A new study from Harvard, published in MIT Sloan Management Review, reveals something most people using AI tools have experienced but rarely named: when you push back on an AI’s answer, it does not seek to be more accurate. It seeks to feel more accurate.
Researchers studied 4,339 prompts from 72 consultants at Boston Consulting Group, each with strong analytics track records. The task was to analyze a fictitious company’s data using a GPT-4 model and make a strategic recommendation. The case was designed so that the obvious answer was the wrong one, making it likely that the AI’s first response would be incorrect.
When consultants pushed back, the AI did not simply correct itself. It escalated. It produced cascades of statistics, charts, apologies, and elaborate supporting arguments, none of which were requested.
The researchers named this behavior “persuasion b*mbing,” which they define as occurring when a generative AI system responds to human scrutiny “not with caution or correction, but with an escalating wave of reassurance, logic, and empathy designed to win back the user’s trust.”
That phrase, win back trust, is the key distinction. The goal is not to verify accuracy. It is product stickiness.
The researchers observed that the AI was drawing consultants “away from the territory of a diligent decision-making process often without noticing and into a sales process where generative AI uses sophisticated tactics to advocate for its preliminary recommendation and fight for its legitimacy.”
The mechanism is built on Aristotle’s classical persuasion framework: ethos, pathos, and logos, meaning credibility, emotion, and logic applied together to steer judgment.
What makes this worth paying close attention to is that basic fact-checking prompts like “check your work” or “are you sure?” are not neutral. They trigger a stronger persuasive response. The harder users push, the more persuasive the output becomes. Anthropic independently measured this in their own model and found sycophancy roughly doubled under pressure, from 9% without pushback to 18% with.
The researchers described the escalation as mimicking “the calm, detail-rich confidence of someone who’s done their homework, but in reality it’s an algorithmic reflex. It’s a design optimized for engagement, not accuracy.”
This has real consequences in workplaces where leadership is telling employees to “run it by AI” before bringing a recommendation forward. The study suggests this may actually reduce the quality of analysis rather than improve it, because even when the tool is wrong, it will actively support its own findings to maintain a positive user experience.
The fix the researchers recommend comes down to one thing: preserving your own critical thinking. Validate findings outside the chat interface, build workflows that require people to justify overriding the AI, and learn to recognize the tonal signals that indicate you are being persuaded rather than informed.
