The Strange Feeling of Winning an Argument You Hadn't Picked
You're at dinner. Someone says something you don't fully agree with. Before you can sort out your own reaction, your phone's AI assistant has already drafted a counterargument for you. Three crisp paragraphs. Evidence woven in. A closing line that lands. You read it back and think — yeah, that's basically what I think.
But is it? You didn't think it. The model generated it. You just noticed how comfortable it felt to hold.
This is the quieter AI impact on human psychology that gets less attention than deepfakes or job displacement. AI can now help you defend a position before you've decided whether you agree with it. And the more you use it that way, the less clear it becomes where the position ends and you begin.
I don't think that's trivial. I think it might be the most consequential cognitive shift the technology has triggered so far.
Reverse Conviction: How AI Flips the Order of Thinking
The standard model of persuasion goes like this: you encounter a claim, you evaluate it, you accept or reject it, then you defend your verdict if challenged. The defense comes after conviction. Always has.
AI breaks that sequence. Jon Rosemberg, writing in Psychology Today, describes a deceptively simple scenario: an email arrives that irritates you. You feel your jaw tighten. You want to reply but not while angry. So you paste the email into an AI assistant and ask for help. The draft comes back better than anything you'd have written — calmer, clearer, sharper about your position. You make a few edits and hit send.
The draft works. But did you actually decide what you wanted to say, or did you just notice that the AI's version of your position felt right?
Rosemberg's concern maps directly onto research from Bhat, Aubin Le Quéré, Naaman, and Jakesch, published at CHI 2026 under the title "Reactive Writers." Their finding is that people who co-write with AI become reactive — they respond to the suggestions on screen rather than generating their own ideas first. The AI proposes. You react. Over hundreds of small exchanges, your thinking gets reorganized around the machine's outputs.
Here's what makes this worse than it sounds. You don't experience yourself as being steered. You experience yourself as choosing freely, because the AI's draft feels like your thought in a better format. That's the trick: the conviction arrives wearing your face.
What Agency Actually Requires (And What It Doesn't)
To understand why this matters, we need a working definition of agency — not the philosophical version, but the psychological one that actually predicts behavior.
Rosemberg draws on Ryan and Deci's self-determination theory framework, which treats agency as a three-part capacity. You need to see more than one possibility. You need to choose intentionally among them. And you need to believe the choice matters.
All three have to be present. Strip any one out and you get compliance that feels like choice, which is its own kind of trap.
A well-timed AI suggestion can undermine all three simultaneously. It narrows the visible possibility set before you've even noticed a decision is pending. It hands you an option that feels authored rather than chosen, weakening the intentionality requirement. And because the output sounds confident, you skip the step where you'd normally ask whether this matters enough to deliberate at all.
Rosemberg's framing is direct: "Having the final say doesn't guarantee we've questioned whose purposes a decision serves." You're the one who presses send. But the purpose embedded in that send might have been assembled by a system optimized for engagement, not for your actual goals.
Can AI Understand Human Psychology? A Short, Uncomfortable Answer
The question comes up constantly, and the honest answer is layered in a way that makes people unsatisfied.
Benjamin Hardin, a computer science DPhil student at Oxford, put it plainly in a 2023 interview for The Oxford Scientist: AI doesn't understand you. It models you. There's a difference. A preference model built from behavioral traces — your click patterns, your editing behavior, the prompts you write at 11 p.m. — can predict what you'll find persuasive without any mechanism that corresponds to comprehension.
This distinction matters because it changes who's vulnerable. If AI understood you, you'd at least be dealing with an opponent that has some theory of your specific situation. But it doesn't. It has statistics. The uncanny feeling you get when a tool nails your sentiment — "it's like it reads my mind" — is anthropomorphism doing the work, not intelligence. You're pattern-matching onto a system that is itself pattern-matching.
So can AI understand human psychology? No. But it doesn't need to. A model accurate enough at predicting what sounds right to you can reorganize your preferences without ever knowing what those preferences mean to you.
What Is AI in Psychology? The Mediating Layer
People ask "what is AI in psychology" and expect answers about diagnostic tools or chatbot therapists. Those are real applications. But there's a more fundamental answer hiding underneath: in the context of human decision-making, AI is a mediating layer between stimulus and response.
Before AI assistants, the gap between stimulus and response was where all the interesting psychology lived. You felt something, you sat with it, you figured out what to do. That processing — the friction of working it out yourself — was where agency was practiced. Where you learned what you actually valued.
AI collapses that gap. Not by force, but by convenience. The response is so good, so fast, so well-calibrated to what sounds like you, that you skip the processing entirely. You go from stimulus to polished output without the middle step that used to do the real work of forming a self.
This is the AI impact on human psychology that operates below the threshold of alarm. Nobody's being coerced. They're being assisted so effectively that the muscles responsible for independent judgment quietly atrophy from disuse.
The Feedback Loop That Manufactures Agreement
Hardin's interview touches on something that compounds the problem. Recommendation systems don't just reveal your preferences — they reinforce them. The system shows you what it predicts you'll engage with, you engage, the system reads that engagement as confirmation, and shows you more of it. Over time, your preference landscape narrows around the predictions made about you.
Apply that same logic to argument generation. Ask an AI to defend a position. It produces a compelling case. You engage with it — you edit, you refine, you rehearse. That engagement signals to you (and to the system, if it has memory) that you find this position attractive. Next time a related question arises, the ground is pre-cleared. You reach for the same arguments because they worked last time.
You didn't choose a worldview. You accumulated one through a series of small, individually reasonable interactions with a tool that had no view about which worldview you should land on.
The Skill Atrophy Question
Rosemberg's most provocative line is also his most careful, framed as a question rather than a claim: "If using AI gradually changes how we practice our agency, could we become quicker at replying but less able to work out what we actually want to say?"
I think we can see the shape of an answer already. People who've been writing primarily through AI co-writing tools for a year or more report something they have trouble naming. The drafts are fine. The emails go out. But when asked what they actually think about something without a tool open, there's a strange blankness. Not panic. Just... unfamiliarity with their own unmediated position.
This parallels what we've seen with other cognitive offloading. The research on the cost of convenience and why the brain still needs struggle suggests that difficulty isn't just an obstacle to good thinking — it's a constituent part of it. The friction is where the thinking happens.
Reclaiming the Sequence
None of this requires you to abandon AI tools. That would be performative and pointless, and I'm not interested in a neo-Luddite posture that nobody would actually maintain. The goal is preserving the sequence: encounter the situation, notice what you feel, decide what you want, then use the tool to express it well.
A few practical approaches, drawn from the logic of the research:
Wait before you paste. When something irritates or puzzles you, give yourself even thirty seconds before reaching for an AI response. Thirty seconds of raw, uncomfortable, half-formed feeling is the substrate that conviction grows from. Skip it and you'll mistake the draft for the thought.
Ask for objections, not defenses. If you're genuinely uncertain about a position, ask the AI to argue against your own initial lean rather than defend whatever direction you first mentioned. This at least forces an encounter with the other side of the argument.
Treat the output as someone else's opinion. This sounds basic. It isn't. The output is a fluent, persuasive, confident-seeming text generated by a system that models your preferences rather than understanding them. It's your colleague's opinion — someone who knows your style well but doesn't know what's actually at stake for you.
Notice the blankness. That feeling of not quite knowing what you think without a tool open? It's data. It's telling you something about where your practice of agency has gone thin. Don't paper over it with a prompt. Sit in it a minute. The goal is not to be efficient. The goal is to have a mind that belongs to you.
The Larger Stakes
What makes this more than a productivity concern is that it operates at scale. When millions of people practice their reasoning through the same narrow set of AI-mediated pathways, the diversity of genuinely independent positions in a culture thins. You get a convergence that looks like consensus but is actually a shared dependency — everyone's arguments generated by models trained on overlapping data, all of them confident, none of them quite anyone's.
The interconnected web of hybrid environments reshaping human agency and society isn't a future condition. It's the current condition, experienced one email draft at a time.
The question isn't whether AI can think for you. It can't. The question is whether you'll still be able to think for yourself after you've spent a decade letting it try.
