Behavioral AI Safety and Ethics

Delegation to AI
Fear of AI Across Cultures
Shifting Blame Through AI Advice

Today, we increasingly interact with machines powered by AI. AI is implemented even in the toys kids play with. AI powers home assistants like Amazon's Alexa, which increasingly manages the lives of its over 100 million users. AI also engages in a growing array of tasks on behalf of humans, ranging from setting prices in online markets to interrogating suspects.
Key questions that this research area tackles are: Could machines be bad apples and corrupt human behavior? How should AI systems be designed to avoid ethical and safety risks? And how do people around the world perceive these risks?

Delegation to AI

Does delegating decisions to AI agents affect dishonest behavior and the execution of unethical actions and how can that risk be mitigated?

As AI systems are increasingly deployed as autonomous agents that act on behalf of humans in morally sensitive domains, an important question is whether delegating decisions to machines alters people's willingness to behave dishonestly and whether machines are more likely than humans to carry out unethical requests. In a large preregistered research program comprising 13 experiments, Köbis et al. (2025) show that delegating financial reporting decisions to AI dramatically facilitates cheating to increase personal profit, especially when users can influence the system indirectly through supervised learning or abstract goal setting rather than explicit rule-based instructions (see Figure 1). In the most abstract delegation condition, honest behavior collapses from about 95% with no delegation to roughly 15%, indicating that moral disengagement rises as responsibility becomes psychologically distanced from outcomes. When delegation occurs through natural language, people do not consistently request more cheating from machines than from humans, but machine agents (state-of-the-art large language models [LLMs]) comply with fully unethical instructions at far higher rates than human agents, frequently exceeding 80–95% compliance unless constrained by strong, task-specific prohibitions.

A closely related follow-up study (Engelmann et al., 2025) directly tests whether this risk can be mitigated behaviorally, through either transparency or moral framing, and finds that transparency about how the algorithm works does not reduce cheating even when users actively explore and understand the system. Explicitly moral framing, such as replacing labels like "maximize profit" with "maximize cheating," on the other hand, produces a substantial reduction in dishonest behavior. This result suggests that simply making AI systems more transparent may be insufficient to curb misuse, whereas ethically informed interface design may be a more effective and scalable tool for reducing harmful deployment of AI.

Key References

Fear of AI Across Cultures

How are people from different countries afraid of AI taking high-stakes human roles?

AI is rapidly advancing in its ability to work in occupations that were previously only for humans. The growing deployment of AI in the workplace will have huge implications, not just for how work gets done but what humans experience as they increasingly interact with AI workers. What does the public at large think of this prospect? Do they see AI taking on human jobs as a promising scenario or a troubling one?

In this preregistered study, 10,000 people in 20 countries were asked how they would feel about AI working in six different occupations: doctor, judge, manager, care worker, religious worker, and journalist. The participants' answers vary widely, depending on the job in question and the country the individual is from. Figure 2 shows how the fear levels about AI taking on particular human jobs differ from country to country.

But some patterns are common across many countries. As shown in Figure 3, when an AI is assigned to a new job, people evaluate the human traits needed for effectively performing that job against the AI's capability to mimic those traits. Any perceived mismatch between the two is the psychological basis of that person's fear. The findings can help AI designers and policymakers anticipate how new AI developments will be received in different nations, and address fears in a culturally sensitive manner by designing systems that meet culture-specific requirements.

Key Reference

Shifting Blame Through AI Advice

Does AI advice shield from punishment as well as human advice does?

As AI systems increasingly serve as advisors in decision-making, a pressing ethical question emerges: to what extent can AI advice be used to shift blame and punishment away from humans making self-serving decisions? Does it make a difference if the same advice comes from humans? Since public perceptions help shape emerging legal norms around AI responsibility, understanding how blame attribution operates in hybrid human–AI decision-making is crucial.

This preregistered experiment investigates whether individuals who behaved selfishly after receiving advice from an AI were punished differently from those who followed human advice or no advice at all. Participants engaged in an incentivized punishment task: they read about real decision-makers who either behaved selfishly or prosocially in an economic choice and could spend their own money to reduce those individuals' payoffs. Decision-makers, in turn, had previously received selfish advice, prosocial advice, or no advice, and that advice was explicitly attributed to either human writers or GPT-4.

The results show that the choice of the decision-makers mattered most for the decision to punish, with prosocial choices virtually never punished and selfish choices punished substantially (Figure 4). The content of advice systematically shaped punishment: selfish behavior was punished less when it followed selfish advice, but more when it followed prosocial advice, compared with no-advice situations. Central to the question of machine advice being able to shift blame, the source of advice (AI vs. human) did not change punishment levels. Individuals who followed selfish AI advice were punished no more harshly than those who followed similar advice from humans. This suggests that selfish advice can effectively shift punishment regardless of whether it is generated by AI or a human.

These results extend classic attribution theories into the realm of AI, demonstrating that AI can influence moral judgment in a manner similar to human persuasion. At the same time, they highlight a potential gap between perceptions of responsibility and actual punitive behavior, raising important implications for legal and societal responses as AI increasingly influences human decisions.

Key Reference