The Expired Syllabus: why we can’t teach 2026 the way we taught 2025 — and what 1985 knew

Every educator I know has the folder. The one with last year’s plans — the sequence that worked, the activities that landed, the learning experiences refined year after year until they ran like a well-rehearsed play. And every year, around now, we open the folder and ask the most tempting question in education: can we just run it again?

This year the honest answer is no. We can’t teach in 2026 the way we taught in 2025 — much less the way we taught in 2015. We can’t hand learners the 2025 syllabus and expect it to produce 2026 learning. Not because everything we knew is suddenly wrong, but because three things every plan quietly assumes have all moved at once: the context our learners live in, the learners themselves, and the tools sitting in their hands. A syllabus is a set of assumptions with a schedule attached. When the assumptions expire, the schedule is the only thing left running.

Let me be precise about what kind of claim that is. It is not a research finding; nobody ran a study on the shelf life of a syllabus. It is a design position, stated from experience — the same kind of judgment an architect makes when the ground under a building shifts. And it comes with a twist I find genuinely hopeful: some of what we need for 2026 is not new at all. It was ordinary practice in 1985, and we lost it along the way.

Three assumptions, all expired

The context changed. What counts as ordinary work — in offices, studios, clinics, newsrooms — has been redrawn while our plans sat in the folder. Producing a first draft, a summary, a translation, a passable image, a working snippet of code: tasks that used to be deliverables are now starting points. A learning experience that trains someone to produce what the surrounding world already treats as a starting point is preparing them for a workplace that closed.

The learners changed. The learner who walks in this year did not discover conversational AI in your classroom, and will not ask your permission to use it. The question “will they use AI?” was settled somewhere else, some time ago, without us in the room. The live question — the only one still ours to influence — is whether they will learn while using it, and that is decided almost entirely by how the activity in front of them is designed.

The tools changed — in kind, not just in degree. The 2015 worry was copy-paste: static text lifted from the Internet, identical for everyone, findable by anyone. The 2026 reality is generative: every learner can now produce, in seconds, an artifact that has never existed before — custom, unique, plausible. Which means something quietly enormous for anyone who designs learning: the deliverable stopped being evidence of learning. A clean essay, a correct-looking chart, a polished presentation — each proves that an artifact was produced. None of them proves that a learner met the subject on the way.

None of this is a reason for panic, and none of it is an argument for banning anything. It is a reason to redesign. And the redesign has a compass much older than the Web.

What 1985 knew

In The Maps We Copied I told the story of the geography teacher who sent us to copy the map of the country’s mountain ranges out of a book. She was not asking for cartographers. She did not care about the map. What she was doing — deliberately, whether or not she had the vocabulary for it — was making us interact with the learning object: the peaks, the rivers, where things sit in relation to each other. The map was the medium. The geography was the lesson.

Sit with that assignment for a moment, because it carries three design decisions we have quietly lost.

First, the artifact was never the point, and everyone knew it. Nobody graded us on the beauty of the map; the copying existed to hold our attention on the thing being learned, hour after hour. Second, the tool was declared and unremarkable. The atlas was open on the desk, in plain sight, and no one called it cheating — because the assignment was designed so that using the tool still required meeting the subject. Third, repetition was not an insult. Tracing, redrawing, correcting, labeling again: the assignment made room for the unglamorous passes through the material that turn acquaintance into knowledge.

And 1985 had a second assignment, one that looks even worse to modern eyes: the hundred algebra problems left as homework for the weekend. That was not a lack of creativity, and it was not a commitment to boredom through naive repetition — I have made the long case for what that volume was actually doing in Deliberate Practice at Scale. It was the teacher’s way of getting us to engage with the learning object — the equations — for a significant amount of time: enough time to learn. The answers were printed at the end of the book anyway, and the teacher could not have cared less about them. What we were expected to produce was the reasoning — to work the process and write it down, step by step, until arriving at the answer was the least interesting part of the page. Process over product, already graded that way in 1985.

An assignment designed so that producing the thing forces you to keep looking at the thing; a tool that is visible instead of policed; time — real, repeated time — spent with the learning object. That was 1985. Every one of those decisions is exactly what 2026 needs, and every one of them fell away in the decades when the deliverable got easy.

The redesign principle: time with the learning object

So here is the whole principle, and I have not found a redesign question that survives without it: rebuild every learning activity so the learner spends significant time interacting with the learning object.

Run any activity in your folder through one test: while the work is being done, where does the learner’s attention actually live — on the learning object, or on the packaging of the deliverable? In 2015, an essay on the causes of a war forced attention through the war: you could not write the essay without passing through the subject. In 2026, the same essay can be produced without attention ever touching the subject at all. The activity did not get worse. Its central assumption — that making the artifact requires meeting the object — expired. The essay prompt from the 2025 folder is not lazy or broken; it is simply a key to a lock that has been changed.

Once you accept the principle, every activity you redesign lands in one of four families. Most institutions, in my experience, reach for only the first and wonder why their learners are bored, or only the second and wonder why nothing sticks — while the third and fourth, where the machine's own limits do the pedagogical work, go unused. This set is the heart of how we work with educators on redesign, and each family does work the others cannot.

Family one: activities AI cannot do

Not “AI-proof” by surveillance — activities whose substance is simply out of the machine’s reach, because they happen in the room, in the body, in the moment. Explaining your reasoning out loud to someone who can interrupt and ask why. Defending a decision you made and hearing it challenged. Building, measuring, growing, observing something physical and recording what actually happened, not what should have. A debate where your opponent’s next move is unknowable. Placing the five highest peaks on a blank outline map, from memory, with every tool closed. Interviewing a real person and being changed by an answer you did not expect.

These need almost nothing: a room, a voice, a blank page. That is not a weakness; it is their equity and their power. The spoken answer and the blank page are ancient technology, and they still do the one thing no generative tool can do for a learner — reveal, to the learner and to everyone present, what is actually in their head. When the artifact can no longer prove learning, these activities are where the proof lives.

Family two: activities that require AI

The second family is the mirror image, and it is where the joy comes in: activities that are impossible — or pointless — without AI. Designed well, they use the machine for exactly what it is genuinely good at: producing custom, unique responses, so that no two learners are ever holding the same artifact; letting learners attempt things that were simply out of reach before — more ambitious, more varied, more theirs; and making the work honestly exciting, because watching your idea take a form you could never have produced alone is exciting, and pretending otherwise wins us nothing.

The design constraint that keeps this family honest is the same one the map teacher used: the brief must demand engagement with the learning object, and a generic answer must be insufficient by design. The AI generates; the learner audits the output against sources, catches what is plausible but wrong, corrects it, and iterates until the result satisfies the evidence — not the model. Every “no, that’s wrong, fix it” is another pass through the subject. The tool amplifies what the learner can attempt; the design guarantees they meet the subject on the way.

Family three: activities where AI is likely to fail — and the learner intervenes

Here is the part we should say with a smile: in 2026 we are still lucky, because there is so much the AI gets wrong. The third family is built on exactly that luck — activities aimed deliberately at what the machine cannot yet do well, so that the learner must intervene: identify the limitation, address it, solve it, validate the result against something real, and iterate until the outcome is actually good. It hallucinates, it fails to represent things adequately, it fails to achieve what it was asked effectively — and every one of those flaws is a gift to whoever designs learning experiences. The machine’s weakness becomes the learner’s workload, and the workload is exactly where the learning lives.

Take the selfie assignment from the maps essay and read it as flaw-hunting: create a selfie of yourself at the summit of each of the country’s ten highest peaks. Not hard at first glance — one prompt, ten images, done before recess. But now require that every picture hold up: vegetation adequate to that altitude, a correct horizon, the sea or valley or city that is really visible from up there actually visible below, clothing adequate to the real temperature at the real summit. Suddenly ten easy images become ten audits — and the auditor has to know the geography. The AI does the rendering; the learner supplies everything the AI cannot be trusted with. That is the template for the whole family.

A working catalogue of useful flaws

These are the faults, inaccuracies and limitations I reach for first when designing a task around AI — each one reliably present today, and each one convertible into work the learner must do in person. (The catalogue is perishable by nature: models improve, and specific flaws will fade. The design move does not — find where the model fails your subject this year, and build the activity on that exact spot.)

• Confident hallucination of facts. Invented dates, names, figures and events, delivered with total assurance. It forces verification against sources the learner must name — and every error caught against a source is domain knowledge gained.
• Invented citations and references. Books, articles and studies that do not exist, or that exist and say something else. It forces the trip to the actual catalogue, library or original text.
• Plausible-but-wrong arithmetic. Multi-step calculations with fluent, intermittent errors. It forces checking by hand — calculation as judgment, not as faith.
• Smoothing conflicting sources into one clean story. Where accounts disagree, the model averages them into a single frictionless “truth.” It forces reading the actual sources side by side and adjudicating the disagreement.
• No local or live knowledge. The model was never in your classroom, your neighborhood market, or your interviewee’s living room. It forces collecting real information — prices, measurements, testimony — that no prompt can substitute.
• Stereotyped gap-filling. What the model does not know it fills with the plausible average — the generic neighborhood, the generic school, the generic family. It forces confronting the output with the reality the learner can actually observe.
• Anachronisms and invented detail in historical material. Costumes, objects and scenes that look right and are wrong. It forces date-by-date, image-by-image verification against period sources.
• Geographically implausible imagery. Wrong vegetation for the altitude, snow on tropical peaks, horizons and shorelines that could not exist from that viewpoint. It forces knowing the real place well enough to catch the fake — the selfie audit above.
• Invented species, places and specifics. Ecosystems populated with organisms that do not live there, or do not exist. It forces species-by-species verification against an authoritative source.
• Dropped and reinterpreted constraints. Ask for five things and receive four, one of them changed. It forces auditing the output against the instruction itself — reading one’s own specification carefully enough to enforce it.
• Inconsistency across iterations. Details that silently drift between versions — a character’s face, a value, a premise. It forces the learner to hold the canon of their own work and defend it against the tool.
• Agreeing with a wrong premise. Feed the model a false assumption and it will often build helpfully on top of it. It forces learners to test claims rather than collect confirmations — including planting deliberate errors and watching who catches them.

Notice what every entry has in common: the flaw does nothing on its own. It becomes pedagogy only when the brief requires the output to survive an audit the AI cannot perform — which is why, in the Challenge’s authoring framework, every brief must declare its friction point: the place where the AI will fail, predictably, before the first team ever sits down. The friction point is not an obstacle in the lesson plan. It is the lesson plan.

And a date stamp, because this essay must practice what it preaches: that catalogue is 2026’s. In 2027 we must revisit it — most likely many of those flaws will have been corrected, and we will need to find new ways to make sure AI is used so that learners still engage with the learning object, and their creativity keeps developing. The expiry that opened this essay applies to the essay too.

Family four: post-AI activities — processing the output with the AI off

The fourth family begins where the generation ends. The AI has produced its text, its image, its plan — and now the screens close. What follows is deliberately non-AI work on the AI’s output, because this is where what the machine produced becomes what the learner knows. Without this family, the other three leak: the artifact stays in the chat session and never moves into the person.

Specific instructions that work — all of them with the tool off:

• Print it and annotate it by hand. Every claim gets a mark: verified (against what source?), unverified, or wrong — with the correction written in the margin.
• Rank three outputs and defend the ranking. Generate three candidate answers beforehand; the post-AI work is choosing the best one and writing why — criteria first, verdict second.
• Rewrite it in your own words for a named audience. The AI’s register is nobody’s; translating it for your little sister, your mayor, or your classmates forces ownership of every idea in it.
• Reconstruct the reasoning. Take the answer and work backwards on paper: what would have to be true for this to be right? Which steps can you actually justify, and which are you taking on faith?
• Localize it. The output describes the generic case; adapt it, on paper, to your classroom, your neighborhood, your data — and note every place the generic version would have failed.
• Explain it aloud and take questions. The oldest post-processing there is — and the bridge back to family one: if you cannot defend it without the tool, the work is not yours yet.

Four families, one test: significant time with the learning object. The first family protects the time; the second makes the time irresistible; the third turns the machine’s weaknesses into the curriculum; the fourth makes sure that what the machine produced becomes what the learner knows.

What it looks like when it runs

This is not a thought experiment for us. The aiLearning Challenge is this argument built as an event, on our Smoother methodology, and its design decisions read like a checklist of everything above. The briefs are written so that a generic AI answer is insufficient by design — a solution that would read identically in any country sits at the bottom of the scale, because it proves no contact with the actual context. The evaluation weighs process over product, because a beautiful artifact can conceal an empty learning process. Iteration is read as evidence rather than hidden as shame — the version trail, the failures, the adaptations are part of what judges look at. And the final gate is a live defense with no AI in the room: family one, closing what family two opened.

Notice how the event walks all four families: it requires AI for ambitious, creative, custom work no participant could produce alone (family two); its briefs carry a declared friction point where the AI will predictably fail and the team must intervene and validate (family three); the audits, the creation record and the version trail are post-processing of the machine’s output, owned by the team (family four); and it finishes with the one thing AI cannot do — a person, standing in a room, explaining and defending what they now know (family one). None of the four works without the others. That is not a compromise between enthusiasm and caution. It is the design.

Change the folder, not everything

I am not asking anyone to burn the folder. Most of what is in it — the relationships, the sequencing instincts, the hard-won sense of where learners struggle — is exactly the expertise the redesign needs. The ask is smaller and sharper: this year, refuse to run the 2025 plan on inertia. Take the activities where, if you are honest, the learner’s attention no longer has to pass through the learning object — and move each one into a family: something AI cannot do; something that requires AI, designed so the generic answer fails; something aimed at a flaw where the AI will stumble and the learner must intervene and validate; or something post-AI, where the output is processed, corrected and owned with the tool off. One activity at a time is enough; the folder compounds.

And this is not only an educator’s conversation. Families are facing the same expired assumptions at home, and I have made the same argument to parents — in Spanish — at Papás Academy: No prepares 2026 como preparaste 2025. Same thesis, other side of the kitchen table.

The 2025 folder is not wrong. It is expired. And the fix is not newer technology — it is an older question, the one the map teacher answered in 1985 without ever hearing it asked: where will the learner’s time with the learning object come from? Design every activity around that answer, and the year takes care of itself.

This essay states my design position and professional conviction, as mine — the yearly-expiry frame and the four-families taxonomy are judgments from practice, not research findings. The map assignment and the evidence around it are laid out, with sources, in The Maps We Copied; the Challenge’s design is documented at /challenge/. — Carlos Miranda Levy

Four perspectives