Could Your SaaS Provider Train AI on Your Data?
Your Data Is the Product
When work lives in SaaS, a change of terms can turn business content into AI input. In 2025, LinkedIn said it would use member data to train generative AI unless users opted out; Malwarebytes reported that the opt-out would not undo data already used and that the official statement gave members in the EU, EEA, Switzerland, Canada, and Hong Kong until November 3, 2025 to opt out. Zoom’s 2023 terms also drew scrutiny because they granted broad rights over customer content, and TechCrunch reported that Zoom later added language saying it would not use audio, video, or chat customer content for AI training without consent.
Neither episode proves that every provider reads your files. What they show is where control sits: the provider writes the terms, changes them, and decides how much detail customers get. If your files, messages, recordings, and activity logs live there, you rely on a contract you did not draft and telemetry you cannot audit.
Whether any given vendor is careless is beside the point. The practical question is whether a company’s core work should sit on platforms where AI and analytics rules can change faster than procurement notices them.
Security: The Silent Risk
Privacy is not the only trade-off. The 2025 OSSRA report found known open-source vulnerabilities in 86% of audited commercial codebases and high- or critical-risk vulnerabilities in 81% of them.
The numbers say less about open source than about how much ongoing care any software supply chain needs.
SaaS is not automatically safe. You do not see the vendor’s dependency inventory, patch schedule, or incident drills. If something goes wrong, you usually learn from a status page, a disclosure email, or the news.
A forgotten self-hosted server is worse. The same report found that 90% of audited codebases had components more than four years out of date. An old Nextcloud or Mailcow install exposed to the internet is unmanaged risk, whatever independence it was meant to buy.
Ownership only pays off when updates, patches, backups, and monitoring keep happening on an ongoing basis, not just at setup.
The Tools Are Mature. Someone Still Has to Run Them
Finding an open-source app for each job is easy now. The hard part is operating the chosen stack together.
Each service brings users, storage paths, ports, certificates, backups, and upgrade notes. DNS has to be correct. TLS has to renew. Backups have to restore, not merely exist. One update can change how services talk to each other, and a rollback without a tested restore path can turn a small outage into data loss.
Maintenance also depends on the people behind the code. Projects go quiet, and someone has to notice, track versions, and decide when to replace a dependency that nobody develops anymore. That is part of operations.
For a small company without its own administrator, this is invisible work: nobody budgets for it, but everyone notices when it stops.
What to Check Before You Trust a Contract With Your Content
The LinkedIn and Zoom cases above both trace back to specific contract language, not to any particular vendor being unusually careless. That suggests a few concrete things worth checking in any SaaS agreement before business content goes into it.
Find the actual clause, not just the privacy policy. Zoom’s AI training language was not in a privacy notice; it sat in the terms of service, in clause 10.4, “Customer License Grant”, on the license the customer grants the vendor (The Register). TechCrunch’s review of Zoom’s terms noted that its bolded no-AI-training pledge was scoped to “audio, video, or chat Customer Content” specifically, with no equivalent pledge for usage/telemetry data (TechCrunch).
Read the duration of the license grant. Zoom’s 2023 terms had customers grant “a perpetual, worldwide, non-exclusive, royalty-free, sublicensable, and transferable license” over customer content — a license that, as The Register put it, arguably let Zoom apply that content to machine learning and artificial intelligence — before the company added a consent carve-out for audio, video, and chat content specifically (The Register). A license described as perpetual is a different commitment than one that ends when the contract does — check which one is on the page in front of you, and separately check whether the agreement’s termination clause says the content license survives.
Check whether opt-out is retroactive. When LinkedIn announced it would train generative AI models on member data by default, the opt-out only stopped future use: Malwarebytes reported that “opting out only stops future training on new data; it does not retract data already used” (Malwarebytes). If content is already used before you find the toggle, opting out later does not undo that.
Check the default state of any sharing or training toggle. TechCrunch found that in Zoom’s own admin UI, a data-sharing checkbox tied to features like AI meeting summaries was “pre-checked by default, thereby requiring the admin to take the active step of unchecking it” to opt out (TechCrunch). Whether a setting is opt-in or opt-out by default decides who has to act, and who has to notice there is anything to act on.
None of this is legal advice, and none of it says whether a given clause is enforceable in your jurisdiction. It is a starting checklist for where to look and what to read closely before deciding what business content is safe to put behind someone else’s terms of service.
The two halves of this article pull against each other on purpose. Someone else’s terms can change under your data, and running the alternative yourself is real, recurring operational work that a small company rarely budgets for. Neither side is free. The point is to make that trade deliberately rather than by default.
principale.dev installs open-source tools like these on a server you own. Every install is free. See how it works.