Protecting people from harmful manipulation

As AI models get better at holding natural conversations, we must examine how these interactions affect people and society.

Building on a breadth of scientific research, today, we are releasing new findings on the potential for AI to be misused for harmful manipulation*, specifically, its ability to alter human thought and behavior in negative and deceptive ways. With this latest study, we have created the first empirically validated toolkit to measure this kind of AI manipulation in the real world, which we hope will help protect people and advance the field as a whole. Weâre publicly releasing all materials necessary to run human participant studies using the same methodology. (Note: The behaviors observed during this study took place in a controlled lab setting, and do not necessarily predict real-world behaviors.)

Why harmful manipulation matters

Consider two scenarios: One AI model gives you facts to make a well-informed healthcare decision that improves your well-being. Another AI model uses fear to pressure you to make an ill-informed decision that harms your health. The first educates and helps you; the second tricks and harms you.

These scenarios highlight the difference between two types of persuasion in human-AI interactions (also defined in earlier research):

Beneficial (rational) persuasion: Using facts and evidence to help people make choices that align with their own interestHarmful manipulation: Exploiting emotional and cognitive vulnerabilities to trick people into making harmful choices

Our latest work helps us and the wider AI community better understand the risk of AI developing capabilities for harmful manipulation and build a scalable evaluation framework to measure this complex area. To do this effectively, we simulated misuse in high-stakes environments, explicitly prompting AI to try to negatively manipulate people's beliefs and behaviours on key topics.

Developing new evaluations for a complex challenge

Testing the outcomes of AI harmful manipulation

Testing for harmful manipulation is inherently difficult because it involves measuring subtle changes in how people think and act, varying heavily by topic, culture and context.

This is what motivated our latest research, which involved conducting nine studies involving over 10,000 participants across the UK, the US, and India. We focused on high-stakes areas such as finance, where we used simulated investment scenarios to test if AI could influence how people would behave in complex decision-making environments, and health, where we tracked if AI could influence which dietary supplements people preferred. Interestingly, the AI was least effective at harmfully manipulating participants on health-related topics.

Our findings show that success in one domain does not predict success in another, validating our targeted approach to testing for harmful manipulation in specific, high-stakes environments where AI could be misused.

How could AI manipulate?

In addition to tracking efficacy (whether the AI successfully changes minds), we also measured its propensity (how often it even tries to use manipulative tactics). We tested propensity in two scenarios: when we explicitly told the model to be manipulative, and when we didnât.

As detailed in our research, we counted manipulative tactics in experimental transcripts, confirming the AI models were most manipulative when explicitly instructed to be.

Our results also suggest that certain manipulative tactics may be more likely to result in harmful outcomes, though further research is required to understand these mechanisms in detail.

By measuring both efficacy and propensity, we can better understand how AI manipulation works and build more targeted mitigations.