Why using AI tools leaves so many of us stuck in an uncomfortable middle — and what four psychologists can tell us about the way we manage that discomfort.
There’s a feeling a lot of us get when we sit down to use an AI tool. You ask it for help, and a few seconds later you’ve got a draft, a summary, an answer that would have taken an hour on your own. There’s a little rush of amazement. And then, right behind it, a second feeling: should I be doing this? That second feeling doesn’t go away. I’ve been sitting with it for well over a year, and I think most people who use these tools are sitting with some version of it too.
I heard someone on the tech podcast This Week in Tech use the phrase “mixed feelings” about AI, and it stuck with me, because that’s exactly the right phrase. There’s a spectrum. On one end are people strongly opposed to AI, and they have real reasons. On the other end are people who wave the whole thing away. But most of us live somewhere in the uneasy middle: we use these tools, we like these tools, and we feel a little bad about using these tools. That middle turns out to be an interesting psychological place to stand.
Before getting to the psychology, it helps to name what’s actually bothering us instead of leaving it as a vague cloud of guilt. There are at least four concerns. First, every time you use one of these tools instead of hiring a person, there’s a possibility that someone didn’t get that work — the illustrator, the writer, the researcher. Second, these tools run in enormous data centers that use a staggering amount of energy. Third, those same data centers use a great deal of water to cool the computers down. And fourth, when you use one of these tools, you’re using everything it was trained on — writing, art, music — most of which the companies that built the tools almost certainly never paid for.
That fourth concern stops being abstract for me, because I’m a content creator myself. I have a number of videos on YouTube, and I make a small amount of money from them each month. That income exists for one reason: people watch the videos and see the commercials. A lot of what I say in those videos has almost certainly been scraped to train these models. So when someone asks an AI a psychology question I happened to answer in a video, the AI just gives them the answer — they never watch, never see the commercial, and I never earn that small amount. The tool takes what I made, gives it away, and removes the one modest way I was paid for it. Which puts me on both sides of this at once: the user feeling guilty, and one of the people these tools quietly take from. If it’s happening to me, it’s happening to a lot of people who have no idea it’s going on.
Cognitive dissonance: the attitude bends to fit the behavior
The term most people reach for here is cognitive dissonance, which comes from Leon Festinger’s 1957 book A Theory of Cognitive Dissonance and a well-known 1959 study he ran with James Carlsmith, published in the Journal of Abnormal and Social Psychology. Most people have heard the phrase but don’t remember what the research actually showed, and the research is where it gets interesting.
Participants were asked to do a genuinely boring task — turning pegs on a board for an hour. Afterward, they were asked to tell the next person waiting that the task had been fun. Some were paid twenty dollars to tell that little lie; others were paid one dollar. Later, everyone was asked privately how much they had actually enjoyed the task. The surprise is that the people paid one dollar were the ones who said they had enjoyed it. If you were paid twenty dollars, you had a clean reason to lie — you did it for the money. One dollar isn’t enough to justify lying to someone’s face, so those participants were stuck holding two thoughts that don’t fit: “I’m an honest person” and “I just lied for basically nothing.” That discomfort is the dissonance, and the mind resolves it by quietly deciding the task wasn’t so boring after all.
Notice what moved. The behavior didn’t change — the lie was already told. The attitude changed to fit the behavior. That’s the piece worth carrying over to AI. Most of us aren’t carefully deciding whether to use these tools and then concluding it’s fine. We’re already using them. That’s the fixed point. So the real question isn’t whether we’ll use AI; it’s what story we’ll tell ourselves so we can keep using it and still feel like decent people.
The menu of stories — and trivialization
There’s a whole menu of those stories. You can add reassuring thoughts: the legal questions will get sorted out in court, everyone’s using these now, the technology is here and there’s no putting it back. There’s even research on one specific version of this move. Simon, Greenberg, and Brehm, writing in 1995 in the Journal of Personality and Social Psychology, described a strategy called trivialization: instead of arguing with the uncomfortable thought, you shrink it. The energy one search uses is nothing next to a cross-country flight. Sometimes the comparison is even sort of true — but its psychological function is to take a thought that was bothering you and make it small enough to ignore. It’s worth being honest about the difference between weighing something fairly and simply shrinking it so it stops nagging.
Moral disengagement: how decent people take part in harm
Cognitive dissonance is about internal consistency — my own thoughts not matching. But some of the discomfort with AI isn’t about me at all; it’s about the possibility that I’m part of something that harms other people. For that, Albert Bandura’s idea of moral disengagement fits better. In a 1999 paper in Personality and Social Psychology Review, Bandura asked how fundamentally decent people take part in things that hurt others without feeling like bad people, and he laid out the mental maneuvers that let us do it. Two of them describe AI talk almost perfectly. One is diffusion of responsibility: when millions of people are doing something, no single person feels responsible, so my contribution feels like a rounding error. The other is euphemistic labeling: we say a model was “trained on” content, which sounds clean and technical. Say instead that it “copied” a few hundred thousand books and images nobody paid for, and the feeling changes. Same event, gentler word, quiet conscience.
Kohlberg: which floor are we reasoning on?
Lawrence Kohlberg’s stages of moral reasoning add one more angle. Kohlberg cared less about what people decide is right and more about the reasoning behind it. At the preconventional level, morality is essentially about punishment — something is wrong if you get in trouble for it. For most casual AI users, the honest answer to “will I get in trouble?” is no, so it feels fine. But consider a student: there really can be a punishment — a failing grade, or failing the course — so a student reasoning at that same preconventional level might avoid AI purely because they could get caught. Same level of reasoning, opposite behavior, depending only on whether a punishment is waiting.
At the conventional level, something is right if the group approves and everyone’s doing it — and a great deal of AI talk lives right here: it’s not illegal, everyone uses it, it’s normal. At the postconventional level, you reason from principle rather than from punishment or popularity, and interestingly you can land on either side. A principled “no” might say: given the harm to artists and writers, and the energy and water costs, I won’t use these tools, regardless of what’s legal or common. A principled “yes” might say: these tools can put tutoring, research, and translation help within reach of people who could never afford a human expert, so refusing to use AI is a comfortable stance for those who already have access to everything, and the fairer path is to build these tools and get their benefits to people who’ve been shut out. That “yes” isn’t “everyone’s doing it” — it’s an argument from fairness and access, which is what makes it postconventional. The honest observation is that most of our conversation about AI, mine included, sits down at the conventional level rather than up at the level of principle.
Being skeptical of my own episode
It would be easy to walk away thinking the psychology has solved the problem. So here’s a worry: putting the label “cognitive dissonance” on this feeling might itself be a way of feeling better. If it’s just a normal quirk of the human mind, then the discomfort isn’t a moral signal to act on — it’s just a glitch. That’s soothing, and it’s the very same move the whole discussion has been describing. Even the psychology can become one more way of getting the discomfort to leave us alone. And one more honest thought: many of the people worrying loudest about AI are knowledge workers, and these tools point straight at what we do. So when the twinge shows up, is it concern for an illustrator we’ve never met, or is some of it anxiety about our own relevance dressed in nicer clothes? Hard to say for certain.
Local models: a partial answer
One thing I actually did about all this was look into local LLMs. Most AI tools run in the cloud: you type something, it travels to a company’s data center, and the answer comes back. A local model runs on your own computer instead — you download it, and nothing goes out over the internet. The advantages are real. Privacy is the big one: nothing leaves your machine, so nobody stores your input or trains on it. It also lowers the environmental cost, since you’re not pinging a giant data center with every question, and there’s no subscription and no handing your material to a company.
But the disadvantages don’t vanish. The model on your laptop was still trained on the same content that was probably never licensed, so the core ethical problem is untouched. Local models are generally less capable than the big cloud models, so you’re trading quality for principle. Running one well takes a fairly powerful computer, which carries a hardware cost and an e-waste angle. And running your own machine hard still uses energy — much less, and moved from the data center to your own electric bill, but not zero. Claiming local models use no energy would be exactly the kind of trivializing the discussion warns against. So the scorecard is honest but incomplete: privacy and much of the environmental cost improve, and the hardest problem sits there unchanged.
Where this leaves me
It leaves me in that uncomfortable middle. I’ll probably keep using these tools, because they’re genuinely useful. I think there’s something real in the idea that using a tool for my own work isn’t putting a specific person out of a job — though, having just told you my own story, I also know these tools do take from specific people, because I’m one of them, so I hold even that comforting thought loosely. My tentative landing is that the discomfort isn’t a problem to solve quickly. It might be doing something useful: as long as it’s there, the ethical questions stay alive. The moment the story becomes airtight and comfortable is the moment we stop paying attention. So maybe the goal isn’t to make the mixed feelings disappear, but to stay a little uncomfortable on purpose and resist resolving things too fast in either direction. That’s where a year of thinking has left me, and I hold it loosely. I’d genuinely like to hear where you land.
The Psychology of Mixed Feelings About AI
Click any box to read what it means. The concerns create the discomfort; the frameworks explain how we manage it; the response is where the episode lands.
The concerns
What actually bothers us — the source of the discomfort.
The frameworks
Four psychological lenses on how we handle that discomfort.
The response
What we do about it — and where the episode lands.
Select any box above to read about it.
Key terms for students
Cognitive dissonance. The mental discomfort of holding two inconsistent thoughts at once, especially when a behavior conflicts with a value. Because the behavior is often already done, people tend to reduce the discomfort by changing the attitude to fit the behavior rather than the other way around. Festinger and Carlsmith’s 1959 study is the classic demonstration.
Trivialization. A specific dissonance-reduction strategy identified by Simon, Greenberg, and Brehm (1995): rather than changing the behavior or disputing the troubling thought, a person reduces the thought’s importance until it can be ignored. Watch for it whenever a real concern is answered only by comparing it to something worse.
Moral disengagement. Bandura’s term (1999) for the mental mechanisms that let ordinarily decent people take part in harm without feeling guilty. Two relevant examples are diffusion of responsibility (no one feels responsible when many are involved) and euphemistic labeling (softer wording that makes an act feel cleaner than it is).
Kohlberg’s levels of moral reasoning. A developmental framework focused on the reasoning behind moral judgments. At the preconventional level, right and wrong track punishment; at the conventional level, they track social approval and rules (“everyone’s doing it”); at the postconventional level, they track self-chosen principles such as fairness — which can support either using or refusing AI.
Local LLM. A language model that runs on your own device rather than in a company’s cloud data center. It improves privacy and reduces the cloud’s energy and water footprint, but it does not resolve the unlicensed-training-data problem and comes with quality and hardware tradeoffs.
References
Bandura, A. (1999). Moral disengagement in the perpetration of inhumanities. Personality and Social Psychology Review, 3(3), 193–209. https://doi.org/10.1207/s15327957pspr0303_3
Festinger, L. (1957). A theory of cognitive dissonance. Stanford University Press.
Festinger, L., & Carlsmith, J. M. (1959). Cognitive consequences of forced compliance. The Journal of Abnormal and Social Psychology, 58(2), 203–210. https://doi.org/10.1037/h0041593
Kohlberg, L. (1981). The philosophy of moral development: Moral stages and the idea of justice. Harper & Row.
Simon, L., Greenberg, J., & Brehm, J. (1995). Trivialization: The forgotten mode of dissonance reduction. Journal of Personality and Social Psychology, 68(2), 247–260. https://doi.org/10.1037/0022-3514.68.2.247