A single educator observation note can contain a child's full name, a photo, a developmental concern and a remark about a family situation. A centre produces dozens of these a day. That is the data an early learning provider is really deciding about when someone suggests running an open-weight model locally to save on AI subscriptions.
Early learning and childcare providers operate under some of the strictest information-handling obligations in Australia, layered on top of standard Privacy Act rules. Working With Children Check records, incident reports involving children, and daily observation notes that often include photos all sit inside the same service. Any AI vendor evaluation in this sector starts from a much stricter baseline than most industry guides assume.
Can a childcare centre use open source AI safely?
Yes, for tasks that never touch child-identifying information, such as drafting generic parent update templates, rostering messages or staff induction summaries. For anything that names a child, including observation notes and incident reports, the model type is not the deciding factor. What matters is a documented data processing arrangement, a legal review against your state's child-safety obligations, and an educator reviewing every output.
That answer disappoints people who expected open source to be the privacy-friendly option by default. Running a model on your own hardware keeps data on premises, but it also makes the centre responsible for securing that hardware, controlling who can see the logs and proving it to a regulator. For most operators, that is the harder half.
What works today
Drafting routine parent update templates and rostering communications, where no child-identifying detail needs to touch the model at all.
Summarising internal policy documents and compliance checklists for staff induction, which involves no child data whatsoever.
Turning a director's rough notes into a newsletter or event notice that goes to every family, with names added by staff afterwards.
These are low-risk on any model, commercial or open-weight. They also recover real hours: the admin time that otherwise lands on a centre director after close.
What we would not recommend, whichever model you choose
Processing observation notes or incident reports that name a child through any model, open-weight or commercial, without a documented data processing agreement and a legal review specific to child-safety obligations in your state.
Using any AI tool to draft communications about a specific child's behaviour or development without an educator reviewing every line before it goes to a parent.
Storing children's photos in a model's prompt logs or fine-tuning data, where they are hard to find, hard to delete and easy to forget about.
Where child-identifying work is in scope, it needs to be designed with the controls in from the start. Our piece on Claude for childcare centres covers programming and family communications with educator sign-off built into every step.
| Task | Child data involved | Risk level | Reasonable approach |
|---|---|---|---|
| Generic parent newsletter | None | Low | Any model, staff add names |
| Staff induction summaries | None | Low | Any model |
| Rostering and shift messages | Staff data only | Low to medium | Governed tool, no child data |
| Learning story drafts | Names, observations, photos | High | Governed service, educator review |
| Incident report drafting | Names, injuries, circumstances | Very high | Legal review before any AI use |
The cost reality for a small operator
Most standalone centres and small multi-site operators do not have the engineering headcount to properly secure a self-hosted open-weight deployment. The free model often ends up more expensive once you add a part-time contractor to patch it, monitor it and answer the auditor's questions. We made the general version of this argument in why most Australian SMBs should not self-host an LLM yet; in childcare the margin for error is smaller still.
By comparison, a properly governed, Claude-based communication drafting tool for a 5 to 10 centre operator typically runs $4,000 to $9,000 to set up, plus a modest monthly subscription, with the compliance documentation built in from day one rather than added afterwards. Claude's enterprise data handling terms do a lot of the compliance work that a small centre cannot replicate in-house.
Questions to put to any AI vendor, open or commercial
Before any tool touches your service, ask these in writing and keep the answers on file:
Where is prompt and output data stored, and for how long?
Is any of our data used to train or improve the model?
Who at the vendor, or in our own team, can read the logs?
How do we delete a specific child's data on request, and how is that confirmed?
What happens to our data if we stop using the service?
A vendor who cannot answer these clearly is not ready for a children's service, and a self-hosted setup where nobody in your team can answer them is in the same position. Aged care providers face a similar narrowing of options, covered in open source AI for aged care and allied health.
As at September 2026, our view for this sector is that the administrative overhead of running your own model safely is rarely worth it below a certain scale. If you operate a childcare or early learning service in Sydney, Melbourne or anywhere else and want a plain-English read on what AI can and cannot touch, book a session with us or use our readiness assessment as a starting point.
