant. Your incident descriptions, injury details, audit findings and corrective actions are used to answer your questions and nothing else.

What data is sent to the assistant?

Only the records needed to answer the question asked. If you ask about near misses at the Mackay site last quarter, the connector returns those near misses. It does not send the incident register, the whole site, or the rest of your account. Your data stays in your Safety Space account and is read when a question needs it. If you want to know what was sent for a given question, the audit trail shows what the assistant accessed.

How do I turn it off?

A Safety Space administrator can disconnect the assistant at any time and access stops. Access also follows the user. If someone leaves and their Safety Space account is deactivated, the assistant can no longer see anything through their login. You do not need to chase down a separate AI account.

How does this fit with Australian and New Zealand privacy law?

Incident and injury records usually contain health information. In Australia that is sensitive information under the Privacy Act 1988. In New Zealand it is personal information under the Privacy Act 2020, with health information getting extra protection under the Health Information Privacy Code. Both regimes expect you to collect it for a purpose, protect it, and not disclose it beyond that purpose.

Here is how the connector lines up with that:

• The purpose stays the same. The data is used to manage safety, which is why you collected it.
• Access stays within your existing permission model, so disclosure does not widen because an assistant is involved.
• The data is not used for training, so it is not repurposed.
• Access is logged and can be revoked, so you can show what was accessed and by whom.

Our privacy policy sets out how Safety Space handles personal information. We provide our hosting region and sub-processor list to customers on request as part of onboarding or a security review.

Two things stay your responsibility. First, consult your workers before you introduce AI tools that change how safety work is done. Work health and safety law in both countries requires consultation on changes that affect health and safety, and workers have a reasonable interest in how their injury records are handled. Second, set a rule that worker health details do not go into public chatbots. The connector exists so that nobody needs to.

Questions to ask any safety software vendor about AI

If you are comparing platforms, ask each vendor these questions and expect plain answers. If an answer is vague, that tells you something.

1. Does the AI assistant inherit the user's permissions, or does it have its own access level?
2. Can it change records? If so, is every change logged against a named user?
3. Is customer data used to train models, by you or by your AI provider?
4. What is sent to the AI provider for a given question: the records needed, or a wider export?
5. How do we revoke access, and does access end when a user account is deactivated?
6. Which AI providers process the data, and where is the data hosted?
7. Can we see what the assistant accessed?

Safety Space's answers are above. For the last two, ask us during your demo and we will walk you through the current providers and hosting arrangements for your region.