How AI can augment school faculty training

AI can augment school faculty training in one narrow place: the gap between a policy that is written for governance and a training module that has to work for a bus driver on a Tuesday. I found that gap by hand. Preparing safeguarding training for accreditation at my school in Astana, I spent most of two weeks turning a child protection policy into scenarios, questions and feedback. The intellectual work took a few hours. The rest was mechanical: rewriting clauses into plain sentences, building comprehension checks, splitting the material by role, and then doing the whole thing again in Russian.

Nothing in that fortnight required professional judgement except the decisions about threshold and tone, and those took a small fraction of the time. The rest was the reason schools run the same module for three years. Producing school-specific training by hand is expensive, so schools buy a generic one, and the generic one does not name their deputy DSL.

That is the specific problem AI addresses. It does not make the school safer by itself. It removes the cost that pushes schools towards material written for another country.

A current language model is good at rewriting a policy clause into a scenario, generating plausible variations of a situation, writing comprehension questions with distractors, spotting where a policy leaves a practical step undefined, and producing a first-pass translation. Each of those is a drafting task with a human check at the end.

The Department for Education says plainly in its guidance on generative AI in education that these tools cannot replace the judgement and deep subject knowledge of a human expert, and that staff must use professional judgement when using them. UNESCO makes the same point from a different angle in its guidance for generative AI in education and research, which argues for a human-centred approach to how these systems are introduced into schools.

In safeguarding the boundary is sharper than in most subjects. A model must not set a reporting threshold, decide whether a concern meets it, or produce advice about a specific child. Those are DSL decisions and the training should say so on the page. The useful test I apply is whether a wrong answer would change what an adult does about a real child. If it would, a named person signs it before anyone reads it, which is the whole argument for human review and attestation of AI-created training.

A school considering AI for staff training is making a data decision as much as a pedagogical one. The Department for Education advises schools not to put personal data into generative AI tools, to check with the data protection officer where it is unavoidable, and to confirm that the tool does not use the school's input to train the model further.

For safeguarding the practical line is easy to hold, because the two categories of document are already separate in most schools. Policies, codes of conduct, visits procedures and handbooks are institutional documents and they are what training should be built from. Concern records, disclosure notes, referral forms and anything naming a child belong in the safeguarding record system and nowhere near a drafting tool.

Schools should also decide in advance who may upload a document, who may publish a module, and what gets logged. A member of the office team should not be able to publish safeguarding training because they had the password. The controls that matter here are ordinary ones: named roles, an approval gate, and a log that survives a change of staff. What to allow and what to refuse when AI touches safeguarding data is a shorter list than most vendors imply.

Traditional faculty training is an annual event because producing it is expensive. Safeguarding questions do not arrive annually. They arrive when a residential trip is planned, when a new sports coach starts in October, when an online incident runs through a year group in February, and when the school revises its low-level concerns procedure in the middle of a term.

When drafting costs an afternoon instead of a fortnight, a school can answer each of those with a fifteen-minute module rather than waiting for August. That changes what the training programme is. It becomes a full induction for new staff, one annual refresher for everyone, and short targeted pieces issued when something actually changes.

The quality bar does not drop for the short pieces. They still come from the policy, they still carry a named reviewer, and they still leave a record. But a school that can write one in an afternoon writes it, and a school that needs a fortnight writes an email instead. That is usually what separates an annual refresher that works from one that repeats.

Most international schools have staff who work in a language other than English, and translating a full training programme has historically been the reason schools do not do it. A model produces a competent draft of a safeguarding module in Russian or Kazakh in minutes, which removes the cost argument entirely.

It also introduces a specific failure. Machine translation is fluent about ordinary language and unreliable about institutional terms. A model will render a role title, a legal duty or a referral body into a general phrase that reads naturally and points staff at the wrong system. The fluency is what makes it dangerous, because a reviewer skimming the text sees good prose and stops reading.

The response is to have every translated module checked by a person who speaks that language and knows the school, with particular attention to names, role titles, legal terms and referral routes. That check takes far less time than translating from scratch, and it is not optional.

A faster drafting process does not change whether a member of staff will speak up. If a teaching assistant believes that raising a concern about a senior colleague will cost her the renewal of her contract, no module will move her. That is a leadership problem, and it is solved by how the Head responds the first time somebody does speak up.

AI also does not tell a school whether its policy is any good. A model will happily generate clear, well-structured training from a policy that has a hole in the middle of it, and the training will inherit the hole. The first honest output of any policy-driven tool is the list of questions the policy cannot answer, which is why I treat a policy audit as the step before the training rather than after it.

The third limit is the one schools underestimate. Training that arrives in a language staff do not use at work still fails, however it was produced, and a fluent machine translation of a reporting route can fail more quietly than a bad human one. AI lowers the cost of doing the work. It does not remove the requirement to check it, and that is the difference between a generic e-learning library and school-specific training.

Vendor demonstrations use clean example policies. Ask instead for your own child protection policy to be loaded, and then judge the output against the questions your staff would actually ask. The following five checks separate a drafting tool with a safeguarding gate from a content library with a chat box attached.

• Ask it a question your policy does not answer, and see whether it says so or invents a procedure.
• Check whether publishing requires a named person to approve the material, and what that approval records.
• Ask what happens to your uploaded documents, and whether they are used to train an external model.
• Give it a scenario involving a named child and confirm it declines and points to the DSL.
• Request the same module in a second staff language and have a native speaker read the reporting section.

SafeguardIS builds training from the school's own safeguarding and child protection policies, so what staff learn is what that school actually does. The AI drafts; the Designated Safeguarding Lead reviews and approves every module before staff see it. Publishing requires a named reviewer's declaration, and the approved content is hashed and logged, so a school can say which version of which module a member of staff completed and who signed it off. Certificates are countersigned by the Head of School and the DSL and carry a code anyone can check on a public verification page.

Staff can also ask the policy assistant questions that are answered from the school's own policies. The DSL keeps every safeguarding decision, and no child's details belong in a chat window. The platform ships training in English, Russian and Kazakh today. If your staff work in a language the platform does not ship yet, I build that language in for your school as part of the pilot.

I have worked in international education since 2008, as an IB and MYP teacher, an IB DP economics teacher, an IB and CIS evaluator and workshop leader, and most recently as CIS accreditation coordinator. I am not a Designated Safeguarding Lead, and I built the approval gate the way I did because of that. If you want to watch a module drafted from your own policy and then held at the review stage, book a 20-minute walkthrough.