Empathy beats engineering

New MIT research ran a simple experiment with 667 people:

First, answer questions alone (math, physics, moral reasoning).
Then, answer similar questions with AI assistance.

Everyone got the same AI.
But got huge differences in their performance boost.
Some people’s accuracy jumped from 55% to 85%.
Others barely moved.

What separated them wasn't technical skill.
It was “Theory of Mind”—the ability to read what the AI understands, catch when it's confused, and adapt your approach question by question. The researchers found ToM predicted the size of your AI boost but had zero correlation with solo performance.

Being smart alone ≠ being good with AI.

Like teaching your mom how to fix her wifi over the phone. You can't see what she sees. You have to build a mental model of what she's looking at, track when she's confused ("Wait, which button?"), and adjust your instructions in real time.

ToM even varied within the same person. When someone focused more on the model’s needs (more clarification, more rapport-building), AI response quality jumped. When they didn’t, performance tanked.

If you're teaching people how to use LLMs:
Maybe we need to stop teaching prompt templates. Start teaching collaboration. Build people's ability to form mental models of what the AI knows, where it'll fail, and what it needs to hear next.

The people who'll win with AI aren't the most technical.
They're the most empathic.


—
Source: Quantifying Human-AI Synergy, Riedl & Weidmann, 2025.