Crosstalk blog

Social AI in the wild

Real examples of AI that works with people, in conversations, presentations, physical spaces and social settings.

Featured · All types

A field guide to the four types of Social AI

Social AI is already working in mediation tools, voice agents, hospitals and city streets. Here is one map of where, organized by the kind of human interaction each system has to handle.

2 min read
Social settings

The AI that helped thousands of people find common ground

Google DeepMind built an AI mediator and tested it with more than 5,000 people. Its real lesson is about what group conversation demands from machines.

2 min read
Conversation

Why your voice assistant still interrupts you

Turn-taking feels effortless to people. For machines, it is one of the hardest problems in conversation, and a new generation of models is finally tackling it.

2 min read
Conversation

Teaching avatars to listen

Good listeners nod, smile and hold your gaze. Meta's Seamless Interaction project shows what it takes to teach a machine to do the same.

2 min read
Social settings

What New York learned from giving older adults AI companions

A state agency put AI companion robots in the homes of hundreds of older New Yorkers. The results point to both the promise and the open questions of Social AI.

2 min read
Physical AI

Moxi walks the halls

A hospital is one of the busiest social environments there is. Diligent Robotics' Moxi shows why robots that work there need social skills as much as mechanical ones.

2 min read
Physical AI

Humanoids are clocking in. Are they ready for coworkers?

The first commercial humanoid robot jobs are here. The most revealing detail is how far they are kept from the people around them.

2 min read
Physical AI

The pedestrian problem

We rarely think of driving as a social activity. But every crosswalk is a negotiation between strangers, and robotaxis are learning to take part in it.

2 min read
Conversation

Meeting assistants and the consent problem

AI notetakers are recording more meetings than ever. The lawsuits now following them show why consent is becoming the foundation of Social AI.

2 min read
Conversation

What 850 hours of strangers talking can teach machines

The CANDOR corpus paired more than 1,400 strangers for unscripted video chats. It is a blueprint for how Social AI data should be collected, and a reminder of how little of it exists.

2 min read