Gemini can help interpret the world you live in, but it's best not to let it keep an eye on your children just yet.

      AI can recognize a child's actions, but interpreting their significance still requires a human expert.

      Google's Gemini models are becoming increasingly proficient at analyzing videos, images, and conversations. A recent study indicates that AI can even detect subtle behaviors in parent-child interactions with notable accuracy. However, the downside is that, despite its ability to observe well, researchers caution against relying on Gemini to determine the meaning of those behaviors.

      It's important to mention that the study utilized Gemini 2.5 Pro, which is not Google's most cutting-edge AI. Future iterations could enhance these results further. Nevertheless, the researchers emphasize the continued necessity of human experts.

      How Gemini excelled in observing children

      How multimodal LLMs can be aligned with Speech-Language Pathologists in parent-child interactions

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

      With Gemini 2.5 Pro, the team developed a process where the AI analyzed three behavior indicators: the direction of the child's gaze, their actions, and what they said. The model accurately identified these observable behaviors with approximately 81% accuracy, closely aligning with how expert clinicians described the same situations. The researchers also observed that Gemini excelled in structured observation, producing detailed logs that could aid clinicians in reviewing lengthy videos more effectively.

      Why understanding children still necessitates human discernment

      The real difficulty arose when the AI was tasked with interpreting the significance of those behaviors. While the model was adept at describing gaze, actions, and vocalizations, its accuracy declined significantly when asked to assess the quality of a child's communication.

      Interestingly, even the experts had differing opinions. One expert emphasized eye contact, another focused on emotional involvement, while a third highlighted communicative intent. This variability meant there was seldom a single correct interpretation for the AI to learn from.

      Consequently, the researchers contend that AI should assist clinicians rather than replace them. It could create timelines, organize behavioral cues, and minimize paperwork, all while trained professionals make the final judgments.

      The team also warned that any future tools aimed directly at parents would require extensive testing before being implemented. Upcoming research will further explore the work in actual home and school environments and will involve autistic children. However, for now, both human judgment and artificial intelligence are essential for understanding a child's development.