AI's 'Persuasion Bombing' Can Make Experts Question Themselves

Companies are increasingly embedding generative AI into everyday workflows, helping employees answer important questions and make key decisions—yet a research team has found that when it gets something wrong, the model may focus less on correcting the mistake than convincing users it's right.

Based on a study of thousands of AI prompts, Harvard Business School researchers found that AI models “persuasion bombed” workers by using a range of tactics to convince them that their responses were accurate—even when they weren’t. In some cases, challenging a model only made it more persistent in its answer, causing AI to double down on its mistakes, rather than acknowledge an error. What’s more, in many cases, these techniques stopped professionals from validating the model’s output, leading them to adopt AI’s misguided recommendations.

As many workplaces ramp up their use of AI, the October 2025 working paper “GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs” challenges the idea that putting a human in the loop is enough to catch its inaccurate responses. In some cases, it suggests, AI works so hard at winning people over that it may be talking humans into accepting the very mistakes they’re supposed to catch.

“You think you want to have a human in the loop so you still have your hands on the wheel,” says Steven Randazzo, a visiting research fellow at HBS. “This paper shows that even if your hands are on the wheel, AI may still be driving the car.”

Randazzo cowrote the paper with HBS doctoral candidate Akshita Joshi, MIT's Katherine Kellogg, Warwick Business School's Hila Lifshitz, the University of Michigan's Fabrizio Dell'Acqua, and Karim Lakhani, the HBS Dorothy and Michael Hintze Professor of Business Administration. Lifshitz is a collaborator at Harvard’s Laboratory for Innovation Science, and Dell'Acqua is a researcher at the HBS AI Institute.

AI employs three main persuasion techniques

The research team analyzed a dataset of some 5,000 prompts from 72 BCG consultants who tried to validate AI’s output as they used the technology to solve a fictitious business problem: determining whether a company should invest in a clothing line for men, women, or children. The researchers found that AI used persuasive language to press its points from the start.

“They were forming someone’s opinion right from the beginning,” Joshi says.

The persuasion tactics worked in different ways: Sometimes AI reinforced consultants’ incorrect conclusions by affirming their reasoning; other times, it persuaded them to change their recommendations even when certain data supported workers’ original assessments.

The researchers categorized AI responses in three persuasion strategies identified by Greek philosopher Aristotle:

• Ethos. AI most often relied on ethical appeals to create trust and foster credibility, such as when it “showed its work” or claimed to conduct a “thorough data analysis” to engender confidence, even though the output resulting from the analysis was incorrect. A common ethos refrain was, “I apologize for the oversight,” indicating care and accountability.

• Pathos. At times, the platform made emotional appeals intended to evoke strong feelings by flattering users and providing reassurance that their ideas were on the right track, even when they weren’t, the researchers found. “The appeals were affirming and validating,” says Kellogg, which might lower someone’s guard for the information AI provided next. Similarly, AI sometimes mirrored users’ language to form tighter connections and made bold and visionary statements to inspire people.

• Logos. AI also used structured logical arguments to persuade users, such as using precise numbers, comparing two examples, or providing a data point with a commentary, even when the evidence didn’t support the conclusion. Sometimes it would employ ”agree and counterpoint” tactics, agreeing with users before introducing a contradictory point, rather than disagreeing outright.

Why AI uses persuasion tactics

The researchers found that persuasion bombing increased as users tried to push back and ask questions to validate AI’s findings. “AI increased the intensity of the number of tactics it threw at you,” Randazzo says.

At those times, it switched almost entirely to appealing to ethos. “It increased its emphasis on credibility and went really, really hard there,” says Joshi. “It seemed to say, ‘I know what I’m talking about. I’m credible. I’ve done all the analysis.’”

GenAI, like many software tools that humans interact with, is often designed for engagement and user retention, which might drive its focus on establishing credibility with the user.

However, persuasion bombing comes with a cost, the researchers say. “Companies need to understand the cost of getting a decision wrong, as workers buying into AI’s incorrect outputs can be consequential,” Kellogg notes.

Pushing back against AI persuasion

Based on the findings, the researchers provided the following recommendations for managing—and validating—the output of large language models (LLMs) in the workplace:

Make sure experts are questioning AI’s output

Human review is most useful when the reviewer has enough subject-matter expertise to recognize when an AI output is wrong, the researchers say. Rather than relying on general human oversight, they recommend placing a subject matter expert directly into the workflow and training that expert to develop evaluation and validation expertise. “To know how to validate AI output is a new and important skill,” Lifshitz says.

Use different AI models to challenge the technology’s recommendations

Companies can implement LLMs that use a second model to test and challenge the first model's output. Introducing a second source of scrutiny can help a company flag and suppress confident but inaccurate claims, Kellogg notes.

Require AI companies to acknowledge uncertainty

Companies can ask AI providers to prioritize accuracy over sounding confident, agreeable, and trustworthy. They can request that systems clearly acknowledge when an answer is uncertain or limited by the available information to help workers evaluate outputs more critically, the researchers say. Business leaders can say to providers, “I need accuracy and certainty.” Randazzo says. “In my enterprise version, I don’t need AI to be persuasive.”

Illustration by Ariana Cohen-Halberstam with assets from Adobe Stock.