Is Gen Z less intelligent? No - and that is the wrong question.

In January a neuroscientist told the U.S. Senate that our kids are less cognitively capable than we were at their age, and that the screens in their classrooms are the reason. The headline did what headlines do. By February “Gen Z is officially less intelligent than Millennials,” circulated widely, and by February 5 the same scientist, Jared Cooney Horvath, had written an essay saying plainly that the data cannot support that position. Both things are true at once: the measurements are falling, and the claim that was popularized is not the one the data or the scientist supports.

Here are the details. PISA 2025, released in September, recorded the lowest OECD averages in reading and mathematics it has ever measured, and the share of fifteen-year-olds who rush through a text and answer wrongly nearly doubled since 2018. The IQ study Horvath leans on, on 98 sixteen-year-olds, found small declines in working memory and verbal comprehension whose confidence intervals include zero, beside gains in fluid reasoning and processing speed whose intervals do not. Something is moving. What is moving is not “intelligence.” Horvath’s own word for it is school-ability: the kind of thinking schooling was built to produce, and the kind his data say schooling is producing less of, in students who spend more years in school than any generation before them.

That is the gap I want to address, because it is a familiar one to me.

Ten years ago I supported instructors in a university computer science department, and I spent a year reading what the research said about why students learn. One line from that reading is significantly relevanty: learning is not a passing of information from an expert to a student, it is a human interaction based in relationship. I also mapped a course redesign at George Mason University that took cheating in an introductory programming course from one student in ten to none. They did not add surveillance. They separated practice from assessment. Students practised together, as much as they liked, and sat a supervised test whenever they were ready, with five tries. When the assignment stopped being the grade, the reason to fake it vanished.

That was 2015, before anyone had a chatbot. Now put a chatbot in the room and delivery amplifies.

If what a school rewards is a polished answer, AI will produce a polished answer, and the student can satisfy the requirement without the understanding, and with almost no effort. That is not a new weakness in students. It is an old weakness in what we have been treating as evidence of thinking, made visible because the artifact can now be produced without the thought – at speed and volume. If what a school rewards is the ability to explain, to defend, to apply the idea to a problem nobody showed you, the same AI becomes something else: the library that lets a curious student reach more and better connections than any library did. The tool is the same. The expectation decides what it becomes.

So I would put the question this way. Not whether a generation is losing its mind to screens, but what we have been accepting as proof that a mind was used and active thinking was involved.

I think the thinking lives in the questioning, the process, the exchange, and where it ends up. I explore this in a article responding to Aliah Davis McHenry’s article “The cost of Sameness” addressing AI’s affect on visual identity and marketing. My article https://www.linkedin.com/pulse/polish-used-proof-alex-dimarco-ybhee/ Polish Used to be Proof applies to learning as well as professional ability. If a machine can create a rapid polish of mediocre content without any engaged thought, a finished essay shows very little of the thinking involved. A student arguing with a machine for an hour, challenging it, being challenged, changing their mind twice and holding their ground once, may have done more intellectual work than the essay ever showed, and may also have done none, because an exchange can be faked as easily as an essay. Both are true, and the second one is the objection I most want teachers to answer. A learner who holds nothing alone cannot tell a confident wrong answer from a right one, and a student who lets the machine supply all the friction is not collaborating. They are being led.

Horvath gives us the right instrument for that worry, and it is a clear view from my perspective. He divides tasks into those where the structure is supplied and you only have to recognize it, and those where you have to build the structure yourself, and his data say the second kind is where the decline sits. There clearly are grades within this as structure itself can be a comprehensive and interactive engagement, and that refocuses the question to what is being observed and evaluated, The question for an AI in the classroom is then exact: is it the scaffold that does the structuring for the student, or the partner that makes the student do it? The one trial that tested this, with a thousand high-school mathematics students, found that unguided access to a chatbot raised practice scores by nearly half and lowered the unassisted exam by seventeen percent, while a tutor built to withhold answers removed most of the harm. Which scaffold the machine is depends on how it was built and what the student is asked to do with it. That is testable, in a term, in a real classroom, by the people who already run them.

Underneath all of this is an older argument that we keep having without naming it: what education is for. In 2015 I drew a map of the purposes of an undergraduate degree, vocational, liberal, research, entrepreneurial, and I left one out that I would add now, the civic one, the capacity to look at the institutions you live inside and ask whether they are working ( https://www.cs.toronto.edu/~alex/images/Undergraduate-teaching-goals-exploration.svg ). Functional preparation is a legitimate purpose. So is the capacity to question, discover and create. The trouble starts when the first becomes the only thing we measure and the second is still what we claim to produce. A generation taught to produce the expected answer is being prepared to work inside existing knowledge. A generation taught to challenge the answer is being prepared to extend it.

I worked this through over one morning with three AI models from three makers, each working under a method I have been developing for making a machine’s claims answerable, and the whole exchange is on the record at conversationswithlimen.ca, with every claim dated, every source marked as reached or not, and a ledger that says who contributed what, including the two models that got Horvath’s distinction backwards. That page is where the details live. It is not where the answer lives, because the people who can answer are the teachers and educators working in classrooms.

If you teach, you are living in the chaos of our present situation as see what is happening. Which of these matches your classroom, and which one does it contradict? What do you already do that makes the artifact stop being the grade? Where does the societal structures and your job structure you into teaching in a way that limits learning. What is one check you could run this term with what you already have that could change the situation and improve outcome? Say so here, or send it to the record, where it will be kept beside everything the models said, under your name or anonymously under your title, with the same standing. The machines have had their turn. The next entry in the ledger should be a teacher’s.