What a Classroom Experiment Taught Me About AI and Human Connection

My neighbor James teaches high school English. He has been at it for nineteen years, which means he has seen every wave of technology come through the classroom and watched teachers either panic about it or pretend it does not exist. He is neither type. When AI writing tools became widely available two years ago, he did something interesting. He gave his students a controlled assignment where half the class wrote an essay themselves and the other half used AI to generate a first draft, then edited it however they wanted. Then he read all of them without knowing which was which.
He told me he could identify almost every AI-assisted essay within the first two paragraphs. Not because they were badly written. Several of them were technically better than the handwritten ones. He could tell, he said, because they had no point of view. They covered the topic thoroughly, considered multiple perspectives, and reached a balanced conclusion. Competent in every visible way. But there was nobody home. Reading them felt like listening to someone describe a painting they had never actually seen.
That conversation kept coming back to me as I started looking more seriously at what it actually means to humanize AI and why so many smart people in technology consider it the defining challenge of this particular moment. James was not describing a problem of grammar or logic. He was describing a problem of presence. And presence, it turns out, is extraordinarily hard to engineer.
The Problem Is Not What Most People Think It Is
Ask someone unfamiliar with AI development what it would take to make AI more human, and they will usually say something about emotions. Teach it to feel things. Give it empathy. Make it understand sadness and humor and sarcasm. That sounds reasonable on the surface. It is also, more or less, the wrong frame for the actual problem.
The researchers working seriously on this are not trying to give machines feelings. That road leads into philosophical territory that nobody has figured out and that would not necessarily solve the practical problem anyway. What they are trying to do is something more specific and, in some ways, more interesting. They are trying to build systems that respond to the full content of what a person communicates, not just the informational layer of it.
When James’s students wrote their own essays, even the weaker ones had something in them: a particular way of approaching the question, a moment of genuine confusion that showed up in the writing, an argument that was slightly off but clearly the student’s own. Those things are not decorative. They are actually how communication works between people. We signal our perspective constantly, through the examples we choose, the things we emphasize, and the questions we are clearly still working out. AI systems trained to produce polished, comprehensive output tend to sand all of that away. The result is technically smooth and humanly inert.
Why the Technology Industry Underestimated This
There is a reasonable explanation for how the AI industry ended up here. The early benchmarks for AI performance were built around tasks where the right answer is relatively clear: translating text, identifying objects in images, playing chess, and answering factual questions. These are problems where more data and better algorithms produce measurably better results, and the field got very good at optimizing for them.
Human communication is not that kind of problem. There is rarely one right answer to the question of how to respond to someone. The best response depends on who they are, what they are actually trying to accomplish, what kind of conversation this is, what has already been said, and dozens of other variables that shift constantly. Training a system to produce responses that score well on comprehensiveness and accuracy does not automatically produce responses that feel right to the person receiving them. Those are different targets, and chasing one does not get you to the other.
The industry is catching up to this. Slowly, in some places. More quickly in others. But the shift in thinking is real: the question is no longer only what can AI do, but how does it feel to interact with it, and why does that difference matter as much as it clearly does?
Real Consequences in Real Situations
This is not an abstract problem. The places where AI is being deployed most aggressively right now are places where the human quality of the interaction carries serious weight.
Healthcare is an obvious one. AI tools are being used to communicate with patients about test results, treatment options, medication schedules, and follow-up care. The accuracy of that information matters enormously. But so does the tone, the pacing, the acknowledgment that receiving a difficult diagnosis is not the same as receiving a software update. Patients who feel processed rather than cared for disengage. They miss appointments. They do not ask the questions they need to ask. The human quality of those AI interactions is not a nice extra. It is clinically relevant.
Education is another. AI tutoring tools are proliferating rapidly, and the evidence on their effectiveness is genuinely mixed. The systems that seem to work best are not the most comprehensive or the most knowledgeable. They are the ones that adapt to how a specific student is responding in real time, that recognize when someone is frustrated versus genuinely stuck versus just moving too fast, and that adjust accordingly. That kind of responsiveness is a form of humanization, even if it does not involve anything resembling emotion.
Customer service, mental health support, legal guidance, and financial advice: the same pattern shows up across every domain where AI is taking on roles that used to belong exclusively to people. The technical capability is often there. What separates useful from alienating is whether the system can engage with the human dimension of the interaction.
The SEO and Content Question Nobody Wants to Ask Directly
Here is something worth saying plainly for anyone working in content or digital marketing. The readers your content is trying to reach are getting better at detecting inauthenticity. Not through any formal analysis. Through the same instinct James had reading his students’ essays. Something feels off and they leave. They do not always know what it was. They just know they did not connect with it.
This is already showing up in performance data for organizations that have moved aggressively toward AI-generated content at scale. Impressions stay flat or improve because the content is optimized for search. Engagement drops because the content does not hold people once they arrive. That gap between traffic and engagement is one of the clearest signals that humanization matters, practically and measurably, not just philosophically.
The push to genuinely humanize AI writing tools is therefore directly relevant to anyone whose job involves producing content that real people are supposed to read and care about. Better tools are coming. In the meantime, the organizations doing best are the ones pairing AI capability with human editorial judgment, using the former for efficiency and the latter to make sure there is still somebody home in the final product.
Back to James
I asked James recently whether any of his students had produced AI-assisted work that genuinely fooled him. He thought about it for a moment. He said yes, one. A student who had used an AI draft but then rewritten it so thoroughly in her own voice that it read as entirely hers. He only found out because she told him afterward, curious whether he had caught it.
What she had done, without using any of this terminology, was humanize the output. She took something technically capable and put a person back into it. That is, at a very human scale, exactly what the whole field is trying to figure out how to do automatically. It is harder than it sounds. It might be the most important thing happening in technology right now. And a nineteen-year English teacher with a pile of student essays understood the problem more clearly than a lot of people who have been working on it in labs for years.