The ability to evaluate what AI produces was never a technical skill. It was a humanities education.
In this post I will:
Show how Melania Trump’s robot Plato accidentally exposed the case for the humanities in an AI world.
Argue that twenty years of defunding the humanities created a population that can use AI but cannot evaluate what it produces, with evidence from the OECD and the automation of Canvas.
Recommend five texts, from Aristotle to Carolyn Forché, that build the evaluative muscle AI cannot replicate (paid section).
Why naming a robot Plato tells you everything
Melania Trump walked into a White House education summit with a humanoid robot and pitched it as the future of teaching:
“Imagine a humanoid educator named Plato. Access to the classical studies is now instantaneous.”
The backlash was immediate and bipartisan. Randi Weingarten, president of the American Federation of Teachers, called it “every parent’s nightmare.” Philosopher Jennifer A. Frey wrote that:
“If you cannot loudly and firmly resist ‘humanoids’ called Plato teaching our youth, you have no place in any movement that calls itself ‘classical’”
Plato argued that education is not about putting knowledge into someone but drawing out what is already there. He taught through questions, not answers. His entire method depended on the relationship between teacher and student.
A robot can answer every question. It cannot notice the student who has stopped asking them.
Why the humanities have been defunded for twenty years
Governments across the English-speaking world have treated the humanities as a luxury. The numbers tell the story discipline by discipline. In the UK, enrolments in language studies fell 45% between 2012 and 2024. In the US, bachelor’s degrees in English fell by more than a third in the same period. Philosophy departments have been merged, cut, or quietly starved. The pattern repeats across every humanities discipline, in every English-speaking country, with only minor variations in the speed of decline.
The justification was always economic. STEM produces workers. Humanities produce thinkers. The market wanted workers.
I have a PhD in atmospheric physics. I spent a decade measuring greenhouse gases from satellites. The technical skills got me the job. The thing that made me good at it was reading poetry. Because poetry trained me to notice what was not being said in a dataset, to sit with uncertainty instead of resolving it prematurely, to read the same output three times and find a different meaning each time. Every critical decision I made as a scientist was an act of interpretation, not computation. I did not know that at the time. I know it now.
Then AI arrived.
The most important skills AI cannot replicate are humanities skills
The skills the market had dismissed turned out to be the ones the machines could not do. Judgement. Interpretation. The ability to read between the lines of something that presents itself as neutral. The capacity to hold two contradictory ideas and decide, slowly, which one matters more. The willingness to sit with ambiguity instead of resolving it.
The OECD Digital Education Outlook 2026 confirmed this with data. A field experiment in Türkiye found that students using AI were 48% more successful at completing tasks, but performed 17% worse when the AI was removed. Performance went up. Learning went down. Nobody noticed because we were measuring the wrong thing.
What the OECD described is the absence of exactly what the humanities train you to do: evaluate, question, resist the first answer, and ask whether the output is actually good or merely fluent.
AI is fluent. Fluency is not the same as good. Goodness requires judgement, and judgement is what the humanities train.
What happens when you automate the thinking tasks
Canvas, the virtual learning environment used by 40% of North American universities, launched an AI teaching agent in March 2026. It automates rubric generation, discussion reviews, and content alignment.
Rubric design is not administrative busywork. It is the act of deciding what matters in a piece of work. Discussion review is not scanning for keywords. It is noticing where a student’s thinking shifted. Content alignment is not matching outcomes to slides. It is asking whether the outcomes are worth teaching.
Automate these and you save time. You also remove the last place where the educator was forced to think carefully about what they were actually doing. Nobody has studied what happens to educators who stop designing rubrics or reviewing discussions by hand. But the pattern the OECD identified in students (displacing cognitive effort leads to weaker independent performance) is unlikely to stop at the classroom door. Every time you hand a thinking task to a machine, the muscle that did the thinking gets weaker. That applies to the teacher as much as the student.
Why the humanities are the moat against AI
The humanities are the moat because they train the one capability AI cannot simulate: the ability to evaluate whether an output is worth anything.
An AI can summarise a paper. It cannot tell you whether the paper should have been written.
An AI can generate a policy brief. It cannot tell you whose voice is missing from it.
An AI can produce a poem. It cannot tell you whether the poem matters.
These are evaluative acts. They require context, experience, values, and the willingness to be wrong. They require a reader who has been trained, slowly, over years, to notice what is absent.
That training happens in the humanities. In literature, where you learn that every narrator is unreliable and every sentence is a choice. In philosophy, where you learn that the strongest argument might still be wrong. In history, where you learn that the people who wrote the record had reasons for what they left out. In poetry, where you learn that the silence between words carries meaning.
Defund these disciplines and you get a population that can use AI but cannot evaluate what it produces. A workforce that is faster but not wiser.
The hardest question in critical AI literacy
There is a question I ask in my curriculum: what is the AI not saying?
It sounds simple. It is the hardest question in critical AI literacy. Because the AI always sounds complete. The output is fluent, structured, and confident. It reads like it has considered everything. The skill is noticing what it has not considered, and that skill has a name.
It is called close reading. It is one of the oldest practices in the humanities. And we have been defunding it for twenty years.
The Slow AI Curriculum for Critical AI Literacy is a twelve-month programme that teaches you to notice what AI leaves out. It is CPD-accredited, grounded in peer-reviewed research, and built around the evaluative skills this post describes. If the argument above resonated, the curriculum is where you put it into practice.
Five texts that build the skills AI cannot replicate
If you want to build the moat, you have to read things that resist easy summary. Below are five texts. They are not recommendations in the usual sense. They are training exercises for the evaluative muscle that AI is quietly making redundant.
I have chosen them because each one does something specific that a large language model cannot.
If you want to read these, get them from your library. I also wrote about why reading, slowly and critically, is the most undervalued AI skill there is in The Best Response to AI Is a Library Card.
Rhetoric, Aristotle (4th century BCE)
Every AI-generated text is an act of persuasion. It is trying to convince you it is correct, comprehensive, and worth your trust. Aristotle gave us the tools to evaluate persuasion 2,400 years ago: ethos (credibility), pathos (emotional appeal), and logos (logical structure).
Read Rhetoric and you will never read an AI output the same way again. You will notice where the ethos is borrowed, where the pathos is manufactured, and where the logos has gaps the fluency is designed to conceal.
This is the oldest technology for evaluating language. It teaches you to see through fluency in ways that no automated detection tool can match.
Frankenstein, Mary Shelley (1818)
The original AI ethics text, written two hundred years before the field existed.
Shelley’s question is not whether the creature is dangerous. It is whether the creator has any right to abandon what he made. Victor Frankenstein builds something, refuses to take responsibility for it, and spends the rest of the novel blaming his creation for the consequences of his own negligence. Replace ‘creature’ with ‘model’ and the plot describes the entire AI safety debate.
Read it for the structure of the argument, not the horror. The horror is incidental. The ethics are permanent.
Do Androids Dream of Electric Sheep?, Philip K. Dick (1968)
Dick’s question is not whether machines can think. It is whether humans can still feel.
The Voigt-Kampff test in the novel measures empathy, the one quality the androids cannot simulate. The test fails. Not because the androids pass it, but because the humans start failing it too.
This is the book to read when someone tells you that AI will never replicate human emotion. Dick’s answer, written in 1968, is: it does not need to. It only needs to make us worse at recognising our own.
Watchmen, Alan Moore and Dave Gibbons (1986)
The question at the centre of Watchmen is not who has power. It is who decides what to do with it when nobody is looking.
Ozymandias makes a unilateral decision to kill millions because he calculated it would save billions. He is intelligent, informed, and wrong in a way that no amount of data could have corrected. The judgement failure is not computational. It is moral.
This is the book to read when someone tells you that AI just needs better guardrails. Moore’s answer, written in 1986, is that the guardrails are only as good as the people who set them. And the people who set them always believe they are the exception.
The Colonel, Carolyn Forché (1978)
One page. A witness poem.
Forché describes a dinner with a Salvadoran military colonel who empties a bag of dried human ears onto the table. The poem does not explain. It does not interpret. It reports.
And in the reporting, it does something that AI structurally cannot: it asks the reader to decide what the facts mean. There is no summary of this poem that captures what it does. You have to read it.
That is the moat in one page.
What these five texts share
These five texts span 2,400 years. They come from different traditions, different centuries, different forms. What they share is that none of them can be reduced to information. They require a reader. They require time. They require the willingness to be changed by what you read.
That is what the humanities teach. That is what twenty years of policy has defunded. And that is what no robot named Plato will ever replicate.
A note on what is missing from this list
I know this list is biased. Five texts, all from the Western canon. The argument I am making applies across every literary and philosophical tradition. Oral storytelling, indigenous knowledge systems, non-Western philosophical frameworks: all of these train the same muscles of interpretation, ambiguity, and critical judgement. The Western canon does not own this. It is simply the tradition I was trained in, and I would rather be honest about that than pretend to a breadth I have not earned.
I talk about decolonising reading lists through tools like the reading list audit, which helps you identify your own biases in what you assign and recommend. If you have texts from other traditions that build the same capacities, please share them in the comments.
Go slow.
Sam, this piece is the closest thing I've seen to the argument I've been trying to make, and I want to push it the one step further I think it needs.
You're right that the humanities are the moat. But I want to ask: what made the humanities work? You say poetry trained you to sit with uncertainty, to notice what wasn't being said, to read the same output three times and find a different meaning. Where did poetry learn to do that?
Aristotle's Rhetoric didn't emerge from nowhere. It emerged from a civilization that had the Delphic maxims, the mystery traditions, the lived practice of public deliberation where the wrong judgment got people killed. Shelley didn't invent the question of what a creator owes his creation. She inherited it from Milton, who inherited it from Genesis. The evaluative muscle you're describing - the ability to notice what's absent, to sit with ambiguity, to ask whether something should have been written at all - that muscle wasn't built by the humanities. It was built by the moral traditions the humanities were drawing on. The humanities gave it a secular vocabulary. The traditions gave it teeth.
You wrote the most important line in this piece almost in passing: "A robot can answer every question. It cannot notice the student who has stopped asking them." That's not an evaluative skill, Sam. That's presence. The ability to see a person, not a performance. And presence wasn't trained by close reading. It was trained by structures that forced you into the room with people when you'd rather not be - the weekly meal with no agenda, the apprenticeship where a master knew you for years, the community where someone noticed you were struggling before you said a word. Those structures weren't humanities programs. They were moral traditions. And we dismantled them decades before we defunded the humanities.
The OECD study you cite is devastating - students 48% more successful with AI, 17% worse without it. Performance up, learning down. But this isn't new. This is the same pattern that played out when we replaced the apprenticeship with the factory school, the Friday table with the television, the elder who told you a story with the therapist who billed by the hour. Each replacement optimized for output. Each one severed a connection that was training judgment without anyone noticing.
You ask "what is the AI not saying?" That's the right question. But the skill to answer it doesn't come from Aristotle or Forche - though both help. It comes from years inside a tradition that taught you what mattered and why. A tradition that didn't ask if you were in the mood. That made you show up whether you felt it or not. That trained the muscle through obligation, not inspiration. The trolley problem is sixty years unsolved by philosophy. Jewish law has been adjudicating it - case by case, context by context, with binding authority and preserved dissent - for millennia. Not because the rabbis were smarter than the philosophers. Because they were operating inside a system that required decisions, not just deliberation.
You said at the end that your list is biased - all Western canon - and you invited other traditions. I'll take you up on that. The tradition I come from says that analytical thinking can tell you whether an argument is valid. It cannot tell you whether the argument matters. And the thing that tells you whether it matters isn't a text on a reading list. It's a way of life. One that we stopped living, and that no amount of humanities funding will restore on its own.
The humanities are the moat. But the moat was fed by a river. And the river dried up long before anyone noticed the moat was empty.
Heartily agree. I studied Sociology and Social Welfare, and then took a job in Finance. My nieces are all over STEM… but, they informed me there is something called STEAM. The A is for Arts. Thankfully that’s at least incorporated!