I want to argue that the collapse of the salary pitch is not a crisis for college. It’s a clarifying event. Because if the only thing you can say for four years and a small fortune is “you’ll earn more,” then when that stops being reliably true, you have nothing left. You built your whole defense on a number, and the number moved. We can trace this line of reasoning back to the Coleman Report in the 1966.

Here’s my claim, stated plainly. The case for college was never really the paycheck. We just let the paycheck do the talking because it was the easiest thing to measure. And the moment we let earnings become the entire justification, we quietly conceded that college is a vending machine: insert tuition, receive salary, and if the salary doesn’t come, demand a refund. No wonder people want their money back. We taught them to see it that way.

But there’s an older answer, and the present moment makes it newly urgent rather than newly quaint. College, at its best, was never primarily a credential factory. It was the place you learned to do the thing machines now do worst: sit with a hard text, weigh a claim, find your own position, and defend it. Learn to read closely and think critically—to be, in a phrase, hard to fool.

Notice what’s happened to the value of that skill while the value of the credential wobbled. We have just flooded the world with confident, fluent, frictionless text that is wrong a meaningful fraction of the time. The single most valuable capacity a person can now have is the one that lets them tell the difference: to read a generated paragraph, or a politician, or a headline, and judge whether it holds. That capacity doesn’t come from a chatbot. It comes from exactly the slow, effortful, deeply uneconomical work that the salary pitch always treated as a side effect.

Which is exactly the habit we are losing, and we’re losing it on purpose. The AI summary is engineered to make the second read feel unnecessary. Why return to a dense chapter when a clean paragraph will hand you “the takeaway” in ten seconds? The pitch is irresistible, and the logic is fatal: it treats a text as a container of information to be extracted once and discarded. Get the point. Move on. Never look back.

But a real text isn’t a container. It’s a structure you understand differently depending on where you stand in it. The first read gives you the plot, the gist, the surface claim. The re-read gives you the architecture: how the argument actually holds together, where it’s load-bearing and where it’s bluffing, and what the author was doing that you were too busy following along to notice. Strip out the re-read, and you don’t get a faster version of understanding. You get a permanent version of the first read: the gist, forever, mistaken for the thing itself.

Here’s my claim, stated plainly. AI summaries do save time, but they also quietly delete the second encounter, and the second encounter was where the thinking happened.

The research keeps circling this. In a 2025 classroom experiment with 405 teenagers in England, published in Computers & Education, students preferred an AI chatbot and rated it as more helpful. Yet those who took notes the old way comprehended and retained more. The comprehension that felt effortful was the comprehension that stuck. The summary optimizes for the sensation of knowing while removing the labor that produces actual knowing. It is a machine for confident shallowness, and confidence without depth is precisely the failure mode a flood of generated text was always going to produce.

Read deeply,

Dr. Genevieve

I’ve met both students every year for a decade; they come from campuses across America.

We talk about the reading crisis as if it fell evenly, like weather. It didn’t. Only 39 percent of the ACT class of 2025 met the college readiness benchmark in reading, and the 2024 NAEP put twelfth-grade reading at its lowest average since the test began in 1992. Look underneath the average, though, because the federal data shows the split: the declines have been steepest among the lowest-performing students, while readers at the top held closer to their ground. The floor is falling faster than the ceiling.

Here’s my claim, stated plainly. The ability to read hard things closely is becoming a marker of class, privileged rather than earned. And we are alarmingly close to assuming that is natural.

A child in a well-resourced school still gets the slow work: whole books, discussion, and a teacher with room to teach reading as a craft. In contrast, a child in a school under relentless test pressure gets reading as decoding: fast, shallow, and optimized for a score. One approach is an apprenticeship to critical thinking, and the other is test prep, sold as the same thing.

We need to name the machinery carefully because it is a documented pattern. Two decades of test-driven accountability under No Child Left Behind and its successors distorted reading instruction in the schools that lived or died by the score. Teachers in those schools have reported the same shift for twenty years: short passages, multiple-choice comprehension, and reading redefined as answer retrieval. Schools insulated from that pressure kept teaching the slow, argumentative, what-do-you-think kind. Same word, two utterly different skills, sorted largely by zip code.

Now the complication is that I cited the evidence myself three weeks ago. I’ll clarify that argument here: the privileged tier is not thriving; it is decaying more slowly. The same Atlantic reporting that anchors this publication found students at elite colleges arriving “bewildered” by whole books. The difference between the tiers is not insulation, but what happens after the stall. When the well-resourced student stalls, someone notices and teaches her through it. When the under-resourced student stalls, the system hands her a shorter passage. Both tiers are losing depth. Only one is offered the way back.

By the college seminar, the divide is invisible and total. One student offers a reading of the text; the other waits to be told the answer, because that is the only kind of reading she was ever trained to do. We call the first student “sharp” and the second “unprepared.” We’re describing a curriculum, and mistaking it for a person.

Now add AI. Label this part prediction rather than record. The machine will not arrive as the great equalizer; it will arrive as an accelerant. The student trained to read deeply will use it the way a strong reader uses anything, as a sparring partner to argue with more effectively. The student who was never taught will use it as a replacement for the reading she cannot yet do, and the muscle atrophies before it forms. Educational tools can only amplify the skills we teach.

A democracy runs on people who can read a claim, weigh its evidence, and evaluate the reasoning that connects them, for example, a ballot measure, a news story, a chatbot’s confident paragraph. If that capacity becomes the property of one tier, we don’t just get unequal students. We get unequal citizens: a class that interrogates the information environment and a class that can only be moved by it. That is not an AI literacy problem (although, to be fair, it has its place). The problem is power wearing a literacy costume.

I don’t have a tidy fix, and I distrust anyone who offers one. But the first step is to stop calling this a talent gap when it is an access gap: to time, to teachers, and to the slow apprenticeship that turns a person who can read into a person who can read deeply and think critically. We decided, district by district, who would get that apprenticeship, and we can decide differently.

What we cannot do is keep pretending the tiers are about who’s smart when they’re really about who was taught. And that is a social choice, not a fact of nature.

A question for the comments: Where did you learn to read deeply, and who in your life never got that chance?

Read deeply,

Dr. Genevive

As educators, we spent twenty years optimizing friction out of school. Shorter readings, smoother interfaces, answers just a search away. Remove the obstacles, the thinking went, and learning flows. I won’t pin the whole decline on the smoothing; phones, a pandemic, and two decades of test policy left their marks, too. But the record is in, and it isn’t kind. Students can decode every word and yet lose the thread of a long argument. They arrive at college, Rose Horowitch reported in The Atlantic in 2024, “bewildered” by the expectation of finishing whole books. (The word belongs to Columbia’s Nicholas Dames.) Does this bother you as much as it bothers me?

So, a counterculture is forming around one insight: some friction is not an obstacle to learning; some friction helps to produce the learning. A paper book is harder to skim, which makes it easier to actually read. Reading aloud is slower, so the difficult sentences and passages can’t be skipped. Working through a text together makes confusion public, and public confusion is survivable. For PublicSource, Jamese Platt profiled Pittsburgh faculty who read passages aloud, line by line, and have students write their own questions before class begins. No one can hide in the silence; that’s a skillful design.

Cognitive science got there thirty years ago. In 1994, the UCLA psychologist Robert Bjork named the pattern desirable difficulties, and the lab results since have leaned lopsided in one direction. Studying that feels easy, rereading the highlighted page, produces memory that fades by Friday. In contrast, studying that feels like a struggle, generating an answer from scratch, wrestling with a passage before being told what it means, produces understanding that stays. Ease is a feeling, not a learning outcome. The students who feel most fluent quickly are often learning the least.

The methods are more precise than paper, good; screens, bad. At the University of Pittsburgh, the sociologist Hillary Lazar pairs stronger and weaker readers in what she calls a one-room schoolhouse, so that comprehension happens out loud, between people, rather than failing silently in private. Call it what it is: the exact effort that reading requires, measured out, and stirred back in by hand.

Here’s my claim, stated plainly. The effort was never the bug; it was the mechanism. Take the difficulty out of reading, and we don’t get easier learning. We get the look of reading without the cognition that made it worth anything.

One objection deserves a straight answer because, no doubt, friction can be a barrier. For instance, the dyslexic reader needs the audiobook, and the working parent needs the flexible format. The boundary that matters runs between the friction in access and the friction in the text. Clear the first completely, but guard the second, because the productive struggle is what helps us learn.

And one caveat the movement’s admirers tend to skip: nobody has published outcomes data on these classroom practices. There are decades of laboratory evidence for the principle, a handful of professors applying it on purpose, and data I’m watching for, but that is not here yet.

Still, I find this counterculture hopeful—and I don't get to say that often on this beat, where most of what crosses my desk is decline. The story of rediscovering deep reading is a bright spot precisely because it isn't a platform. It won't scale or break like one, either. Good. Let the tech titans go target something other than a core function of civil society.

Deep reading was never going to come back by removing effort; it comes back by designing the right kind of effort, and perhaps more importantly, by valuing that effort. The professors who went old-school looked most closely at where we are and decided that the way forward runs through the difficulty, not around it. Even though the ingredients look like 1990, the recipe is the most forward-looking thing on campus: a paper book, a slow voice, and a classroom where nobody can skip the thinking.

A question for the comments: what’s one piece of useful friction, in reading or anything else, that you’d put back if you could?

Read deeply,

Dr. Genevive

Recently I ran a genAI’s answer to “Is college still worth it?” through CERIC and watched a confident paragraph fall apart under five questions. A reader wrote back with a fair challenge: I still prompted the machine to get that paragraph. Doesn’t that just prove we need to use the tool better?

It’s a reasonable position, and it’s everywhere. The whole AI-literacy industry is built on it—the workshops, the certifications, “prompt engineering” as a résumé line. Learn the right incantations, the pitch goes, and the tool will serve you. I won’t pretend I’m above it; I’ve been tuning prompts for years, especially as the models turned into decent research assistants. But I want to make the case for the opposite skill, the one almost nobody is teaching: knowing when not to prompt at all.

Start with what a prompt costs. Not money; most chat boxes are free. The cost is hidden in the moment before the answer arrives, that interval when you don’t yet know, when you’re holding a question open and turning it over, feeling mildly uncomfortable and slightly lost. The interval feels like a problem to solve as fast as possible. It is, in fact, where the learning lives. A mind gets sharper and more curious inside the not-knowing. The prompt is a command that ends it early.

This is why the deep-reading parallel is exact. A hard paragraph and an unanswered question are the same kind of object: both ask you to stay inside the difficulty long enough to be changed by it. The genAI summary and the prompt are the same kind of escape: both hand you a clean result while quietly soothing the reason you came. Ask the machine, and you get the conclusion without the journey that was supposed to build the muscle to reach it.

So here is the line. Real AI literacy isn’t better prompting; it’s the discernment to know whether you should be writing the prompt at all, whether this is a task worth outsourcing or a thought worth holding.

The rough rule: outsource the work that has a correct answer and asks nothing of your mind—for example, formatting a citation, organizing a list, hunting down more sources, or expanding a brainstorm you’ll curate anyway. Refuse to outsource the work whose whole value is the doing—such as evaluating an argument, finding your own position, or reading an essay that’s hard precisely because reading it changes you. When the correct, factual answer is the point, prompt away. When the thinking is the point, the prompt is theft, and you are the one being robbed.

Imagine where that theft happens. A student is assigned a page of Kant—it’s dense, recursive, and genuinely hard—and trying to read it the night before an essay is due loads the acute discomfort of not understanding quickly. The reading is a fork. One path: paste the page into the chatbot, get a clean response, feel the relief, move on, and retain nothing. It’s junk food for the mind. The other path: stay, reread, underline the verb, and argue with the sentences until they yield. It’s slow food for the mind. Both paths end the discomfort. But only the second ends it by teaching the reader to think. The machine stands at that fork offering the exit, and the exit always looks like help.

Is this just dressed-up Luddism, refusing a tool because it’s new? I don’t think so, and the test is simple: a Luddite refuses the tool everywhere. I’m asking for refusal in one place, the place where the struggle is the product. Use genAI to book the flight, debug the script, or summarize the meeting you missed. Refuse it where the wrestling was the whole point. That isn’t technophobia; it’s knowing what a tool is for, which is the most pro-technology stance there is, and the most pro-human.

So build the habit nobody taught us, because nobody had to: taking a half-minute pause before the prompt to ask one question. Am I looking something up—what stinging jellyfish live near Hawai‘i’s beaches—or am I trying to skip the deciding: should I take the job? The first is free. The second costs the exact skill no model can value.

My question is unfashionable because what I’m arguing for is slow and difficult, and slow and difficult are what the machine was built to abolish. But the people who actually understand anything five years from now won’t be the ones with the best prompts. They’ll be the ones who can still sit in the discomfort of not knowing, who can tell when the smart move is to close the tab and stay with the hard thing a little longer.

The real thinking work lies in exploring the not-knowing. Don’t let anyone sell you on the idea of pushing a button that skips it.

A question for the comments: where’s your line? What do you happily hand to genAI, and what would you never let it do for you?

Read deeply,

Dr. Genevive

It told me no. Summaries hand over the plot but bypass critical thinking, thematic analysis, and the slow comprehension that only comes from wrestling with the text. I agreed with every word.

Which is exactly the problem.

An answer that flatters what you already believe is the one you’re least likely to check. So I ran it through CERIC (Claim, Evidence, Reasoning, Implications, Context), the same five questions I’d want a student to put to anything a machine tells them. Here’s where a “correct” answer turns out to be hollow.

Claim

Watch how much it hedges. Summaries typically deprive students. They give the basic plot. And then, a few lines later, the same answer concedes that summaries can be “a useful supplement,” even helpful “for students with learning disabilities.” So which is it: a theft of learning or a study aid? The answer wants credit for the strong position and the safety of the soft one. A claim you can’t restate in a single sentence as declarative knowledge isn’t a claim; it’s a mood.

Evidence

This is where it falls apart. The answer arrives armed with six citations, which looks like rigor until you actually read them. One is a blog from a fax software company. One is a business that sells book summaries; asked whether book summaries are bad for you, it is not a neutral witness. One is the marketing page for Evernote’s own AI summarizer, which says AI captures fiction well—a source cited in support of a claim it directly contradicts. Not one is a study. The machine dressed a stack of content marketing in the costume of a literature review, and the footnotes are doing fake work.

Reasoning

Notice the structure: a tidy “What Students Miss” set against “What They Gain (At a Cost).” It’s a rigged scale. The losses are stated as established facts: reading “strengthens memory” and “builds analytical skills,” with no mechanism and no measurement. The gains are scare-quoted into suspicion, and the conclusion was loaded into the format before any evidence showed up.

Implications

Then comes the tell. After warning that summaries gut learning, the answer cheerfully offers to help me use AI “as a supplement” to generate discussion questions and clarify confusing chapters. The critique dissolves into an upsell for the very behavior it just supposedly diagnosed. If you took the answer seriously, you’d close the tool. The attention economy would prefer you keep it open.

Context

And it answers all of this about no book in particular, no student in particular, and no course in particular. A plot summary of an airport thriller and a summary of Beloved are not the same reading situation. What “the reading was meant to teach” depends entirely on what was assigned and why, and the machine, knowing none of that, answered as if it were one thing.

Here’s the part that should unsettle us. The students I’ve been calling the honor-roll offloaders are doing exactly what Gemini did here: producing a fluent, well-formatted, plausibly cited artifact that looks like understanding and was assembled without any of the work understanding requires. The summary and the summarizer share a flaw. The answer made some points I agreed with, and yet, it was still hollow. If you can’t tell those two things apart, the machine has already won an argument it never actually made.

That’s what CERIC is for. Not to reach the right conclusion, but to know whether the argument holds.

Read deeply,

Dr. Genevive

A 2026 preprint, “Patterns of Student Cognitive Offloading to AI in Higher Education,” analyzed 3,047 ChatGPT messages from 46 undergraduates and found that the heaviest use was not among the strugglers but among the high performers. The honor-roll students were offloading the most. (It’s a preprint, not yet peer-reviewed, and the sample is small, but the pattern is worth sitting with.)

This breaks the story we’ve been telling ourselves. The campus genAI panic has always cast the cheater as someone desperate: overwhelmed, underprepared, gaming a system they can’t otherwise beat. That student exists. But the data point somewhere more uncomfortable: for the best students, genAI dependence is rarely a desperate move and more often an optimization strategy.

And why wouldn’t it be? We built the incentives ourselves. We told a generation that the grade is the goal, that the transcript is the prize, that the point of the assignment is the artifact you hand in. A bright, rational student hears that message and does the obvious thing: she finds the fastest route to the artifact. AI is simply the fastest route yet. She understands the game perfectly.

There’s a particular cruelty in this for the conscientious student. She isn’t lazy. She’s overloaded: five courses, a job, the ambient dread that any slip will cost her the future she’s been told depends on this transcript. GenAI doesn’t tempt her the way it tempts the struggling student; it rescues her. It’s the only thing offering to give back the hours hyperachievement stole, and it asks for nothing except the part of the work that was invisible anyway. Who refuses that bargain? Only someone who was told what it actually costs, and almost certainly, no one has told her because the data are only beginning to emerge.

And it appears the cost is deferred, which makes it so easy to pay. The thinking she skipped this week shows no visible deficit. The essay is fine. The grade is fine. The deficit surfaces years later, diffuse and unattributable in the form of an adult who can generate a competent document and cannot, when it matters, tell a sound argument from a slick one. The bill arrives with no name on it.

Now look at the same problem from the registrar’s side, where it stops being anecdote and becomes a transcript. This spring, a University of California, Berkeley team analyzed 500,000 grades across eight years and found that in writing- and coding-heavy courses, the share of A’s rose about 30 percent after ChatGPT—and the jump was driven almost entirely by take-home work, not by proctored exams. As the study’s author put it, you now have “a C student who is now an A student.” The grade went up; the thinking didn’t. EDUCAUSE Review named the paradox in 2025: “better results, worse thinking.”

So here’s my claim, stated plainly. The honor-roll offloader is not a discipline problem; she is a measurement problem. We spent a century building an education system that grades outputs, because outputs are easy to grade and were reasonably hard to game. GenAI has revealed that a graded output is no longer evidence of anything happening inside a head, and now the transcripts say so at scale. Our best students were the first to notice. Of course they were; noticing is what made them our best students.

The institutions have noticed too, and their alarm is the tell. Princeton is ending 133 years of unproctored exams; Harvard’s faculty voted to cap A’s at a fifth of each class. Both are confessions that the transcript has stopped certifying what it claims to. But capping grades treats the symptom. The real question is the one the Berkeley author asks: do grades still tell us what students can actually do?

What we should be developing is reasoned judgment: the capacity to sit with an argument’s claim and reasoning and decide whether it holds. You cannot offload that to a machine, because the machine was trained on the internet, where reasoned judgment was never required to post. Outsource your thinking to ChatGPT, and you become more reliant; learn to interrogate what it hands you, and you become harder to fool. Same tool, opposite results.

The fix isn’t detection; that’s a tax on trust the cleverest students will always outrun. It’s assignments where overreliance carries a penalty, and the thinking is the deliverable: done in front of you, explained out loud, and graded explicitly, as the thing itself. If we make the thinking the deliverable, the shortcut stops being short.

Until we do that, the most capable people in the room will keep making the most rational possible choice. We won’t like where it leads them, and for now, the transcript won’t even tell us it happened.

A question for the comments: If your best students are the heaviest AI users—and the grade can no longer reveal who learned what—is that a cheating problem or a curriculum problem, and what would you actually change?

Read deeply,

Dr. Genevive

This is the third piece in a series that began when I argued the old college model is dead. Earlier this week I promised you a method for taking an argument apart. Here it is.

The method has a name: CERIC, which stands for Claim, Evidence, Reasoning, Implications, and Context. It was built for reading scientific papers, but it works just as well on anything a chatbot zaps out. In five simple elements, we can watch output come apart.

So I gave a free GenAI chatbot—the one built into Google search—the question millions of families are typing right now: “Is college still worth it today?” Here is its answer, verbatim.

Yes, a bachelor’s degree remains an excellent return on investment for most students. Graduates outearn non-graduates significantly, enjoy higher employment rates, and gain access to robust benefits like health insurance and retirement plans. However, the payoff depends heavily on your choice of major, school costs, and personal career goals. [1, 2, 3]

The value of a college degree hinges on several key factors:

The Financial Advantage

• The Earnings Gap: Bachelor’s degree holders typically earn over $1 million more in median lifetime earnings compared to those with only a high school diploma.

• The College Wage Premium: In California, the median annual wage for a bachelor’s degree holder is $96,000—double the median wage for high school graduates.

• Job Security: College graduates are significantly less likely to experience unemployment, even during economic downturns. [4, 6]

The Catch: Not All Degrees Are Created Equal

• Majors Matter: Fields like computer science, engineering, and nursing offer high starting salaries and strong demand. Conversely, majors in the arts or humanities may have lower median earnings, though they develop highly adaptable critical thinking and communication skills.

• Debt: Excessive student loan debt can quickly neutralize the financial advantages of a degree. Online consensus emphasizes the importance of minimizing debt by exhausting grants, scholarships, and considering community college transfer programs. [3, 8, 9, 10, 11]

When College May Not Be the Best Path

If taking on high debt for a general major doesn’t appeal to you, there are highly lucrative alternatives:

• The Trades: You can bypass student debt and earn excellent livable wages in skilled trades such as plumbing, welding, carpentry, or electrical work.

• Certifications & Academies: Police academies, fire science programs, and specialized tech bootcamps can launch you into high-paying, stable careers much faster than a traditional four-year program.

• Military Service: The armed forces provide career training, healthcare, and GI Bill benefits that can fund further education later in life. [13, 14, 15, 16]

If you decide to pursue higher education, maximize your return on investment by starting at an affordable community college to knock out general education requirements, and focusing on high-demand fields. Researching projected career demand and regional salary averages using the Bureau of Labor Statistics can help ensure your time and tuition translate into a strong financial and professional future. [3, 9, 17]

AI can make mistakes, so double-check responses.

C — What’s the claim?

Strip the hedges, and the claim is plain: college is a good financial investment for most students. A confident “Yes” up top, then the retreat. The payoff “depends heavily” on major, cost, and goals; a later section lists when college isn’t the path at all. The bold verdict is defended only until you press it, then falls back to ground that can’t be wrong—a motte-and-bailey, for you logic lovers.

But the real move came one word earlier. The question was whether college is worth it. The machine quietly swapped worth for return on investment. That’s the swap to watch: not a false claim, a substituted one. Everything downstream defends the cheaper question.

So whether college makes you a better thinker, citizen, or human survives only as a footnote: arts and humanities majors “develop highly adaptable critical thinking and communication skills,” which the answer immediately prices as a low-value workplace asset. The civics case wasn’t deleted. It was financialized.

And mind the quantifiers. Every figure is a median or an average—the middle of the pack, not the share of students for whom college actually pays off. “For most” is an unearned jump from central tendency to proportion.

E — What’s the evidence?

Three kinds of numbers, all for four-year graduates: a $1M lifetime gap, a $96K California wage (double the high-school median), and lower unemployment. Check them, and they mostly hold. PPIC really does report the $96K as a doubling, and the $1M lifetime gap is, if anything, conservative next to Georgetown’s estimate nearer $1.2 million.

But look at the framing. It calls college “an excellent return on investment” while supplying only the return. The investment—tuition, time, interest—is missing until “debt” turns up late, as a caveat. You can’t certify a ratio from its denominator alone. The $96K is one state’s figure (California), quietly drafted to back a nationwide verdict. The “robust benefits like health insurance and retirement plans” belong to particular jobs, not to the diploma, and arrive uncited. Plenty of plumbers have both.

And those earnings come from people who finished college decades ago, before AI began reshaping the entry-level market this year’s applicant is about to enter. The evidence looks genuine. It’s also dated, gross rather than net, and silent on the automation already moving through the labor market—a $1M headline never once set against four years of forgone wages and loan interest. Past performance, no disclaimer about future results.

R — Does the reasoning hold?

The core inference is: graduates earned more; therefore, college caused the earnings. That’s the oldest confound in education research: selection bias. The second inference was it paid in the past, so it’ll pay in the future. That’s extrapolation—inductive over-reach, for you logic lovers—and it assumes the future shows up shaped like the past.

Some of the premium is the degree; some is just who earns one. People who finish college differ from people who don’t in ways that move earnings on their own. And the premium describes only the finishers: roughly a third who enroll never complete, many carrying debt and no degree, a risk the average quietly erases.

Then the answer refutes its own headline. It admits debt “can quickly neutralize the financial advantages of a degree.” Yet, most students borrow. “Excellent ROI for most” and “advantage neutralized” cannot both stand. It switches yardsticks, too, judging college by lifetime earnings while selling the alternatives as “much faster” to enter. Its closing advice to “maximize your return on investment” simply assumes the ROI frame the question never granted.

I — What follows if it’s true?

Okay, let’s hypothetically grant the core argument (CER). What logically follows? That past graduates of past colleges earned a premium. It would not follow that the college you’re weighing, at the price you’ll pay, in the market your graduating year meets, is worth it.

The biggest implication lurks in the part the machine treated as an afterthought: if returns swing by major and institution, then which college and which major make the entire decision, not whether. Yet the “highly lucrative alternatives”—trades, bootcamps, the military—get waved through with no numbers at all, carrying the very selection and survivorship problems the answer just flagged for college.

Two more issues. If the payoff “depends on personal career goals,” then there is no general answer, which contradicts the confident “Yes.” And if majors and institutions are the real drivers, the headline advice to save on gen eds at a community college aims at the wrong variable. The prescription doesn’t follow from its own diagnosis.

C — What’s the context?

Who wrote this? A model trained on the open internet, where the college-ROI genre is owned by universities marketing themselves, lenders marketing loans, and listicles recycling both. The machine filled the gaps with Reddit, Quora, and Medium posts, which is exactly what this answer leans on. The machine has no stake in your kid’s future. Its sources did.

Now count the citations: 1, 2, 3, then 4 and 6, then 8 through 11, then 13 through 17. Five, seven, and twelve never appear. A sourcing apparatus that can’t keep count of its own sources is a tell, all by itself.

Watch the words drift inside a single section, too. The alternatives are “highly lucrative” in one sentence and pay “excellent livable wages” in the next—lucrative downgraded to livable while you weren’t looking. Fluency is what the machine optimizes. Sleight-of-hand is what we got.

What just happened

We asked five questions in ten minutes. A reasonable-sounding answer turned out to hold a swapped claim, evidence that’s real but dated and missing the key trend we’re worried about, reasoning built on selection bias and extrapolation, a conclusion resting on more assumptions than it admits, and a source list pockmarked with self-interested promotion and anonymous forums.

And the machine’s answer wasn’t bad. As a place to start—to pull those statistics and read deeper into current trends—it’s useful. The danger was never the answer. It was a reader relying on it whole.

Here’s the claim, stated plainly. AI generates text; only a reader generates judgment. The machine even named critical thinking—once, as a consolation prize for arts and humanities majors. It had its own product backward. The more fluent text we’re handed, the more of that reasoned judgment we have to bring ourselves.

Try it this week. Ask any AI a question you actually care about, run the five CERIC questions on its answer, then reply or comment and tell me what you caught. Next walkthrough, we’ll point CERIC at something harder—maybe you’d like to suggest an article in the comments.

Read deeply,

Dr. Genevive

In 2024, The Atlantic published a piece whose headline did half the work: “The Elite College Students Who Can’t Read Books.” Rose Horowitch interviewed 33 professors, and their stories rhymed. Nicholas Dames, who has taught Literature Humanities at Columbia for decades, described students arriving “bewildered” by the expectation of reading multiple books in a semester. Others described the quiet workaround already underway: assign excerpts instead of books, chapters instead of novels, shrink the syllabus until it fits the students you have rather than the students you remember.

Let me do to this essay what I’d ask you to do to any text: read it deeply before accepting my argument.

The honest caveat first. Horowitch’s evidence is anecdotal: 33 professors’ experiences raise important questions, but they are not a dataset. Her own piece concedes that no comprehensive data exist on the trend, and critics said so loudly. Thirty-three frustrated professors is a pattern of complaints, not proof of a broader decline. If the story stopped there, you’d be right to hold it at arm’s length.

But it goes further.

When quantitative data arrived, it was worse. The Nation’s Report Card—NAEP, the closest thing America has to a measurement of what students can actually do—released its 12th-grade results in September 2025. Reading scores fell to their lowest level since the test’s current form began: only 35 percent of high school seniors were proficient in reading, and 45 percent were below even the basic level. These are the students arriving on campus this fall. New results released this month show that younger students are recovering, while teenagers continue to slide. The narrative and the numerical data now converge on the same thing: the students walking into freshman seminars are, on average, weaker readers than any cohort in decades—and the strongest readers are pulling further away from everyone else.

The workaround makes it worse. Faced with students who can’t finish books, colleges are choosing, mostly silently, to stop assigning them. Understandable. And also exactly backward. Deep and critical reading isn’t a prerequisite for college; it’s the product of it. The cognitive scientist Maryanne Wolf has spent a career showing that the reading brain is built, not born: the circuitry for sustained, analytical, empathetic reading forms only through practice and dissolves without it. Shrink the syllabus, and you’re not accommodating the deficit; you’re institutionalizing it.

And the timing could not be more perverse. This is happening at the precise moment AI makes shallow reading worthless. The one capacity that can’t be automated—sitting with a difficult text, tracking its claims, examining its evidence, evaluating its reasoning—is the one colleges are inadvertently removing from the curriculum when they remove deep reading because it is hard to teach, especially when students simply don’t do the reading.

There’s a counter-signal worth watching. Princeton just assigned Wolf’s Reader, Come Home—a book about the loss of deep reading—to its entire incoming class. Read that move for what it is: one of the most selective universities in the world telling its newest students, before they unpack, that the skill most at risk is the one it considers foundational. Princeton can see the same NAEP tables as everyone else. The difference is the response: name the problem and teach into it, rather than shrink the syllabus around it.

Here’s my claim, stated plainly. The students are not broken. They are unpracticed, in exactly the way a generation raised on skimming, scrolling, and now summarizing by machine would predictably be. A skill nobody teaches doesn’t get learned. That’s not rocket science; it’s a syllabus decision. Which means the crisis is also an instruction problem, and instruction problems always have answers.

That’s what the method is for. CERIC (Claim, Evidence, Reasoning, Implications, Context) is a structure for the thing students were never taught to do: stay inside a text long enough to evaluate its arguments. It works on a chatbot’s answer (we will do this later this week), it works on a news story, and it works on the books that have quietly vanished from the syllabus. In coming posts, I’ll show how it works in an actual classroom, including with students who arrive saying they’ve never finished a book.

Colleges can keep shrinking the reading until there’s nothing left to assign. Or they can decide that teaching people to read deeply and critically is what they meant to do all along.

Next week — CERIC vs. the Machine: The five-question method, run head-to-head against the chatbot that’s replacing the reading.

Read deeply,

Dr. Genevive

If you teach, I’d like to hear it straight: What’s happened to the reading load in your courses over the past five years? Please reply or comment. Anonymized patterns may show up in a future post.

Sources referenced:

• The Elite College Students Who Can’t Read Books — The Atlantic

• Critique of the Atlantic piece — Three Percent, University of Rochester

• NAEP 12th-grade scores decline — Chalkbeat, Sept 2025

• New NAEP results — Chalkbeat, June 2026

• How to practice deep reading — NPR Life Kit

• Princeton selects Reader, Come Home as 2026 Pre-Read

I know this firsthand. In my own doctoral program, I discovered that no one had ever taught me, or most of my classmates, how to read a research paper. We were expected to already know.

Colleges are struggling to figure out what to do with the next generation of students, and the panic has a common root: the implicit contract of higher education has broken. The contract was straightforward. We transfer information, students demonstrate they absorbed it through essays and exams, and everyone agrees to call that an education. AI severs both halves at once. The essays and exams were only ever proxies for learning, and those proxies can now be gamed in seconds. And the thing they stood in for—the processing of information itself—AI now does better, faster, and for free. So the deal fails from both ends: we can no longer tell who learned, and it no longer matters whether they did. If that's what college is for, college is over.

But that was never what college was for. It was for something harder to name and harder to teach: the ability to read a claim and evaluate it. To ask what the evidence is, whether the reasoning logically connects the evidence to the claim, what follows if the claim is true, and who wants you to believe it. That skill has a delivery mechanism, and it isn’t the lecture. It’s deep, critical reading followed by discussion or discourse.

The symmetry is uncomfortable. The technology that makes reading feel optional is the same technology that makes critical readers indispensable. AI produces fluent, confident, sometimes-wrong text at infinite scale. Every hour of your life now involves judging machine-generated claims. A society that can’t read deeply is defenseless.

So this publication makes one argument, from different angles, with a method attached: the more we rely on AI, the more we need people who know how to read and think critically. And that skill can and should be taught.

The method is called CERIC, and it’s five questions a deep reader asks of any text: What’s the Claim? What’s the Evidence? Does the Reasoning hold? What are the Implications? What’s the Context? Five questions, endlessly deep, and entirely AI-proofAI can answer them. The point is that you can. Reasoned judgment, unlike information, doesn’t transfer. It has to be built, one text at a time.

That’s what we’ll do here:

• Explore what college is for now

• Walk through the CERIC method that you can use on a syllabus, with your students, or alone with a difficult text.

• Practice deep and critical readings of texts that deserve our attention—including, sometimes, the output of the machines themselves.

The old college is dead. The case for deep readers and critical thinkers has never been stronger. Welcome to The Deep Reader.

Read Deeply,

Dr. Genevive

If you teach, parent, or read — subscribe. Next week: The Students Who Can’t Finish a Book.

A question for the comments: if the old model is dead, what’s the one thing college still does that nothing else can, or is even that gone?