“Safe” sounds like a yes-or-no property of the software. Under Canadian privacy and security guidance, it's closer to a due-diligence checklist a business has to run for itself.
Short answer
It depends on how it's used, not on the product name. The specific risk Canada's Cyber Centre names is that “users may unknowingly provide sensitive corporate data or personally identifiable information (PII) in their AI queries and prompts.” Whether that risk actually materialises turns on the account type in use, what employees are told they can type into it, and whether the safeguards around it match the sensitivity of that information — not on anything the tool decides on its own.
Canada’s Cyber Centre lists “privacy of data” as one of eight specific risks of generative AI: “Users may unknowingly provide sensitive corporate data or personally identifiable information (PII) in their AI queries and prompts” (Cyber Centre, ITSAP.00.041). The full list of eight, and what to do about each, is set out at AI and cyber security basics; this risk in particular is the one that decides whether an ordinary chat tool is safe for a specific business’s actual use, because it doesn’t depend on an attacker at all — it happens through normal, well-intentioned use.
Once information does go into the tool, PIPEDA’s safeguards principle sets the bar: “Personal information shall be protected by security safeguards appropriate to the sensitivity of the information,” and clause 4.7.2 adds: “The nature of the safeguards will vary depending on the sensitivity of the information… More sensitive information should be safeguarded by a higher level of protection” (PIPEDA, Schedule 1, clauses 4.7.1–4.7.2). A public chat tool being used for generic drafting carries a different risk profile than the same tool being fed client financial records or health information — the standard scales with what’s actually at stake, not with the tool’s reputation.
The underlying discipline here is the same one a business would run before taking on any third party’s handling of its data: does the party actually handle it responsibly, on the facts, rather than on the strength of its name. A Treadstone Law article on assessing a business’s data practices before an acquisition frames that discipline plainly — “does the target actually handle it responsibly?” — as a general due-diligence principle, not one written with AI adoption in mind, but the same underlying question applies before adopting a tool as before acquiring a company (Treadstone Law, on data due diligence generally).
In practice that checklist is short: which account or tier is being used and what its own terms say about retaining or training on inputs; what employees have actually been told not to paste into it; and whether anyone has looked at what categories of information have gone in so far. None of that is answered by the product’s name, and none of it is answered once — see should a small business worry about AI security for how this scales (or doesn’t) with the size of the business asking.
Turning “what's safe to type into this” into an actual usable policy employees will follow is where this becomes an operations question, not a one-off decision.
