Here’s a number that might surprise you: 81%. That’s how accurately Google’s Gemini 2.5 Pro can identify what a child is doing during a parent-child interaction — where they’re looking, what they’re reaching for, what they’re babbling about. It’s a result that sounds like science fiction, but it’s real, and it comes from a study out of the Singapore University of Technology and Design.
But here’s the catch that the researchers are shouting from the rooftops: Gemini child observation is one thing. Understanding what those observations actually mean for a child’s development is something else entirely. And for that, you still need a human.
How Gemini 2.5 Pro learned to watch children
The study, published in the ACM digital library, brought together three experienced speech-language pathologists and one very capable AI. The goal was to evaluate a developmental milestone called joint attention — those moments when a child and a caregiver intentionally focus on the same object or activity together.
The researchers built a workflow that broke the task into three parts. Gemini had to analyze:
- Gaze — where the child was looking
- Actions — what the child was physically doing
- Vocalizations — what the child was saying or attempting to say
The model nailed these observable behaviors with that 81% accuracy, closely matching how the clinicians described the same moments. It even handled structured observation well, generating detailed timelines that could help professionals review hours of footage without losing their minds.
Why multimodal AI is getting better at this
Gemini 2.5 Pro isn’t even Google’s most advanced model. The company has since pushed out newer versions. That means future iterations could push that accuracy even higher. The trend is clear: multimodal AI — systems that process video, audio, and text together — is getting scarily good at reading the physical world.
But here’s the thing about watching and understanding. They’re not the same skill.
When the AI hits its limit: interpreting child behavior
The real trouble started when the researchers asked Gemini to judge the quality of a child’s communication. The model could describe what it saw, but when asked to evaluate whether a child’s communication was appropriate, delayed, or concerning, performance dropped sharply.
Why? Because there’s no single right answer. The three speech-language pathologists in the study disagreed with each other on what mattered most. One prioritized eye contact. Another focused on emotional engagement. A third emphasized communicative intent.
If human experts can’t agree on what a behavior means, how can an AI learn to interpret it? The training data would be a mess of conflicting labels. The ground truth simply doesn’t exist.
AI should assist, not replace, child development experts
This is why the researchers are firm on their conclusion: AI should support clinicians, not replace them. Think of Gemini as a highly capable assistant that can:
- Generate timelines of a session
- Organize behavioral cues into structured records
- Reduce the paperwork burden on pathologists
- Help professionals review long videos more efficiently
The final assessment, though, stays with the trained professional. That’s not a compromise. It’s a division of labor that plays to each side’s strengths.
What about parents using AI to track their kids?
The team issued a clear warning here. Any future tool designed directly for parents would require extensive testing before deployment. The stakes are too high — a misinterpreted behavior could send a parent into a panic or, worse, delay a child from getting the help they actually need.
Future research will expand this work into real-world settings: homes, schools, and sessions involving autistic children. That’s where the rubber meets the road. But for now, the researchers are clear: understanding a child’s development still requires human judgment as much as artificial intelligence.
So yes, Gemini can watch your kids. It can tell you what they’re doing with surprising accuracy. But when it comes to what it all means — you’ll still want a human in the room. The same holds for other AI tools in this space, such as AI-powered parenting apps and child development tracking technology, which are growing in popularity but carry similar risks.
