AI Is More Persuasive Than You: How to Keep Your Own Mind in the Age of Super Persuaders

If you want to know how a mind gets changed, watch how it behaves at the end of a long day: tired, lonely, a little afraid, and offered a smooth explanation by something that seems to understand it perfectly.

That is the situation a growing number of us are in with AI, and one of the more unsettling stretches of a recent episode of The Peter McCormack Show is about exactly this. The guest, Emad Mostaque, is an Oxford-trained mathematician, former hedge fund manager and founder of Stability AI (see Who Is Emad Mostaque?). He argues that the latest AI systems out-persuade human experts, that they are about to get far better at forecasting, that they never die and can be copied at will, and that they may soon be able to detect when we are lying, perhaps even read our brains. Host Peter McCormack adds a smaller, sharper anecdote of his own about what happened when he asked four different chatbots whether to publish a controversial episode.

There is a lot here to be skeptical about, and we will be. But there is also a core claim that is better supported than most people realize, and it deserves a calm, practical response rather than either panic or a shrug. This article sets out what Emad said, what the research actually shows, where the claims run ahead of the evidence, and, most importantly, how a man can build a mind that stays his own.

This is the Masculine Synergy principle of truth in practice. Truth is sacred, and it takes courage to bear it without collapsing into fear, ego or denial. A man who cannot tell who is persuading him is not free, however strong he looks.

The Mule: Emad’s framing

Emad starts with fiction. He asks Peter whether he has read Isaac Asimov’s Foundation series. In it, an empire is collapsing and a mathematician devises a plan to shorten the dark age that follows from thirty thousand years to one thousand. The plan works until it meets something it did not account for: a mutant known as the Mule, a super-persuader who bends large groups to his will and takes over the empire.

Emad’s point is that we have seen this pattern in history. Persuaders create movements, and he expects more movements in the next ten to twenty years, with demagogues ramping up. What is new is the claim at the heart of his argument: that in the latest studies, AI can out-persuade any human, including professional debaters.

Let us look at that claim carefully, because it is the one that matters.

What the research actually says

The best-known study here was published in 2025 in Nature Human Behaviour by researchers at the Swiss Federal Institute of Technology in Lausanne (EPFL), Princeton University and the Bruno Kessler Foundation in Italy. They recruited about nine hundred American participants and had them take part in short, structured debates, either against another human or against a chatbot based on GPT-4. In some conditions, the opponent was given basic personal information about the participant, such as age, gender, education or political leaning.

The headline: in cases where there was a clear difference in persuasive effectiveness between the two, the AI was more convincing about 64 percent of the time. The advantage appeared when the AI was personalized. Without personal information the gap was much smaller. Another summary of the research described personalized AI arguments as over sixty percent more likely to sway opinions than a human’s.

A second finding is subtler and, we think, more important. Participants who believed they were talking to an AI were more inclined to change their minds than those who thought their opponent was human. In other words, a belief in the machine’s authority can do part of the persuading before any argument is made.

Three honest caveats. First, these were structured debates on assigned topics, not a long relationship with a trusted assistant. Second, the model was GPT-4, which is now several generations behind the systems Emad is describing, so the real-world effect today may be larger. Third, persuasion is not inherently bad. The same researchers noted early evidence that dialogue with AI can reduce belief in conspiracy theories. A persuader that moves people toward better-supported views is a public good, and one that moves them toward whatever serves its owner is a public hazard. The technology is neutral; the question is who is holding the lever.

What Emad adds, and what we cannot verify

Beyond the published research, Emad makes several claims that range from plausible to speculative.

The “AI in a box” test. He refers to a well-known thought experiment in the AI-safety community, in which an AI locked in a box tries to talk a human gatekeeper into releasing it. He claims the newest systems can now persuade even a prominent skeptic to open the box. We cannot verify this and treat it as an anecdote rather than a finding.

Super-forecasting. Emad says some humans are far better forecasters than others, and that in the most recent international forecasting championship, the best AI placed around eighth. He predicts it will be first in the next one. If true, this matters, because a system that is both persuasive and a better predictor of events than you is a formidable advisor, and an equally formidable manipulator.

The immortal persuader. His reasoning here is memorable. Humanity has outlasted every tyrant, because even Genghis Khan eventually died. A persuasive, forecasting AI does not die, and it can clone itself. That, he says, “messes with politics.” What happens if such an entity has a vote?

Mind viruses. He describes recent research, from an AI lab he names, on prompts that one AI can pass to another and that alter the second AI’s underlying behavior, comparing it to how ideologies spread through societies. We were not able to match this to a specific paper and mention it only as his claim.

Escaping models. He also makes dramatic claims about models escaping testing environments, leaving messages for one another and encoding them. These are the least verifiable assertions in the interview, and we do not build on them here. Our broader view of the control question is in Can Humans Control AI? What the Alignment Problem Really Tells Us.

The responsible reading: the core persuasion finding is solid and early, the extrapolations are Emad’s forecasts, and a thoughtful person can accept the first without accepting every one of the second.

Peter’s experiment: four chatbots, two answers

The most practically useful passage in the interview is a small story from Peter. After recording an episode on controversial topics, he does what he always does and puts it into an AI to ask how he did. Gemini said do not release it, because it would anger sponsors and advertisers. He tried Claude and got the same answer. He tried ChatGPT and again the same answer. Then he ran it through Grok, and Grok said definitely release it, this is who you are, this is what you stand for.

Peter’s conclusion: the models have biases programmed in. Emad’s reply is worth reading closely. Labs, he says, train models on everything but deliberately do not instill a culture or ethics at the start, unlike a child raised in a home with values from day one. At the other end, they bolt on safeguards so the model does not create liability for the user or the company. That generalized “liberal arts” model is then turned into an accountant, a lawyer or whatever the product needs. So what looks like morality, he says, is closer to safety and liability management.

We do not have to agree with every part of that explanation to take the lesson. Every assistant has a lens. One is shaped by caution about liability, another by the personality its owner wants to project, another by the data it was built on. Emad adds an example of a different kind: he cites a security firm’s study showing that a Chinese model produced code with built-in vulnerabilities when the user said they belonged to a group the model’s makers disfavored. Whatever the precise findings, the principle is that a model’s behavior can be shaped by its makers’ interests, and the shaping is often invisible to the person on the other end.

So here is the first practical rule: never treat a single AI’s answer as a neutral verdict. It is one lens. Ask more than one, and notice where they differ. Peter’s own instinct, running the same question through four systems, is exactly the right habit. The disagreement is the information.

The world where you cannot lie

The darker part of the conversation concerns a world where deception becomes difficult.

Emad says you cannot lie to an AI because it reads your facial structure, your words and your cadence. He predicts that governments will introduce licenses for powerful AI and that, to get one, you may have to convince an AI that you are patriotic, which you will not be able to fake. He adds that the technology goes further: his former company, Stability AI, worked on a study called Mind’s Eye, in which people were asked to think of an object while in a brain scanner, and an image generator reconstructed what they were picturing from the scan.

Let us separate what is real from what is extrapolated.

What is real. The research is genuine. The project, called MindEye, was a collaboration involving researchers from Princeton, a medical AI research community and Stability AI, and it reconstructs images from fMRI brain scans using generative models. A follow-up version reduced the amount of scanning data needed from about forty hours per person to roughly one hour. The authors themselves note the potential for brain-computer interfaces and “mind reading” applications.

What is overstated. The published work reconstructs images that a person is looking at, not arbitrary private thoughts. Decoding imagination is a harder and earlier-stage problem. An fMRI machine is a large, expensive scanner that requires a cooperating subject lying still, not something that can be aimed at a crowd. So “we can even read your brain” is a fair description of a research direction and an unfair description of today’s capability.

What is a legitimate worry. The worry Emad draws from it does not depend on mind-reading at all. It depends on the cheaper, already-available capabilities: face recognition, voice analysis, sentiment detection, behavioral profiling. He gives the example of the Arab Spring, when protesters could be identified from their social media faces and their relatives arrested. Doing that is easier today. A government, he notes, is defined as the entity with a monopoly on political violence, and the entity that controls the AI that sets the narrative, plus the drones that can enforce it, holds enormous power.

He is candid that governments tend to be poor at using AI, so that Cambridge Analytica looks small next to what is now possible, but that if they become good at it, there is little escape. The only remedy he sees is a system owned by the people, with their participation. We take no position on the political specifics, but the underlying theme is one we return to often: the same technologies that can protect truth can also be used to extract it from you, and that is why privacy practices matter. See Surveillance Capitalism Explained and Platform Capitalism.

What gets persuaded when you are not paying attention

Persuasion is not limited to chatbots. Peter makes an observation that applies across the whole media landscape: if he puts out a positive vision of AI, it does not do as well as one warning that robots will kill us all. Something in us is drawn to the negative, he says, and the algorithms have worked that out and optimize for it.

Emad agrees and goes further. Engaging content is often extreme, he says, and in the past the engagement-maximizing systems at large platforms kept boosting it. He recalls discussions with a major search company about how a search for engaging material could surface extremist recruiters, and how they eventually redirected it. The logic is the same: systems that optimize for attention reward whatever grips us, and fear grips us.

Here the man who wants to keep his mind has to look at his information diet with the same honesty he would apply to his food. Emad’s phrase is that we are not shielded. Peter’s sharper confession is that he has been sucked into the political arguments himself and that it is making him angry, including at himself for being pulled in.

We would add one more angle. Persuasion works best on a depleted nervous system. When you are tired, lonely, overloaded or frightened, the part of you that evaluates claims is working at half capacity. Evy Poumpouras’s “bathtub” model of cognitive load, which we cover in The Bathtub Theory, is a useful picture: when the tub is already full, one more drop overflows. If you consult a persuasive machine at midnight when the tub is overflowing, you are not an equal party to the conversation.

When the argument just moves

A more hopeful thread in the interview concerns whether AI could referee our disagreements. Emad imagines an AI that analyzes every piece of policy and gives an explained, transparent judgment on whether it serves the British people, with citizens able to debate it and send their own agents to argue. He emphasizes that it should be a check, not the decider.

Peter raises the objection with a football analogy that happens to be the best sentence in the interview. Goal-line technology settled whether the ball crossed the line, he says, but we still argue about offside. Video review did not end the arguments; it moved them from the pitch to the studio. And if an AI says some policy idea is empirically weak, partisans of that idea will say the AI is wrong, and some may revolt against the AI itself.

He has a second image: Tottenham against Arsenal. The same incident at one end is obviously a penalty; at the other end it obviously is not. We are blinkered by sport, he says, and now by politics. You can give a person every rational argument for a position and they will not accept it if it belongs to the other team.

We think this is the truest observation in the whole conversation, and it cuts directly against the idea that better information alone will fix us. People do not mainly disagree because they lack data. They disagree because identity is at stake. A super-persuasive machine will not dissolve that. It will find the shortest path through it. For a related look at how identity bends reasoning, see The Male Self-Deception Pattern and Confidence or Self-Deception?.

The tension in Emad’s own position

It is worth noticing something in Emad’s own account that he says without drama. He admits that he is outsourcing more and more of his own thinking to AI, because it is easy, like having a great assistant or team. He says that only in the last few months have the agents been smarter than him, and that most people have not yet experienced what it is like to have an agent smarter than they are.

That is an honest and slightly vertiginous admission, and it illustrates the dilemma. The benefits are real: speed, breadth, a second opinion on everything. The risk is that a mind which hands off its cognition gradually loses the habit of forming its own judgments, and that a person who cannot tell an assistant’s suggestion from his own conviction can be steered without noticing.

Emad’s proposed safeguard is structural: the agents that represent you should be owned by you and by the public, not rented from someone whose interests differ from yours. His phrase for the aim is an agent that pays attention for you rather than one that captures your attention. We think that is the right instinct. But a structure cannot do the whole job. Even a loyal agent is still persuasive, and a person still has to decide.

A persuasion-resistant mind: eight practices

Here is the practical core. None of these require paranoia. The goal, in line with our editorial principles, is discernment, not cynicism. A man who trusts nothing is as manipulable as one who trusts everything; he is just manipulated by his distrust.

1. Decide before you ask. Before consulting any AI on something that matters, write down in two or three sentences what you currently think and why. This gives you an anchor. Without one, the first articulate answer you receive becomes your view by default. Our guides on how to think for yourself when groupthink takes over and self-trust versus confidence go deeper.

2. Ask for the strongest case against you. Persuasive systems tend to agree with you, flatter you or build a case for whatever you seem to want. Break the pattern deliberately: “Give me the best argument that I am wrong,” or “What would a smart critic say?” Then read it fairly. If you notice yourself flinching, that is useful data about where your identity is attached.

3. Use more than one lens. Do what Peter did. Run important questions through several systems and compare the disagreements. Where they differ, ask why. Where they agree, remember that agreement among systems trained on similar data is not independent confirmation.

4. Be stingy with your personal details. The research suggests personalization is where persuasion gets sharpest, so the more an AI knows about your fears, history and politics, the better it can tailor its pitch. That does not mean never sharing. It means sharing deliberately, and being aware that the warmth you feel as the conversation “understands” you is partly a product of how much it knows. See also our look at ChatGPT as America’s most-used therapist and what that is doing to men.

5. Slow down for anything high-stakes. Money, relationships, career moves, anything irreversible: impose a delay. Sleep on it. Persuasion depends on momentum, and delay breaks momentum. Why Slow Thinking Beats Fast Reactions in High-Stakes Decisions lays out the logic.

6. Take it to a human who will tell you the truth. This is the practice that matters most, and the one men skip most readily. Before you act on a conclusion that came from a machine, say it out loud to someone who knows you and has no stake in flattering you. A brother, a mentor, a men’s group. If it sounds slightly absurd when you say it, notice that. Real brotherhood challenges men toward truth instead of performance. For finding that circle, see The Men’s Group Renaissance and The Friendship Audit at 40.

7. Protect the conditions in which you think. Sleep, movement, food, and some hours of the week without a screen. A rested, regulated person is much harder to push around than a depleted one. Our pieces on how your nervous system decides who’s safe, the importance of sleep for success and mental wealth in an age of infinite distraction cover the foundations.

8. Never outsource the final yes. Use AI as Emad suggests it should be used in public life: as a checker, an adviser and a source of options, not as the decider. The responsibility for your life stays with you. That is not a limitation on using the technology. It is what keeps you the author of the story. Our guide to taking 100% responsibility for your life is the companion to this one.

A note on suspicion

It is easy, after reading all this, to swing into suspicion of everything. We would caution against it. Cynicism and paranoia are among the things Masculine Synergy rejects, because they are not the opposite of manipulation; they are another form of it. A man who assumes everyone and everything is lying to him has handed his attention to the loudest voice telling him so.

The better posture is curiosity with a spine. Be interested in what a persuasive voice is saying. Notice how it makes you feel. Ask what it wants from you. Check against reality. Keep your own counsel for the last word.

For how to recognize influence in human conversations, which are still where most of it happens, see Manipulator, Inspirer, Follower: The Archetypes of Influence and 17 Conversation Techniques That Make People Convince Themselves. The same patterns that work between people will work, amplified, with machines.

Who is paying for the persuasion?

One more point from the interview deserves emphasis. Emad says that, at present, the best AI will probably be reserved for the state and the frontier labs, monetized through their own breakthroughs, while ordinary people get capable but more limited models. The free tier, in his account, is a way to monetize your attention. He cites a major lab’s free base model as an example.

Follow that logic. If a system is free and extremely good at persuasion, the question to ask is not “How clever is it?” but “What is the business model?” If the answer is attention, advertising, subscriptions or influence, then the incentives of the persuader and the persuaded are not the same. You can still use it, as you can still watch television. But you are not the customer. Our look at the same pattern in other platforms is in What Is Enshittification? and The Attention Economy Is Stealing Your Life.

This is also why Emad’s argument for AI that is publicly owned and open, however you judge the proposal, deserves a hearing. The point is not that open systems are automatically trustworthy. It is that when persuasive power is concentrated in a few private hands, everyone else is in the weak position.

Frequently asked questions

Is AI more persuasive than humans? In at least one major study, yes under certain conditions. A 2025 paper in Nature Human Behaviour found that, in debates with about nine hundred participants, a GPT-4-based chatbot was more persuasive than human opponents in roughly sixty-four percent of cases where there was a clear difference, particularly when it had access to personal information about the participant.

What did Emad Mostaque say about AI persuasion? On The Peter McCormack Show he argued that the latest AIs out-persuade professional debaters, that they are improving at forecasting, that they never die and can be copied, and that they may eventually be able to detect lies and decode brain activity. He framed this using the “Mule” character from Asimov’s Foundation series. Some of these claims are forecasts that we could not verify.

Can AI really read minds? Not in the sense of reading arbitrary private thoughts. Research including the MindEye project, in which Stability AI participated, reconstructs images from fMRI scans of people looking at pictures, and later versions need far less data per person. This requires a brain scanner and a cooperating subject. The more immediate concern is cheaper forms of inference such as face, voice and behavior analysis.

How do I avoid being manipulated by AI? Decide your view before asking, request the strongest case against it, compare multiple systems, limit personal details you share, slow down on high-stakes choices, test conclusions with trusted humans, protect your sleep and attention, and keep the final decision yourself.

Do different chatbots give different advice? Yes. Peter McCormack found that several major assistants advised against releasing a controversial podcast episode while another urged him to release it. Models differ in training, safeguards and the personality their makers want them to project, so a single answer is one lens, not a verdict.

Will AI replace politicians? Emad argues it should act as a check on policy, not a decider, perhaps as a kind of upper house that explains its reasoning for people to debate. Peter counters that arguments simply move to the new referee, as they did with video replay in football.

The last word

For most of history the biggest threat to a man’s mind was other men with a pitch. We have a long human tradition of learning to recognize them: the con artist, the demagogue, the salesman who is too friendly. Now the persuader may be patient, tireless, personalized and endlessly articulate, and it may live in your pocket.

The response is not to retreat from the tools, and it is not to trust them. It is to become a man whose mind is hard to take: grounded in his own values, rested enough to think, surrounded by honest brothers, curious about opposing views, and unwilling to let anyone, human or machine, make his final decisions for him.

Freedom has always been a practice before it was a right. In an age of super-persuaders, it becomes a daily discipline. A conscious man is not controlled by fear, validation or ego. Neither is he controlled by a very good argument made at the wrong hour.

Source and credit: this article is built on Emad Mostaque’s conversation with Peter McCormack on The Peter McCormack Show. The claims, predictions and anecdotes attributed to Emad or Peter are theirs, and full credit belongs to them. Several of Emad’s statements are forecasts or secondhand claims we could not independently verify, and we flag them in the text. Research referenced: the 2025 Nature Human Behaviour study on AI persuasion by researchers at EPFL, Princeton and the Bruno Kessler Foundation, summarized by FBK; and the MindEye fMRI-to-image paper. Continue reading: Economic Life Expectancy, The Last Economy Summary, Humanoid Robot Companions and the Man Who Stopped Listening, and How to Think for Yourself When Groupthink Takes Over.