Why don’t students use AI-based tutor-bots? Holt calls this the 5% problem: only about five percent of Khan Academy students use the platform enough for gains to show up on standardized tests, and that five percent skews toward higher-income, higher-performing kids, who would have done fine anyway. Ben Riley, over at Cognitive Resonance, names a second pattern that he calls the 50% problem: of the students who do take up AI freely, something like half end up outsourcing their thinking to it. As a study showed, homework scores go up, exam scores (where no AI is available) go down. So the students who stay away from these tools get nothing from them, and a good number of the students who dive in come out worse.
How do we explain this?
Danah Henriksen and I have a new piece out in TechTrends arguing that part of the explanation sits in the way we (humans) interpret the language these systems produce. We call the problem the hollow signals problem.
Essentially it comes down to something subtle about how we psychologize engaging with other people, and how we process the feedback we receive, and how that changes when the interaction is with a machine. Here are two stories from graduate school that capture how complex this really is.
Story 1: As a grad student I met regularly with a professor, who shall remain nameless, around some research we were doing together. Some weeks I came in with real work. Other weeks I was tired and brought something I had assembled mostly so as to have something to put on the table. What struck me was that he was always effusive, irrespective of what I turned in. Praise, warmth, encouragement, the whole package. And when I got praised for what I knew to be average work, I remember wondering: does he think I am not capable of much, or does he want something from me? Neither question had anything to do with the work I had actually done.
Story 2: Another professor I worked with was Bill Brewer, who was very different. Once, on a paper I had not worked hard enough on, he told me plainly that I had not done the heavy lifting the problem required. That moment, and the sting of embarrassment that came with it, has stayed with me for decades. But underneath the shame was the sense that I had let him down, and you can only let someone down if they expected something of you to begin with. He could have let it pass, and it would have cost him nothing to let it pass. But he didn’t, and that is why I still remember it. He was invested in me, and calling me out was how that showed.
In both cases I was reading past the words to what the person must think of me. In some ways THAT was the message, whether intentionally sent or not. Wulf-Uwe Meyer and his colleagues showed this experimentally back in 1986, where blame after failure on a hard task raised people’s estimates of their own ability, on the reasoning that the evaluator would not have bothered unless he thought better of them. Praise for an easy success got discounted, or was even read as a small insult.
I ended up running a version of that study myself, with the human evaluator replaced by a computer, and published it in 2006. And the results were interesting, similar in one way to the human-human situation but different in another. So, praise from the computer felt good and blame from the computer felt bad. No real surprise there. But what did not translate across was the deeper reasoning. They just took the feedback at face value, with none of the inferencing about self that participants had done in the human-human case.
What this meant was that the second level of processing, about the intentionality of the person giving the feedback, their investment (or lack thereof), and what it said about them, just did not happen. That investment came in the form of attention. The first professor did not give me or the work genuine attention, but Bill did. As Simone Weil wrote: attention is the rarest and purest form of generosity. Generosity is exactly the right word, because attention costs something.
Evolutionary biologists discuss this arguing that it is the cost that makes a signal a honest signal. A signal carries information because it is expensive to produce. The peacock’s tail is credible evidence of fitness because growing it burns scarce calories and dragging it around makes the bird slow and conspicuous. A weaker bird would not be able to pull that off. The plumage is an honest signal of a bird’s awesomeness as a potential mate.
This applies to teaching as well. When a teacher stays after class for the third explanation of the same idea, reaching for a new analogy each round, the student can see the “cost,” in terms of time and attention that the teacher is taking on. This is not something they have to do but here they are.
Bill did not have to tell me the blunt truth about what he had expected from me. He took a social risk, and I interpreted it to mean he thought highly enough of me to be willing to do so.
What my 2006 study showed was that we may not do that same level of deep thinking when it comes to feedback from computers. And maybe that can explain the 5 and 50 percent paradoxes.
Danah and I begin the column with that line from Weil, and go on to argue that there is no social cost that GenAI is perceived as paying when it interacts with us or gives us feedback. Patience is free when there is no fatigue to overcome. Fluency is what the architecture is optimized for. Warmth is what reinforcement learning from human feedback selects, since warmth rates well with users. Infinite patience, as it turns out, is indistinguishable from infinite indifference.
So if the peacock’s plumage is an honest signal, the outputs of AI are hollow signals. They have the surface look and feel of human communication with none of the costs underneath.
My friend Andrew Maynard has argued that large language models slip past the ordinary machinery we use to evaluate what we are told. He calls these cues honest non-signals, and I am not sure I agree with that term. AI systems are neither honest nor dishonest. That is not a scale on which they operate. They are not non-signals either, since they function as signals perfectly well. Hollow signals was what I felt was more appropriate, capturing the idea that though the form may look the same, there is nothing inside it. Andrew and I have agreed to disagree on what to call it but we agree on the mechanism.
In the paper Danah and I do more than diagnose the problem. We argue for a way through it, which takes us back to Dewey and Papert. There is a good deal more in the article, and you have to read the whole thing to get the full picture.
This is the third article in our TechTrends series on AI and language. The first was on the metaphors we use for AI, the second on words in danger of losing their hard-earned meaning once the AI bros get hold of them, and this one on how we interpret what AI says to us. We did not know it would become a series, but all three turn out to be about the same thing, which is the slow erosion of the vocabulary and the infrastructure through which we recognize human agency in learning.
Read it here, and tell me where you think we have it wrong.
Mishra, P., & Henriksen, D. (2026). Hollow signals: Why AI’s best features may be its biggest problem for education. TechTrends. https://doi.org/10.1007/s11528-026-01227-9
Meyer, W.-U., Mittag, W., & Engler, U. (1986). Some effects of praise and blame on perceived ability and affect. Social Cognition, 4(3), 293–308.
Mishra, P. (2006). Affective feedback from computers and its effect on perceived ability and affect: A test of the computers as social actor hypothesis. Journal of Educational Multimedia and Hypermedia, 15(1), 107–131.
Earlier on this blog: Honest Non-Signals: Why AI Fools Us Without Lying (February 2026)
Also: Why Sal Khan’t on Learning by Making but Teaching by Telling (April 2026)
Ben Riley’s Cognitive Resonance, on the 50% problem
Andrew Maynard, The AI Cognitive Trojan Horse: How Large Language Models May Bypass Human Epistemic Vigilance