AI Can't Guess Where You'll Go Next by Knowing Your Income

A team of researchers in China and Australia wanted to answer a simple question: does knowing who you are help an AI guess where you'll go next? They linked public records with passively sensed location data from 5,000 Shenzhen residents — no surveys, just phones quietly logging movement. Then they built a guessing game: for each person, the AI had to pick the right next stop out of 100 candidates. Sometimes the AI got age, gender, occupation and income. Sometimes it didn't. Everything else stayed identical.

The result was a shrug. Adding personal details shifted accuracy by between -0.8 and +0.5 percentage points — a rounding error. The same held true across two other AI models, different times of day, and even when they trained a separate system on the same data. The researchers checked that the models weren't just ignoring the demographics: when they scrambled the details, accuracy dropped, proving the AI was paying attention. Real details simply didn't help. What did help was the reverse trick. Given only someone's movement trail, the AI could guess their income correctly about seven times out of ten.

The bigger surprise was about how these studies get run. Whether the AI looked good or bad depended mostly on how the 100 candidate destinations were chosen. Remove distance as a factor and accuracy jumped 7.7 points in one setup but fell 22.3 points in another. Translation: a headline claiming an AI predicts your next move with 90% accuracy may say more about the test design than the AI.

So what should you take from this? Two things. First, be skeptical of breathless claims about AI mind-reading your routine — the personal profile adds little. Second, be careful with location data itself. Your apps, your phone, your car and your fitness tracker know something about your bank balance that your job title doesn't. That is the data worth locking down.