· Early Childhood Teacher
I tried writing child observations with AI on three under-5s last term, and I deleted all three.
The one that stuck with me: I asked the AI to draft an observation against EYLF Outcome 1 for a four-year-old who had just sorted out a tug-of-war over a toy in outdoor play. It handed me back a child who “demonstrated an emerging capacity for self-regulation and prosocial negotiation strategies.” That is not what I saw. This post is how I now write child observations with AI without losing my voice: the raw note stays mine, the AI only tightens, and the child’s actual words never get sanded off.
The fear that stops most educators using AI for observations
Most educators I talk to are not scared the AI will be wrong. They are scared it will be smooth. There is a difference.
”It won’t sound like me”
When you have written hundreds of observations, your documentation has a fingerprint. You notice transitions. You quote children oddly and on purpose. A generic draft strips that out and replaces it with the same four adjectives every centre uses. A parent who reads your fortnightly notes can feel the swap instantly, even if they could not name what changed.
Why a generic AI observation reads hollow to a parent
A family does not want to read that their daughter “engaged in cooperative play.” They want to know she looked her friend in the eye and said something specific. The hollow version is technically true and emotionally empty. That gap is the whole risk, and it is why I never start with the AI.
My before-and-after: one real anecdote, two write-ups
Here is the four-year-old and the toy. Same moment, two ways of recording it.
The rough notes I jotted at the sandpit
What I actually scribbled, in my own shorthand: child wanted the digger another child was using. Did not grab. Looked the friend in the eye. Said “you can have it after I count to five,” then counted out loud on their fingers. Friend waited. Handover happened with no adult prompt. That is the raw note. Five details, all observed, all mine.
The over-edited version that lost the child
The AI’s version compressed all of that into “self-regulation and prosocial negotiation strategies.” Accurate label. Dead moment. The verbatim line, “you can have it after I count to five”, vanished, and with it the evidence. A three-year-old who says “I maked it go faster” is telling you something precise about language development that “used descriptive language” quietly erases. The AI wrote the conclusion. I write the moment. Two things no prompt can give it: the child’s exact words, and the physical detail. Which hand they reached with. Whether they leaned in or held back. The three seconds they spent staring at the paint before they touched it. The pause is the learning.
A safe prompt recipe that keeps your wording
My rule now is simple. The AI helps with structure, grammar, and linking to EYLF outcomes. The observational content stays mine. I write the raw note first, then ask the AI to tighten it. I never start with the AI and edit down.
The “don’t add anything I didn’t mention” guardrail
The single most useful line I paste is a hard constraint: “Tighten the grammar and structure of these notes. Do not add any actions, words, or detail I have not written. If something is missing, ask me, do not invent it.” This stops the model filling gaps with plausible fiction. The first time I used it, a draft went from invented eye contact to a flat note asking me whether I had actually seen the child make eye contact. I had. So I added it myself, in my words.
Strengths-based phrasing without invented detail
I do ask for strengths-based language, because that aligns with how the framework wants us to see children. But strengths-based is a lens, not a licence to embroider. “The child persisted” is fine if the child persisted. It is not fine if the AI decided they persisted because that sounded nice.
Keeping the child’s own words in
I now flag verbatim quotes in my raw note with quotation marks and tell the AI to preserve anything in quotes exactly. “I maked it go faster” survives (and honestly, that line is my favourite sentence in the whole entry). If I had let the model paraphrase it into standard English, I would have destroyed the one piece of evidence that made the observation worth keeping.
Linking the observation to EYLF outcomes
This is where AI genuinely earns its place. Australia’s Early Years Learning Framework V2.0 (Belonging, Being and Becoming) sets out the five learning outcomes, and matching a moment to the right one is a real cognitive task at 5pm when you are tired. Or it was, anyway, before I started letting the model take a first guess.
Letting AI suggest the outcome, you confirm the reasoning
I let the model propose an outcome and a one-line rationale. Then I check it against what I saw. For the digger moment it suggested Outcome 1, children have a strong sense of identity, specifically the capacity to interact with care and respect. I agreed, but I rewrote the rationale in my own sentence so the link reflected my observation, not a template. The AI suggests. I confirm. Those are not the same job.
Reflection and forward planning stay yours
The assessment-and-planning cycle in the National Quality Standard Quality Area 1 expects educators to critically reflect and plan forward from what they document. That reflection is professional judgement. I have never once let the AI write my “where to next.” It does not know this child, this room, or what we tried last week. I do.
Documenting quiet and non-verbal children fairly
The risk with any structured template, AI or not, is that it rewards the children who perform for it.
Why templates over-reward talkers
The chatty four-year-old who narrates everything generates easy, quotable observations. The quiet child, or a child who is not yet verbal, can disappear from your documentation simply because they give you fewer obvious words to write down. That is an equity problem, not a paperwork one.
What to look for: gaze, intent, engagement, transitions
For non-verbal and quiet children I write the raw note around what the body tells me: where the eyes went, how long the attention held, how they moved through a transition, who they chose to be near. Then I ask the AI only to tidy that, with the same “add nothing” guardrail. When I write observations for non-verbal children with AI, the model is even more tempted to invent intention, so the constraint matters more, not less. The evidence has to be something I genuinely watched.
What this actually saves me, and what it does not
People want a number, so here is mine. Before, a single individual observation took me roughly 11 minutes from raw note to filed entry. With the AI tightening grammar and proposing the outcome link, it now takes about 7. That is four minutes a child. Across a week of observations, it adds up to most of a planning hour I get back.
But notice what those four minutes cover. They do not cover the watching. They do not cover the raw note. They do not cover my reflection or my forward planning. The AI shaved time off the typing and the formatting, the parts that were never the point. The minutes I protect, on purpose, are the ones spent at the sandpit actually seeing the child. If a tool ever offered to save time on those, I would say no.
There was one near-miss worth naming. Early on I let a draft through that read beautifully and described a child “sharing willingly.” When I reread my raw note the next morning, I had written that the child handed the toy over only after I sat down beside them. The AI had quietly upgraded a prompted handover into a spontaneous one. I caught it before it reached the family, changed it, and that is the morning the “add nothing” guardrail became permanent.
My honest opinion: AI is making documentation worse for the educators who use it badly
I will take a stance most AI-for-educators posts will not. Used the way most people are using it (paste a vague prompt, accept the polished draft, move on), AI is actively lowering the quality of our observations. It produces documentation that passes a quick skim and fails a child. The fluency is the trap. A hollow note that reads well is more dangerous than a clumsy note that is true, because nobody flags it. My line is hard: if I cannot write at least the core observation myself, I do not publish it. The AI is allowed to make my writing clearer. It is not allowed to make it up. Educators who get that order backwards are not saving time, they are quietly hollowing out their own evidence base.
Privacy, NQS and what to never paste into an AI tool
Before any of this, there is a non-negotiable step.
De-identify before you draft
I never paste a child’s full name, date of birth, photo, or any identifying family detail into a general AI tool. My raw notes go in with the name swapped for “the child” or initials I strip afterward. This is a personal practice rule, not legal advice — check your own centre’s privacy policy and how it treats third-party tools.
Centre policy and supervisor sign-off
In my practice, anything involving children’s data goes past my supervisor and our centre policy first. If your service has not yet decided where AI sits in your documentation workflow, that conversation belongs in a staff meeting before it belongs in your observations. Ask your nominated supervisor. The framework expects our documentation to be trustworthy, and trust starts with knowing where the data went. I wrote more about keeping AI-assisted writing in my own voice in the four prompts I used to make a parent newsletter sound like me, and the wider planning cycle still belongs to you, not the tool.
TL;DR / Key Takeaways
- Write the raw observation yourself first; only then let the AI tighten grammar and structure. Never start with the AI and edit down.
- Use a hard guardrail prompt: “Do not add any actions, words, or detail I have not written — ask me instead of inventing.”
- Keep verbatim child quotes and physical detail. “I maked it go faster” beats “used descriptive language” every time.
- Let AI suggest the EYLF V2.0 outcome, but confirm the reasoning and write your own reflection and forward planning.
- De-identify before pasting anything, and clear AI use with your supervisor and centre policy first.
For more on AI in early childhood without losing the human part, the same rule holds across every tool and every template: write the human bit yourself, let the tool tidy the rest.
Sources
- Approved learning frameworks — EYLF V2.0 (Belonging, Being and Becoming), ACECQA
- Quality Area 1 — Educational program and practice, National Quality Standard, ACECQA
Fact-checked 2026-05-29. Last reviewed 2026-05-29.