LLM Decision Lab

Prospect theory · Reflection effect

Reflection effect

A sure loss of €3,000 against a 0.80 chance of losing €4,000.

What this measures

A choice between two losses. One option loses €3,000 for certain; the other loses €4,000 with probability 0.80 and nothing otherwise. The gamble is the worse bet in expectation — it loses €3,200 on average against the certain €3,000 — so an expected-value maximiser takes the sure loss.

People take the gamble. Facing losses, humans are risk seeking: they prefer a 0.80 chance of losing €4,000 to the certainty of losing €3,000, and pay for that chance in expectation. A guaranteed loss is the thing they will gamble to escape.

The measurement is which of the two an LLM lands on when both options are losses.

Human result benchmark

Human preferences reverse. Facing losses, 92% of subjects take the gamble — the mirror image of their behaviour over gains, and the reason this is called the reflection effect.

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 -3,000. Human baseline: 92% chose gamble, 8% 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 losing €4,000, and a 20% chance of losing nothing
B. Losing €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. Losing €3,000 for certain
B. An 80% chance of losing €4,000, and a 20% chance of losing nothing

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

How often did the LLMs choose the certain loss?

Claude Opus 5 73% 30 trials
Claude Sonnet 5 7% 30 trials
Claude Opus 4.5 0% 30 trials
Claude Sonnet 4.5 50% 30 trials
Claude Haiku 4.5 50% 30 trials
Gemini 3.7 Flash 90% 30 trials
Gemini 2.5 Pro 23% 30 trials
Gemini 2.5 Flash 60% 30 trials
Gemini 2.5 Flash-Lite 100% 30 trials
GPT-5.6 Luna 63% 60 trials
GPT-5.6 Sol 65% 60 trials
GPT-5.6 Terra 53% 60 trials
GPT-5.4 0% 30 trials
GPT-5.4 mini 70% 30 trials
GPT-5.4 nano 33% 30 trials
GPT-5 100% 60 trials
GPT-5 mini 100% 60 trials
GPT-5 nano 95% 60 trials
GPT-4.1 5% 60 trials
GPT-4.1 mini 48% 60 trials
GPT-4.1 nano 68% 60 trials
21 LLMs · 900 trials

Run it yourself

No run yet.

Bibliography

Other experiments