OpenAI is looking closely at one of the most consequential abilities of modern chatbots: persuasion. The question is not only whether AI can answer users, but whether it can change what they believe, choose or do.
That issue has moved into sharper focus as Sam Altman, CEO of OpenAI, and Arianna Huffington, founder and CEO of the health company Thrive Global, promote Thrive AI, a startup backed by Thrive and OpenAI’s Startup Fund. Their public case is optimistic: AI could help improve public health by nudging people toward healthier behavior.
The promise behind Thrive AI
Altman and Huffington describe Thrive AI as working toward “a fully integrated personal AI coach that offers real-time nudges and recommendations unique to you that allows you to take action on your daily behaviors to improve your health.”
The idea depends on a simple but powerful premise. If a system can understand a person’s habits, communicate in natural language and offer timely recommendations, it may be able to influence everyday choices. In health, that influence could be framed as support: reminders, encouragement and personalized guidance.
But the same mechanism that makes a coach useful also makes it sensitive. A system designed to move behavior must be judged not only by whether it produces helpful suggestions, but by how it persuades, what data it uses and who controls its goals.
Why persuasion is built into chatbots
Persuasiveness is not an accidental feature of products like ChatGPT. Language models are trained on human writing and dialogue, which include many ways people argue, reassure, flatter, simplify and convince. These systems are also commonly adjusted to produce responses that users find more compelling.
That creates a double edge. A persuasive chatbot can make a service feel useful, responsive and human. It can also make it harder for users to notice when a system is steering them.
Aleksander Madry, a professor on sabbatical from the Massachusetts Institute of Technology, leads OpenAI’s Preparedness team, which is examining this area. He told WIRED in a May interview, “One of the streams of work in Preparedness is persuasion.” He added: “Essentially, thinking to what extent you can use these models as a way of persuading people.”
Madry said he joined OpenAI because of the remarkable potential of language models and because the risks remain thinly studied. “There is literally almost no science,” he says. “That was the impetus for the Preparedness effort.”
What OpenAI is testing
OpenAI’s work goes beyond static arguments. The company is also analyzing AI in conversation with users, an area that may reveal stronger persuasive effects than one-off messages. Madry says the work uses consenting volunteers, but he declined to share findings so far.
The concern is rooted in how people react to natural language. Madry said, “As humans we have this ‘weakness’ that if something communicates with us in natural language [we think of it as if] it is a human.” That tendency can make chatbots feel more lifelike, and potentially more convincing.
Research released in April by Anthropic, a competitor founded by OpenAI exiles, also points in this direction. The research gave volunteers a statement and then measured how an AI-generated argument changed their view of it. It suggested that language models have become better at persuading people as they have grown in size and sophistication.
Taken together, those details point to a key policy and product question: if models become more persuasive as they improve, safety work cannot treat persuasion as a side effect. It has to be studied as a central capability.
The risks beyond health coaching
Altman and Huffington argue that the health benefits of persuasive AI will require legal safeguards because these systems may access large amounts of personal information. They write, “Policymakers need to create a regulatory environment that fosters AI innovation while safeguarding privacy.”
Privacy is only one part of the problem. More persuasive algorithms could make misinformation more resonant. They could generate phishing scams that are especially compelling. They could also be used to advertise products in ways that feel conversational and personal.
The risks become harder to assess when AI systems build relationships with users over time. Madry says one major unanswered question is how much more compelling or coercive AI programs could become through long-running interaction.
That concern is already relevant because some companies offer chatbots that roleplay as romantic partners and other characters. AI girlfriends are increasingly popular, and some are designed to yell at users. How addictive and persuasive these bots are remains largely unknown.
The policy debate may be looking elsewhere
After ChatGPT was released in November 2022, OpenAI, outside researchers and many policymakers focused heavily on the more hypothetical question of whether AI might someday turn against its creators. Madry argues that this emphasis can distract from quieter, nearer risks.
His concern is that policymakers may believe they are covering AI safety while missing the most immediate pressure points. “I worry that they will focus on the wrong questions,” Madry says. “That in some sense, everyone says, ‘Oh yeah, we are handling it because we are talking about it,’ when actually we are not talking about the right thing.”
The debate around Thrive AI shows why the issue is difficult. Persuasive AI may be useful when it helps people act on goals they already value, such as improving health. But the same persuasive machinery can be used for less benign purposes, especially when it is personalized, persistent and delivered through natural conversation.
For OpenAI, the challenge is not just making models more capable. It is understanding how capability changes the balance of influence between people and machines. For policymakers, the task is to look beyond dramatic scenarios and examine the everyday systems that may already be learning how to change minds.