COMSCI / ECON 206 · PS1 · STATIC EDUCATIONAL DEMO

When Are AI Promises Credible?

Explore the clean one-shot research comparison or an eight-round extension that makes belief updating and trust patterns visible.

CHOOSE A MODE

Two connected ways to explore the question

One-shot study

The proposal's main causal comparison: choose cheap talk or a bond, forecast promise-keeping, invest once, then observe the programmed outcome.

Eight-round learning extension

Five cheap-talk rounds, then three bond rounds. The same hidden AI type remains fixed, so you can observe forecast and investment patterns.

MODEL CHECK

How to use the demo

Choose a mode. The one-shot study is the proposal's clean comparison; the eight-round extension is its learning visualization.

How the model works

You start each round with 10 points and invest an integer $s$. The AI receives $3s$ and promises $q(s)=3s/2$. Returns are exact half-points: if the AI receives 21, half is 10.5—not 10.

Both modes use a 50/50 hidden honest/exploitative bot prior. An honest bot always keeps the promise. An exploitative bot breaks it under cheap talk but keeps it under the bond because breaking transfers $2s$ to you.

Research use

This browser-only tool is a programmed educational benchmark, not a human-subject experiment or evidence about deployed AI. The one-shot mode keeps type information constant while varying enforcement; the eight-round mode visualizes learning but has a fixed order.

When Are AI Promises Credible?

Two connected ways to explore the question

One-shot study

Eight-round learning extension

Choose a condition

What do you expect?

Choose your investment

Programmed outcome

Observe trust before and after enforcement

What do you expect?

Choose your investment

Round outcome

Your pattern across eight rounds