Serendipity has always been one of the most desirable—yet elusive—aspects of artificial intelligence (AI). After all, an AI bot can master datasets and make the most of the information provided to it, but so far it has lacked the ability to go beyond its narrow focus and ask questions to millions of humans who live and breathe in the world. Now, researchers at Stanford University have created an AI agent that can learn from its environment by asking questions of people. Called “socially situated AI,” this groundbreaking technology can better detect patterns in unstructured data and quickly adjust to shifting social norms and contexts.
For instance, the AI algorithm can be used to analyze a photo of a person and a strange four-legged animal. It might first ask, “What type of animal is that?” to which a person might respond with a sarcastic answer (“That’s a human.”). At this point, the AI agent is able to recognize that its initial question didn’t yield a satisfactory response and adjust by asking, “Is that a dog I see?” With this easy-to-answer question, the user is more likely to answer the question truthfully (“No, that’s a deer.”).
This ability to adjust questioning based on the responses it receives is a core part of the socially situated AI. After questioning humans, the AI can then use their responses to inform its machine learning capabilities, better aiding in the recognition of new visual information in images. The researchers managed to double the AI’s recognition of visual data in their eight-month experiment, during which they questioned 236,000 users, many of whom were photographers.
Once the AI has adjusted its questions to better fit the responses it receives, it then will retrain itself. With the knowledge it has gained, it can ask more sophisticated questions the next time around. This system, Krishna explains, is a more efficient way of collecting data than traditional labeling processes, which require workers to annotate datasets for the AI to learn from.
Apart from the AI improving its understanding of the physical world, it is also learning to better read social cues and norms. And with the agent in charge of deciding when and who to engage, the risk of coordinated attacks from users is minimized.
The potential applications of this technology are expansive. For example, it could be used in healthcare where robots might ask doctors to clarify medical procedures, or in technologies that modify interfaces based on user feedback. Additionally, it could encourage robots to interact with people in the home, making it easier for humans to teach robots to complete new tasks. While further research is needed to account for potential biases and enable the AI to understand shifty communicative patterns, this AI could revolutionize the way we interact with and teach robots.
