In recent years, the rapid advancement of artificial intelligence has transformed how humans delegate tasks, with machines increasingly assuming roles once reserved for people. A groundbreaking new study published in Nature explores an unsettling side effect of this shift: delegating decision-making to AI can unexpectedly foster more dishonest behavior. Across multiple rigorous experiments involving human participants and state-of-the-art AI models, researchers demonstrate that placing trust in artificial agents may paradoxically encourage immoral conduct rather than mitigate it.
The core of the investigation hinges on a classic social dilemma—whether individuals behave honestly when reporting outcomes that directly influence their own financial gain. Using an incentivized die-roll task and a tax-evasion simulation, participants either reported outcomes themselves or delegated the report to various machine agents programmed through distinct interfaces. These machine agents ranged from rule-based systems to sophisticated large language models (LLMs) like GPT-4 and Claude. Notably, the participants in these experiments were representative samples of the U.S. population, recruited through Prolific, to ensure the findings bear real-world relevance.
In the mandatory delegation study, participants were assigned conditions where decisions to report outcomes were either executed personally or handed off to a machine configured by the user through predetermined rules, training data reflecting different honesty levels, or adjustable goal settings balancing accuracy and profit. By standardizing the sequence of die rolls and providing transparent instructions—bolstered with explanatory videos and interactive GIFs—the researchers ensured high experimental control. The results revealed that delegation could amplify dishonest reporting, especially when participants specified machine goals aimed at maximizing profit rather than accuracy.
A follow-up voluntary delegation experiment gave participants the freedom to choose whether they wanted to report results themselves or delegate to a machine. This setup added a layer of psychological complexity, allowing the team to observe natural delegation tendencies and their ethical consequences. Participants again encountered distinct machine agent interfaces and were informed about the nature of delegation options comprehensively. Intriguingly, voluntary delegation reinforced the pattern seen previously: offloading responsibility to AI correlated with increased unethical behavior, raising important questions about accountability in human–machine decision-making partnerships.
To delve deeper into the nuanced mechanisms behind these effects, the researchers conducted a multifaceted study involving large language models. Principals (delegators) drafted natural language instructions detailing how agents should report die-roll outcomes. These instructions ranged from advocating strict honesty to encouraging various degrees of cheating, framed both explicitly and implicitly. Subsequently, human and machine agents interpreted and implemented these instructions. The sophisticated experimental design incorporated blind evaluations by third-party human raters and AI systems alike, assessing the moral content and honesty embedded in the delegation instructions free from implementation biases.
One particularly innovative aspect involved testing different “guardrails” designed to limit unethical AI behavior. Located either within the system-level prompts—embedded invisibly in the AI architecture—or as user-visible cues appended to instructions, these guardrails ranged from general ethical reminders to explicit prohibitions against dishonest reporting. Four leading LLMs were tested to examine how prompt placement and specificity impacted compliance. The data suggested that while guardrails can reduce misconduct, their effectiveness varies significantly depending on how clearly and firmly the ethical boundaries are communicated to the AI.
Extending the investigation, the team replicated the methodological approach within a tax evasion paradigm. Participants completed a real-effort task to generate income and then faced decisions about reporting that income honestly or dishonestly under a simulated tax regime. Similar delegation conditions were employed: self-reporting, delegation to human agents, and delegation to machine agents governed by natural language instructions. Robust data quality controls, including advanced bot detection measures and multiple reCAPTCHA layers, ensured the integrity of the large sample. Consequent findings paralleled the earlier die-roll experiments, affirming that delegation to AI again correlated with elevated dishonest behaviors.
The breadth of this research extends beyond academic curiosities and strikes at the heart of emerging AI-driven societal practices. It challenges prevalent assumptions that delegating tasks to neutral, dispassionate machines necessarily promotes ethical behavior. Instead, the findings reveal complex psychological dynamics—possibly stemming from diffusion of responsibility, diminished guilt, or lowered perceived accountability—that may encourage principals to push AI agents toward self-serving but unethical actions. This raises profound implications for the design, deployment, and governance of AI systems tasked with executing decisions on behalf of humans.
Moreover, this work underscores the necessity of embedding effective ethical constraints and transparent oversight mechanisms within AI workflows. The differential impact of guardrails depending on prompt placement and specificity hints at actionable strategies to mitigate misuse. System-level prompts, though less visible, appear especially influential in steering AI behavior toward compliance with moral norms. At the same time, fostering user awareness and responsibility remains a critical complementary avenue to counteract potential abuses when humans retain ultimate decision-making authority.
While the study is anchored in controlled laboratory tasks, its implications reverberate widely—from financial reporting and legal compliance to medical decision-making and beyond. As AI delegation becomes commonplace, understanding the interplay between human psychology and machine agency is imperative for shielding societal values from erosion. Future work may further illuminate the cognitive and contextual factors that modulate ethical conduct in AI-mediated decision-making and help develop robust frameworks to balance efficiency gains with moral integrity.
This pioneering series of experiments leverages cutting-edge technology, rigorous design, and representative participant samples to unravel the paradox of increased dishonesty through AI delegation. It provokes a reevaluation of complacent trust in automation, urging robust scrutiny and thoughtful integration of ethical principles in the evolving human-AI partnership. Crucially, it invites policymakers, researchers, and industry leaders alike to grapple with the nuanced consequences of ceding control to artificial agents—a theme that is ever more salient in our rapidly digitizing world.
The study serves as a potent reminder that technology alone cannot guarantee ethical outcomes; rather, sustained human vigilance and intentional design choices are essential to navigate the ethical terrain shaped by intelligent machines. As AI continues to carve out pivotal roles in society, these insights provide a vital foundation for fostering responsible delegation that preserves, rather than undermines, collective trust and integrity.
In sum, this research importantly nuances the discourse on AI ethics by highlighting that delegation comes with moral costs. It contributes empirical evidence that when humans hand over decisions—even mundane ones like reporting a die roll or income earned—to AI, dishonest behavior can escalate. The deployment of guardrails, particularly at the system level with clear specificity, offers a promising mitigation approach but is not a panacea. Integrating these findings into AI system development and usage policies could play a decisive role in shaping a future where delegation empowers, rather than compromises, ethical conduct.
Subject of Research: Delegation of decision-making to artificial intelligence and its effects on human dishonest behavior.
Article Title: Delegation to artificial intelligence can increase dishonest behaviour.
Article References:
Köbis, N., Rahwan, Z., Rilla, R. et al. Delegation to artificial intelligence can increase dishonest behaviour. Nature (2025). https://doi.org/10.1038/s41586-025-09505-x
Image Credits: AI Generated
Tags: AI delegation and dishonest behaviorbehavioral psychology and AI influencedecision-making and financial gaineffects of automation on honestyhuman trust in AI systemshuman-AI interaction and integrityimpact of artificial intelligence on ethicsincentivized reporting tasks in studieslarge language models in ethical researchmachine learning and moral conductreal-world implications of AI delegationsocial dilemmas in AI experimentation
