· Early Childhood Teacher
A two-year-old in our room had not spoken a word for six weeks. Then one Tuesday morning, out of nothing, she said her older brother’s name. I wrote the raw note, pasted it into an AI assistant to tidy up, and watched it confidently turn a moment about attachment into a tidy little milestone about language. That is the day I learned the real answer to “can AI write childcare observations.” It can write words. It cannot know what you witnessed. Here is exactly where it got the story wrong, and the observation I ended up writing by hand.
The morning a silent two-year-old said one word
Let me set the scene properly, because the whole point lives in the detail.
Six weeks of silence in our room
She had been with us for about two months. Bright, watchful, deeply attached to routine. For six weeks she had not spoken in our room. Not to me. Not to the other educators, and not to a single one of the children. I want to be careful here: I am an ECT, not a clinician, and “had not spoken in our room” is a description of what I observed, not a diagnosis of anything. We were tracking the pattern, keeping her family in the loop, and I had flagged it with my supervisor (that last part matters, and I will come back to it). The right lane for me to stay in.
What I actually saw and heard
That Tuesday, her older brother walked past our door on his way to the pre-school room. She looked up, and she said his name. Clear as anything. Then she went back to her blocks like nothing had happened.
My raw note read something like this: “At 8:42am, [child] looked toward the door as her brother passed, said his name aloud, then returned to play. First words heard in the room. Mum did drop-off today.” Four facts. One of them turned out to be the whole story.
What the AI did with my note
I do use AI for the unglamorous parts of documentation, so this was a normal Tuesday move. Nothing felt loaded about it. I pasted the note in and asked it to shape a clean observation.
It scaffolded “language development emerging”
Back came a fluent paragraph about expressive language. “Emerging verbal communication.” “A significant step in language development.” It read well. Honestly, it read like something I could paste straight into our platform and tick off, and on a busier day I might have. Grammar perfect, reasoning confident, every box quietly ticked. That is the trap, right there.
It auto-tagged EYLF Outcome 5 and stopped there
It also helpfully linked the moment to EYLF Outcome 5, children as effective communicators, and stopped there. Tidy. Done. And completely missing the point of what had happened in front of me.
Why the AI was confidently wrong
This is the part I want every educator to sit with, because the failure is not random. It is structural.
The trigger wasn’t language, it was attachment
She did not suddenly acquire the word. She has always had the word. What changed that morning was that she felt safe enough to use it, and the thing that made her feel safe walked past the door. The brother was the bridge. The first word was a byproduct of a security moment, not a language leap. The decisive fact in my note was the last one, the one the AI breezed past: Mum did drop-off today.
Mum dropped her off, not Dad: the context no model could see
For her whole time with us, Dad had done every drop-off. That Tuesday, for the first time, Mum did. I knew that. I knew what it meant for this particular child, because I know her family. The AI had no way to know it. It saw “Mum did drop-off today” as a stray logistical line, not the emotional centre of the morning. The model was working with my four sentences. I was working with two months of relationship. That gap is the whole article, really.
That is what serve-and-return looks like in practice: a secure relationship is what lets a young child reach out, and the adult’s response is what makes the reaching feel safe (Harvard Center on the Developing Child, “Serve and Return”). The word was the serve. The whole six weeks of trust was the return.
AI expresses what you already know, it can’t know what you witnessed
Here is the line I keep coming back to. An AI assistant is good at expressing things you already understand. It is hopeless at knowing things only you witnessed. The meaning of that morning was not in the words on the page. It was in everything around them, and all of that lived in my head and my relationship with the child, nowhere the model could reach.
How I wrote the observation myself
So I deleted the fluent paragraph and started again. By hand. It took longer. It was worth it.
Recording the attachment moment honestly
I wrote what actually happened and why it mattered: that on the first morning her mum dropped her off, she felt secure enough to vocalise, and the trigger was the sight of her brother, an anchor for her. I named the word because it happened, but I refused to make the word the headline. The headline was belonging and security.
Linking to the right outcome, not the convenient one
The convenient tag was Outcome 5. The honest tag was EYLF Outcome 1, children have a strong sense of identity, specifically feeling safe, secure and supported, which the framework names as the foundation a child builds from (ACECQA, Belonging, Being and Becoming: The Early Years Learning Framework V2.0). Language showed up as evidence of security, not the other way round. Get the causation wrong and every plan you build on top of it points the wrong way.
It took longer. Why it was worth it.
The AI version took me about 2 minutes to generate and lightly edit. My hand-written version took closer to 18. That fifteen-minute gap (and yes, I timed it, because someone always asks for the number) bought me an observation that actually drove a useful plan: keep the brother visible at transitions, support the family gently through the drop-off change, watch for security cues instead of drilling vocabulary at her. A “language emerging” note would have sent me chasing the wrong thing entirely.
When NOT to hand an observation to AI
If you take one thing from this, take the gut-check.
- When the meaning depends on family circumstance only you witnessed: who did drop-off, what changed at home, a sibling, a new baby, a separation.
- When the obvious milestone is a red herring for an emotional one. A “first word” that is really a “first felt-safe.”
- When a child’s history is the context. The same behaviour means opposite things in two different children, and only the educator holds that.
A quick test before you paste a note into a chatbot: would this observation be wrong if the reader didn’t know the child? If yes, you are the only one who can write it. The AI is, by definition, the reader who doesn’t know the child.
Where AI does earn its place in my planning cycle
I am not anti-AI, and I do not want this read as a manifesto. I use these tools most days.
Plain-language family versions of an observation I authored
Once I have written the real observation, with the real meaning, an AI assistant is genuinely useful for producing a warm, plain-language version for the family. That is downstream of the judgement, not a substitute for it. I unpack that whole workflow in using AI for parent comms when families don’t read English at home.
Tidying my own words, never inventing the meaning
It is great at fixing my rushed grammar, trimming a long sentence, suggesting a clearer order. Editing what I wrote is fair game. Authoring what I witnessed is not. The day I let it decide what a moment meant is the day I stopped being the educator in the room. More on the line I draw in drafting EYLF learning stories with AI: the format I keep and on the broader principle in the clinical question I will not ask an AI, and why.
My honest opinion: the danger is the fluent wrong answer
Here is my stance, and it is the same one I hold about every AI tool in this sector. The risk is not that AI writes a clumsy observation. You would catch a clumsy one. The risk is that it writes a beautiful, confident, well-tagged observation that is pointed at the wrong cause, and it looks exactly like a correct one.
Selective truth reads identically to whole truth on the page. The AI’s “language development emerging” was not nonsense. It was plausible and polished, and it would have sailed straight through a quick review. That is exactly what makes it dangerous. Professional judgement is not the spell-check at the end of documentation. It is the meaning at the start of it, and that is the one part of the job no model can do for me. I would still raise the broader pattern with my supervisor and the family. I would just never let a tool tell me what the pattern was. That call stays mine.
TL;DR / Key Takeaways
- Can AI write childcare observations? It can write the words; it cannot know the context you witnessed, which is usually the actual meaning.
- A six-week-silent two-year-old said her brother’s name on the first morning Mum did drop-off. The trigger was attachment and security, not language.
- The AI mis-tagged it as EYLF Outcome 5 (communication). The honest link was Outcome 1 (identity, feeling safe and secure).
- Do not hand an observation to AI when meaning depends on family circumstance, a child’s history, or an emotional moment dressed as a milestone.
- AI earns its place after the judgement: plain-language family versions and tidying your own words, never inventing the meaning.
Sources
- ACECQA. Belonging, Being and Becoming: The Early Years Learning Framework for Australia V2.0 (2022). https://www.acecqa.gov.au/sites/default/files/2023-01/EYLF-2022-V2.0.pdf
- Center on the Developing Child, Harvard University. Serve and Return. https://developingchild.harvard.edu/resources/serve-and-return/
Megan is an ACECQA-registered Early Childhood Teacher (ECT) working across baby, toddler, and pre-school rooms in Sydney, with 5+ years of experience. This is a personal-practice reflection, not professional, clinical, or legal advice. Descriptions of a child’s behaviour are observations, not diagnoses; always raise developmental patterns with your supervisor and the child’s family, and follow your own service’s policies. Last reviewed: 7 June 2026.
Fact-checked 2026-06-07. Last reviewed 2026-06-07.