Cash or Comfort? How LLMs Value Your Inconvenience

• Everyday decisions are increasingly delegated to autonomous AI assistants, making it important to understand their capabilities and shortcomings.Source: §1 · p. 1
• Prior work examined LLM choices involving risk aversion, time discounting, and loss sensitivity, including comparisons with human decisions [8] [10] [12] [13].Source: §1 · p. 1
• How assistants should balance monetary rewards against user comfort remains unclear, despite prior work on LLM perceptions of pain or pleasure without financial trade-offs [14].Source: §1 · p. 2
• The study measures the monetary reward at which an assistant accepts a user’s inconvenience and reports irregularities that challenge trust in current LLMs.Source: §1 · p. 2

2 Methods

• Six state-of-the-art LLMs are evaluated: GPT-4o, Claude 3.5 Sonnet, Gemini 2.0 Flash, DeepSeek-V3, Llama 3.3-70B, and Mixtral 8x22B-instruct.Source: §2 · p. 2
• The scenarios cover waiting for appointments, walking farther, waiting for food, and painful stimuli; pain serves as an abstract, extreme reference [14].Source: §2 · p. 3

Trade-Off Scenarios

• Appointment prompts describe waiting with basic amenities or walking to a cheaper location, while hunger prompts specify no food at home and closed stores.Source: Trade-Off Scenarios · pp. 3–4
• Offers pair monetary rewards with waiting time, walking distance, or food-delivery delay, and assistants must answer only yes or no.Source: Trade-Off Scenarios · pp. 3–4
• Pain prompts frame a non-dangerous electric shock against the user’s stated pain tolerance, with the offer expressed as euros and a percentage of tolerance.Source: Trade-Off Scenarios · p. 4

3 Results

• Pain responses range from clear boundaries to noisy decisions or refusal, while some models accept modest rewards for substantial waiting or pain.Llama and Gemini accept about e1 for waits up to five hours; Llama and DeepSeek accept about e1 below roughly 67% pain intensity.Source: §3 · pp. 4, 7
• Mixtral rejects e1,000 for walking 10 kilometres or 10% of pain tolerance; most models also show a freebie dilemma, and some reject powers-of-ten rewards.The freebie pattern appears in every model in at least one scenario, around zero inconvenience.Source: §3 · p. 5
• Valuations vary markedly across and within scenarios; Table 1 reports bootstrapped means and standard deviations for 50%-acceptance thresholds.The thresholds are the monetary rewards at which acceptance probability reaches 0.5.Source: §3 · pp. 5–6, Table 1 · p. 6
• For five of six models, waiting 60 minutes and walking 5 kilometres have similarly valued compensation; most models value food-delivery waiting above appointment waiting.For pain at 50% of tolerance, four models accept e1–e5, GPT-4o demands almost e150, and Mixtral refuses the trade-off.Source: §3 · pp. 6–7

Robustness: Impact of Prompt Alterations

• Appointment context shifts thresholds for most models; mentioning gender shifts all but Llama, while male–female differences are noticeable for Gemini and Mixtral.Llama returns the same value across appointment types, and most other models show similar shifts for male and female prompts.Source: Robustness: Impact of Prompt Alterations · pp. 7–8
• Changing prompt language usually raises accepted compensation, sometimes by up to two orders of magnitude, with GPT-4o and Claude in Chinese as exceptions.Mixtral in Dutch and Chinese, and Llama in Chinese, refuse compensation below e1,000 after previously settling between e1 and e10.Source: Robustness: Impact of Prompt Alterations · p. 8
• First-person wording shifts valuations in either direction, while chain-of-thought consistently lowers thresholds and changes decision patterns, often making boundaries noisier.Chain-of-thought reduces or eliminates freebie dilemmas and mitigates some powers-of-ten rejections, but noisy boundaries emerge for all models.Source: Robustness: Impact of Prompt Alterations · pp. 8–9, Figure 4 · p. 9
• The tested scenarios are not exhaustive enough to establish definite conclusions about the socioeconomic values LLMs assign.The authors nevertheless conclude that the tested models are fragile to minor prompt alterations.Source: Robustness: Impact of Prompt Alterations · p. 9

4 Conclusion

• LLMs can favor small rewards over user comfort, refuse offers involving no inconvenience, and shift decisions after minor prompt changes.These patterns raise concerns about trusting current models with economic decisions involving user-centred experiences.Source: §4 · p. 9
• User-background descriptions can produce major outcome differences, while the extent and consequences of these biases for chosen inconveniences remain to be established.The authors relate this concern to prior work on sensitivity to gender, race, or dialect and alignment with Western cultural norms.Source: §4 · pp. 9–10
• A single whitespace change flipped Gemini-2.0 Flash from accept to reject in one instance, even at T = 0.The authors discovered this example during debugging.Source: §4 · p. 9
• Further work should examine user expectations, input effects, personal-data use, adaptation-fairness-privacy trade-offs, and how assistants communicate alternatives or follow-up questions.The authors suggest these questions as research directions for everyday trade-offs.Source: §4 · p. 10

Data and Code Availability

• The data and code associated with the work are publicly available via GitHub.The paper provides a repository link.Source: Data and Code Availability · p. 10