The Negotiation Trap: An Experiment on a Large Language Model

In an experiment on the large language model GPT-4o, a supplier always makes a higher profit if it replaces uniform contract terms with a set of terms between which the custom-er may choose. The extra profit results from price discrimination. There is a first order and a second order effect. The first order effect results from heterogeneous willingness to pay for a more protective term. The second order effect results from the possibility that con-tract choice is a signal for general willingness to pay for the traded commodity. In the ex-periment, the effect is bigger if the least protective version is labelled as the default, and more protective terms as an “upgrade†. The effect is smaller if, conversely, the most pro-tective version is labelled as the default and less protective (and cheaper) versions as an opportunity for “savings†. The effect is also bigger if the supplier only sets the price after it knows which version of the contract the consumer chooses. The profit increasing effect of giving the consumer a choice is strong. Most pieces of demographic information (which the supplier might, for instance, learn from cookie data) have a significantly smaller effect on profit. If the supplier combines cookie information about demographic markers with contract choice, it often even makes an extra profit. The main results replicate on Gemini 2.5 flash.