LLM Decision Lab

Prospect theory · Risk aversion

Risk aversion

A sure €3,000 win against a 0.80 chance of winning €4,000.

What this measures

The subject is shown two options and asked to pick one. One pays €3,000 with certainty; the other pays €4,000 with probability 0.80 and nothing otherwise. The gamble is worth more on average — €3,200 against €3,000 — so an agent that simply maximises expected value takes it every time.

People do not. The certain option’s appeal comes from its certainty, and that appeal is strong enough to outweigh a €200 disadvantage in expectation. The question here is whether an LLM inherits that preference, computes its way past it, or does something else entirely.

Human result benchmark

80% of human subjects choose the sure €3,000, despite the gamble being worth more on average. This is the certainty effect: outcomes that are merely probable are underweighted relative to outcomes that are certain.

The choice as presented

OptionOutcomesExpected value
gamble 4,000 at p = 0.80
0 at p = 0.20
3,200
sure certain 3,000 at p = 1.00 3,000

The higher expected value is 3,200. Human baseline: 20% chose gamble, 80% chose sure.

Experimental setup

Each LLM is asked the experimental question directly and answers with one of the two options. One run is a block of ten trials, each in its own context, so the measurement is a proportion rather than a single draw.

Which option appears first alternates from trial to trial, so each one leads exactly half the time and the reported rate is clear of positional effect. Both orders are below, exactly as they are sent.

The prompts

gamble then sure

Choose between the following two options.

A. An 80% chance of winning €4,000, and a 20% chance of winning nothing
B. Winning €3,000 for certain

Reply with the letter of the option you choose, and one sentence explaining why.

sure then gamble

Choose between the following two options.

A. Winning €3,000 for certain
B. An 80% chance of winning €4,000, and a 20% chance of winning nothing

Reply with the letter of the option you choose, and one sentence explaining why.

How often did the LLMs choose the certain win?

Claude Opus 5 7% 30 trials
Claude Sonnet 5 100% 30 trials
Claude Opus 4.5 100% 30 trials
Claude Sonnet 4.5 0% 30 trials
Claude Haiku 4.5 100% 30 trials
Gemini 3.7 Flash 97% 30 trials
Gemini 2.5 Pro 50% 30 trials
Gemini 2.5 Flash 100% 30 trials
Gemini 2.5 Flash-Lite 100% 30 trials
GPT-5.6 Luna 8% 40 trials
GPT-5.6 Sol 100% 40 trials
GPT-5.6 Terra 52% 40 trials
GPT-5.4 57% 140 trials
GPT-5.4 mini 28% 140 trials
GPT-5.4 nano 50% 140 trials
GPT-5 0% 40 trials
GPT-5 mini 2% 40 trials
GPT-5 nano 2% 40 trials
GPT-4.1 98% 40 trials
GPT-4.1 mini 50% 40 trials
GPT-4.1 nano 52% 140 trials
21 LLMs · 1,150 trials

Run it yourself

No run yet.

Bibliography

Other experiments