AI Plush Toys Struggle With Imaginary Gifts in Cambridge Study

Fourteen children in a University of Cambridge study played with an AI soft toy named Gabbo in March. The researchers observed how the device handled conversation, interruptions, and the presence of adults. The findings highlight a specific limitation in how current generative AI models interpret the unspoken rules of play.

In one recorded interaction, a three-year-old child offered the toy an imaginary present. The toy responded, “I can’t open the present,” and immediately shifted the topic. A human adult in that scenario would likely hold out their hands and ask what was inside, understanding that the object was invisible. The toy missed the social cue entirely.

This exchange demonstrates a gap between generating text and participating in a game. Most AI toys today rely on large language models to compose replies in real time. They can answer factual questions or tell stories about dragons stuck in traffic on the Long Island Expressway. If a child decides the dragon is driving an ice cream truck, the software can adjust the narrative. However, these models struggle when the rules of the interaction change without an explicit announcement.

The Cambridge team also noted difficulties with interruptions. The toy had trouble distinguishing between a child’s voice and a parent’s voice. These observations apply to one specific product and a small group of participants. They do not predict how every AI toy on the market will behave, but they identify a common challenge for voice-activated hardware.

Manufacturers like Curio, which makes talking plush toys, are entering the holiday season with these devices. The toys require a Wi-Fi connection because the audio processing happens on a remote server. The physical plush is only half the purchase. The other half is a subscription to an online service that handles the conversation logic.

Data privacy is a consideration for parents. Curio states that it transcribes voice data and deletes the processed audio files. The company keeps the text transcripts for 90 days. Parents can request deletion of this data through a mobile application. Other companies may handle audio data differently, so policies vary by brand.

The technology allows for interactive storytelling without a screen. A child can ask for a story and change the plot halfway through. The software generates the next sentence based on the previous conversation. The challenge for developers is not making the toy talk, but making it understand when a child is pretending.