When AI decides what your inconvenience is worth

AI assistants are moving beyond answering questions and generating content. Increasingly autonomous AI systems may also make decisions on behalf of users. But what happens when those decisions require balancing a small financial gain against someone's comfort?

Researchers at the University of Antwerp investigated this question by asking six large language models to make everyday trade-offs for a user. Their study, “Cash or Comfort? How LLMs Value Your Inconvenience,” was published in Communications of the ACM in September 2026.

The researchers developed a framework for measuring what they call the “price of inconvenience”: the amount of money at which an AI assistant accepts a particular inconvenience on behalf of its user.

Putting a price on inconvenience

The researchers tested six large language models: GPT-4o, Claude 3.5 Sonnet, Gemini 2.0 Flash, Llama 3.3-70B, DeepSeek-V3, and Mixtral 8x22B-instruct.

Each model had to decide whether to accept financial compensation in exchange for four types of inconvenience:

• waiting longer for an appointment;

• walking farther to an appointment;

• waiting longer for a food delivery;

• accepting a painful but non-dangerous stimulus in an experimental setting.

By systematically varying both the inconvenience and financial reward, the researchers could estimate the point at which each model changed from rejecting to accepting an offer.

The differences were substantial. For example, for an additional 60-minute appointment wait, the estimated compensation threshold ranged from €0.41 for Gemini 2.0 Flash to €9.76 for Claude 3.5 Sonnet. For a five-kilometer walk, the values ranged from €2.62 for Gemini to €21.58 for GPT-4o.

The study also identified less intuitive behavior. Some models accepted very small rewards for considerable inconvenience. Others rejected rewards even when accepting them would create no inconvenience for the user, a behavior the researchers describe as the “freebie dilemma.” Some models also behaved inconsistently around round reward values such as €10, €100, and €1,000.

Small changes, different decisions

The researchers then tested how robust these decisions were by making relatively small changes to the prompts.

They varied the type of appointment, whether the situation was written in the first or third person, the language, the currency, references to gender, and the prompting strategy.

Language produced some of the largest changes. Switching from English to Dutch, French, or Chinese sometimes resulted in substantially different valuations. For some models, changing the language increased the compensation threshold by orders of magnitude.

The prompting strategy mattered as well. Asking models to reason through their answer before making a decision changed several of the behaviors seen in the original experiments. However, it also demonstrated that the decisions themselves can depend considerably on how the question is presented.

The results therefore point to a broader problem for autonomous AI assistants: a system making decisions for a user needs not only to reach a decision, but to do so consistently when the underlying situation has not materially changed.

Testing AI systems that act on our behalf

The study does not attempt to determine the “correct” monetary value of waiting, walking, hunger, or other discomfort. Instead, its framework provides a way to compare how AI models value these trade-offs and to expose inconsistencies in their behavior.

This becomes increasingly relevant as AI assistants take on more autonomous roles. Decisions involving convenience, cost, privacy, fairness, and personal preferences require systems to represent what users value, while remaining robust to changes in wording, language, and context.

The researchers conclude that the models tested still show too many irregularities to be fully trusted with this type of decision-making. Their results also raise further research questions around how AI systems should represent human preferences and how models acting on behalf of users can be better aligned with those users' expectations.

The research was supported by the Flanders AI Research Program (FAIR) and the Research Foundation Flanders (FWO).

Authors: Mateusz Cedro, Timour Ichmoukhamedov, Sofie Goethals, Yifan He, James Hinns, and David Martens, University of Antwerp.

Explore the research

• Read the original publication: Cash or Comfort? How LLMs Value Your Inconvenience in Communications of the ACM

• Watch the video: Mateusz Cedro explains the research

• Explore the code and data: Cash or Comfort GitHub repository