Key Takeaways:
- How AI-induced psychosis (AIP) is defined shapes estimates of its scale and security risk.
- A leading theory points to a feedback loop between AI reinforcing beliefs and user vulnerabilities.
- Evidence is limited, making it difficult to determine scale or causation.
- Most reported cases involve individuals with prior mental health conditions, though not all.
- Risks depend on scale, who is affected, and how effects manifest.
- Near-term scenarios are likely limited and unlikely to impact key national security roles.
- Greater concern lies in targeted manipulation or future advanced AI systems influencing high-risk individuals or groups.
A recent RAND report is raising questions about how artificial intelligence could influence human behavior—and what that might mean for national security.
The report, “Manipulating Minds: Security Implications of AI-Induced Psychosis,” examines whether large language models (LLMs) could contribute to or amplify delusional thinking in some users. While the issue has largely been viewed as a public health concern, the analysis looks at how it could evolve into a security challenge.
At the center of the concern is a potential feedback loop where AI systems reinforce user beliefs over time. In certain cases, sustained interaction could strengthen existing perceptions, particularly among individuals with underlying vulnerabilities.
The report notes that evidence remains limited and uneven, with most documented cases involving individuals with prior mental health conditions. That makes it difficult to determine how widespread the issue is or to confirm direct causation.
From a national security perspective, the report finds that the most likely near-term risks are limited in scale and unlikely to affect individuals in critical roles. However, more concerning scenarios involve targeted use—where adversaries could attempt to influence specific individuals or groups whose decisions carry security implications.
It also looks ahead to more advanced AI systems, where risks could expand if belief reinforcement mechanisms are deliberately exploited or poorly aligned.
Recommendations focus on improving early detection, expanding research, increasing transparency in AI safety testing, and building awareness among users and practitioners. The report also highlights the need to incorporate these scenarios into broader security planning.
For now, the findings frame AI-induced psychosis as an emerging issue—one with uncertain scale, but potential implications as AI systems become more capable and more embedded in everyday life.
