Gemini is capable of understanding the world around you, but it's best not to let it watch your children just yet.

      AI can recognize what a child is doing, but interpreting its significance still requires a human expert.

      Google’s Gemini models have demonstrated impressive capabilities in understanding videos, images, and conversations. A recent study indicates that AI can accurately identify nuanced behaviors in parent-child interactions. However, the researchers emphasize that although Gemini can effectively observe actions, it should not be relied upon to interpret the meaning of those behaviors.

      It is important to note that the study utilized Gemini 2.5 Pro, which is not Google’s most advanced AI. This suggests that forthcoming models may yield even better results. Nonetheless, the researchers assert that human expertise remains vital.

      Gemini's impressive observational skills regarding children

      Researchers from the Singapore University of Technology and Design collaborated with three seasoned speech-language pathologists to assess parent-child interactions, centering on a developmental milestone known as joint attention. This concept pertains to instances when a child and another individual deliberately share focus on the same object or activity.

      Using Gemini 2.5 Pro, the research team developed a method that instructed the AI to analyze three behavioral indicators: the direction of a child’s gaze, their actions, and their vocalizations. The model accurately identified these observable behaviors with approximately 81% accuracy, aligning closely with the descriptions provided by expert clinicians. Additionally, the researchers highlighted that Gemini performed well in structured observations, producing detailed records that could aid clinicians in reviewing lengthy videos more effectively.

      Why is human judgment still essential for understanding children?

      The real difficulty arose when the AI was tasked with interpreting the implications of those behaviors. Although the model could describe gaze, actions, and speech, its performance significantly declined when it was required to assess the quality of a child’s communication.

      Interestingly, the experts themselves often had differing opinions. One expert emphasized eye contact, another prioritized emotional connection, while a third focused on communicative intent. Consequently, there was often no singular correct interpretation for the AI to learn from.

      This underscores the researchers’ belief that AI should assist clinicians rather than replace them. AI could help create timelines, organize behavioral indicators, and minimize documentation, while trained professionals would make the ultimate assessments.

      The team also warned that any future tools specifically designed for parents would necessitate much more testing before being implemented. Future studies will broaden the research to include real-world home and school environments and involve autistic children. For now, comprehending a child's development still relies heavily on human judgment alongside artificial intelligence.