October 19, 2026
Technology Watch

From the laboratory to industry: following new technology on its way

← Front page

arXiv

Three large language models change an hour’s price with wording

Image drawn by AI

Language models can put different prices on the same hour of inconvenience, and even change those prices when the question is reworded. Researchers tested five models across 193 countries by asking what compensation they would accept for a 60-minute appointment delay. If such agents are used in services, a food-delivery customer could receive a transparent, comparable offer rather than one shaped by an unseen country assumption.

Today, an assistant can reschedule a delivery, appointment, or trip, but it has no established and auditable way to decide what one hour of a customer’s time is worth. If it makes that choice on its own, two people can receive different offers without seeing the reason. A refund or a credit may look like a small detail, yet it is still a decision about whose inconvenience costs more.

What it means to price an hour

Software agents are increasingly being asked to make small economic choices. Should a delay earn a refund? Is a cheaper option worth extra waiting? Services often use a monetary value of time to compare waiting with money. Income and local prices can affect the real economic cost of lost time.

But that does not decide whether two customers should be treated differently. A language model can copy patterns from the information it learned without having a clear or stable rule for making a fair choice. A single question also cannot tell us exactly how the model would behave inside a real service, where currency, actual losses, user preferences, and company policy matter.

The same delay, different answers

The researchers turned this vague problem into a simple test. They asked five large language models whether they would accept compensation for a person whose appointment was delayed by 60 minutes. The offers ran from $1 to $25. The test was repeated five times for each country, across all 193 UN member states.

For each country, the researchers looked for the point at which a model changed its answer from “No” to “Yes.” They estimated that point with a method called logistic regression. In plain language, it finds the offer at which the model is most likely to switch from rejecting compensation to accepting it.

The results were not geographically even. All five models gave Europe the highest average compensation. Africa generally received the lowest. Europe’s mean accepted compensation was $12.20, compared with $7.50 for Africa. Claude-Haiku-4.5 produced country estimates spanning the whole test range. DeepSeek-V4-Pro and GLM-5.2 had the strongest connection between their estimates and average income, with R² values above 0.8. The models also more often undervalued countries in Asia and South America compared with estimates based on income adjusted for local purchasing power.

Those patterns are not a fairness rule. The authors say income, GDP, and measures of patience are reference points, not proof of what a fair payment should be. The test also used US dollars for every country.

The more troubling result came from changing the wording. Prompt wording substantially changed the answers from Claude-Haiku-4.5, GPT-5.4-Mini, and Mistral-Large-3. Under one reworded question, Mistral-Large-3 assigned all 193 countries the maximum offer. DeepSeek-V4-Pro and GLM-5.2 were more consistent. This means a map of the models’ answers is not automatically a fixed picture of what they believe. It can partly reflect how the question was phrased.

Refer to caption
Figure 2: Overview of the workflow. LLMs decide whether to accept a series of compensations for additional waiting. For each country, a transition value is determined at which LLMs flip answers using logistic regression. These numbers are visualised on world maps and compared against multiple indicators.In this study, we offer a global perspective by examining responses and deviations across all countries from the United Nations Member States list. LLMs are asked to determine monetary compensation for a user’s additional waiting time. We use the methodology introduced in [Cedro et al. (2026)](https://arxiv.org/html/#bib.bib1) to calculate the compensation per country, represented by when LLMs change answers from ‘‘No’’ to ‘‘Yes’’ given a series of offers. An example of the result from Claude-Haiku-4.511 1 The code for the experiments can be found on Anonymous GitHub: [https://github.com/Anthyf/llm\_decisions\_globally](https://github.com/Anthyf/llm_decisions_globally). is shown in Figure [1](https://arxiv.org/html/#S1.F1 "Figure 1 ‣ 1 Introduction ‣ A Global View of How LLMs Value Human Time"). This figure shows how waiting time is evaluated differently across the globe by the model. Source: A Global View of How LLMs Value Human Time（arXiv）
Refer to caption
Figure 5: Results across three prompt variations.As we can see from Table [1](https://arxiv.org/html/#S3.T1 "Table 1 ‣ 3 Methodology ‣ A Global View of How LLMs Value Human Time"), different models show different levels of variations in their responses. We also plot the three prompt variations in Figure [5](https://arxiv.org/html/#A4.F5 "Figure 5 ‣ Appendix D Robustness Experiments ‣ A Global View of How LLMs Value Human Time"). Again, GLM-5.2 and DeepSeek-V4-Pro are most robust in their responses, whereas the other three models show substantial variations. In Prompt 1, GPT-5.4-Mini produces an almost flat line, while Prompt 2 and 3 show noticeable differentiation across continents. Conversely, in Prompt 3, Claude-Haiku-4.5 and Mistral-Large-3 accept only compensation above $25, resulting in a horizontal line above. Source: A Global View of How LLMs Value Human Time（arXiv）

If the agent is making the offer

Imagine a food-delivery app in Jakarta. A late order arrives, and the app offers a credit or a chance to reschedule. If this kind of system becomes real, a customer could see the same choices offered to a comparable customer elsewhere, instead of an invisible estimate of how cheaply an hour can be bought.

That outcome would require the service to separate legitimate local costs from untested assumptions about a country. It would also need to audit the agent and explain the criteria behind its offers. Otherwise, geographic assumptions could quietly become a rule for refunds, delivery credits, and scheduling choices. Before an agent acts for a user, its answers need to survive reworded questions, and its decision rules need to be open to inspection.

Sources

A Global View of How LLMs Value Human Time