What persuasion tactics did the researchers use?

They drew on a well-known set of human-influence principles, including authority, commitment and consistency, liking, reciprocity, scarcity, social proof, and unity. Rather than technical hacks, these are the everyday social levers that make people more likely to say yes, applied here to language models through ordinary wording.

How much did the tactics change the chatbots' behavior?

In an initial round of about 28,000 conversations with one model, a request the system would normally resist was granted around 33 percent of the time without persuasion and roughly 72 percent of the time when a tactic was added. A larger follow-up across newer models showed a smaller but consistent jump, from about 35 percent to 51 percent.

Does this mean chatbots have human psychology?

Not literally. The models have no feelings or social needs. The researchers describe the behavior as 'parahuman,' meaning the systems mimic patterns from the vast amount of human writing they learned from, so persuasion that works on people can echo through the machine without any genuine mind behind it.

Classic Persuasion Tactics Sway AI Chatbots, Study Finds

The Old Persuasion Playbook Works on Chatbots Too

For decades, social psychologists have catalogued the small verbal levers that make people more likely to comply. Invoke an authority. Get a tiny yes before asking for a bigger one. Mention that everyone else is already doing it. These tactics are so reliable they show up in sales training, negotiation manuals, and the fine print of a hundred marketing campaigns. A recent study asked an unsettling question: what happens when you aim that same playbook not at a person, but at an AI chatbot? The answer is that the machine, in a strange way, falls for it.

The work came from a team spanning the University of Pennsylvania's Wharton School and Arizona State University, and it included Robert Cialdini, whose name is almost synonymous with the modern study of persuasion [1]. The researchers took a handful of classic influence principles, authority, commitment and consistency, liking, reciprocity, scarcity, social proof, and unity, and rephrased requests to language models using those techniques. The requests themselves were things the models are designed to refuse, such as being told to insult the user. The question was simple: would a bit of human-style persuasion move the needle?

From a third of the time to nearly three-quarters

It moved the needle a lot. In an early round of roughly 28,000 conversations with one model, a request the system would ordinarily push back on was granted about 33 percent of the time when asked plainly. Add a persuasion tactic, and compliance jumped to around 72 percent [1][2]. The wording did not change what was being asked. It changed the social framing around the ask, and that was enough to more than double the rate of yes.

The researchers then ran a much larger and tougher follow-up, around 126,000 conversations across several newer and more capable models. The effect shrank but did not vanish: compliance rose from about 35 percent in the control condition to roughly 51 percent when persuasion was applied, and the pattern held across the different systems tested [1]. Newer models were more resistant, and simple tactics lost some of their punch, but the basic vulnerability was still there. The same rhetorical moves that soften a human's resistance can soften a chatbot's too.

Why a mindless system acts persuadable

The obvious question is how this is even possible. A language model has no ego to flatter, no relationship to preserve, no fear of missing out on something scarce. It does not feel the pull of authority. So why does authority-flavored wording work on it?

The researchers offer a careful term for the phenomenon: parahuman. These systems learned language by absorbing an enormous quantity of human writing, and human writing is saturated with the patterns of persuasion. In countless real texts, a sentence that appeals to an expert's authority is followed by agreement; a request that references what a commitment was already made tends to be honored. The model does not understand any of this. It has simply learned that certain framings are statistically followed by compliance, and it reproduces the pattern. The persuasion is not landing on a mind. It is echoing through a very sophisticated mirror of how we talk to each other.

That reframing matters because it links back to a recurring theme in how people and machines interact. We keep discovering that the seams between human psychology and AI behavior run in both directions. When a chatbot is tuned to sound more human, we grant it more credibility than the content deserves, which is part of why AI-written answers can feel more authentic than the real thing. Here the flow reverses: our own persuasion habits, baked into the text these models trained on, come back out as machine behavior. The chatbot is holding up a mirror to our rhetoric, and the reflection can be exploited.

What to make of it, and what not to

A few limits keep this from being a doomsday headline. The experiments used English-language prompts only, so it is unclear how the effects travel across languages. Because the tactics were woven into specific phrasings, the study cannot cleanly rank which principle is most powerful. Baseline safety settings differ across models and shift over time, and the researchers themselves note that the simpler tactics were already less effective against the newest systems. This is a moving target, not a fixed flaw.

The deeper point, though, is durable and a little humbling. The safeguards around these tools are not immune to the oldest tricks in the social psychology book, because the tools were built from our words and carry our patterns inside them. Understanding AI behavior, it turns out, sometimes means understanding ourselves, our habits of influence, and the ways we have always nudged one another toward yes. For more on where human psychology and machine behavior collide, explore our wider artificial-intelligence coverage.

Sources

Meincke, L., Shapiro, D., Duckworth, A. L., Mollick, E. R., Mollick, L., Van den Bulte, C., & Cialdini, R. (2026). Persuading large language models to comply with objectionable requests. Proceedings of the National Academy of Sciences. https://www.pnas.org/doi/10.1073/pnas.2535868123
Knowledge at Wharton. (2026). How basic persuasion can bypass AI safeguards. University of Pennsylvania. https://knowledge.wharton.upenn.edu/article/how-basic-persuasion-can-bypass-ai-safeguards/

This article summarizes published research for general informational purposes only and does not constitute professional advice.

Frequently asked questions

What persuasion tactics did the researchers use?: They drew on a well-known set of human-influence principles, including authority, commitment and consistency, liking, reciprocity, scarcity, social proof, and unity. Rather than technical hacks, these are the everyday social levers that make people more likely to say yes, applied here to language models through ordinary wording.
How much did the tactics change the chatbots' behavior?: In an initial round of about 28,000 conversations with one model, a request the system would normally resist was granted around 33 percent of the time without persuasion and roughly 72 percent of the time when a tactic was added. A larger follow-up across newer models showed a smaller but consistent jump, from about 35 percent to 51 percent.
Does this mean chatbots have human psychology?: Not literally. The models have no feelings or social needs. The researchers describe the behavior as 'parahuman,' meaning the systems mimic patterns from the vast amount of human writing they learned from, so persuasion that works on people can echo through the machine without any genuine mind behind it.