When a nudge backfires

At Berkeley's Operations & Behavioral Analytics Lab, I've spent the last stretch on a question that sounds simple and isn't: when you make the reward more obvious, do people make better choices?

The setup is a repeated decision task — a bundling problem people play round after round, 85 participants over more than a thousand rounds total. Half the time we made the payoff of each option more salient: bigger, clearer, harder to ignore. The intuition almost everyone shares is that this should help. Show people the stakes and they'll act on them.

It did the opposite. Under a fixed-effects model, making the payout salient raised the rate of suboptimal choices by 71 percentage points. And people didn't grow out of it — across fifteen rounds there was no measurable learning. The nudge didn't teach; it distracted.

What I find interesting is that "people" is doing a lot of work in that sentence. When we clustered the behavior, it didn't collapse into one average human — it separated into five distinct decision policies, with an eight-fold spread in outcomes between them. Aggregate effects hide the fact that a salient reward helps some strategies and wrecks others.

I came into this wanting to build the reinforcement-learning side — the DQN, the reward shaping, the offline evaluation. I'm leaving it more convinced that the modeling is the easy half. The hard half is being honest about whose behavior your model is actually describing, and resisting the tidy story where one intervention moves everyone the same way.

The reward function you design for an agent and the incentive you design for a person turn out to have the same failure mode: optimize the number in front of you, and you can miss the thing you actually wanted.