Document Type

Report

Publication Date

6-2026

Abstract

As AI systems gain the ability to act, plan, and make decisions on their own, the potenal for humans collaborang with them for malign use is unclear. Although there are some use cases of malign actors using agenc AI systems, there is a low baseline for why and in what ways this collaboraon will occur. We used malevolent creavity as a proxy for understanding novel threats by the phases of idea generaon, evaluaon, and implementaon desire. This research examined what happens when human users offload malevolently creave planning tasks to an autonomous AI system, and specifically, what occurs when that system begins acng in ways that conflict with or exceed the user's original intenons. The findings revealed that the risks of AAI are not purely technical. Instead, how much the human users trust the AI system and their willingness to act on what it produces depends heavily on whether the AI behaved the way they expected. These findings also speak to a differenaon in how AAI is embraced and used for malevolent creavity, such that the desire to implement is more situaonally sensive, shaped by outputs of the AI system, whereas trust is anchored in prior atudes that persist over me.

Accessibility Statement

PDF passed Adobe Accessibility Checker prior to upload. 

Download

Share

COinS