What AI Made Cheap
A level-by-level read of Bloom's taxonomy in the age of generative AI: which rungs a model now completes for free, which ones resist it, and why 'the AI can do it' is a design signal, not a threat.
Pricing Each Rung
Think of every Bloom’s level as having a cost to the student — the cognitive effort it takes to produce acceptable work. Generative AI didn’t change what the levels are. It changed their price. Here’s the ladder, re-priced:
Remembering — near zero. A model recalls definitions, dates, and facts more completely than any student. An assignment whose deliverable is recalled information is now an assignment whose deliverable is free.
Understanding — near zero. Summaries, paraphrases, explanations, classifications. This is the single most common thing traditional homework asks for, and it’s the thing chatbots were, in effect, purpose-built to do.
Applying — low, but conditional. Execute a known procedure on a familiar problem type — solve the equation, write the standard function, translate the passage. AI does the textbook cases well. It gets less reliable as the application moves away from patterns it has seen, which is a crack you can design into.
Analyzing — rising. Distinguishing, contrasting, dissecting a specific artifact. A model can produce plausible analysis, but its quality now depends heavily on the student’s ability to feed it the right material and judge whether the analysis is actually true of this case. The thinking starts to move back to the human.
Evaluating — high. Argue, judge, rank, critique, defend a choice under real constraints and take responsibility for it. A model will generate an argument, but it has no stake in the judgment and no accountability for being wrong. The moment the task requires owning a defensible position, the human is back in the loop.
Creating — highest. Generate something genuinely new that fits a specific, constrained context — this class, this dataset, this community, this problem nobody has framed yet. AI is a powerful instrument for creation and a poor author of it.
The pattern is a diagonal: cost to the student rises as you climb, and AI’s ability to substitute for the student falls. Where those two lines cross is roughly “the line” from Module 1.
Kallas draws the line sharply: she grants that AI is capable of doing nearly everything on the ladder — remembering, understanding, applying, and even analyzing — “with a few exceptions: they are not very good at doing evaluation and creation.” AI can generate new outputs, but only “out of statistical probability,” whereas “human creation is completely something else.” In her framing, evaluating and creating are the rungs where the human brain is still unique — especially ethical evaluation and genuine, context-bound creation. Treat analysis as contested middle ground you can design either way; treat evaluation and creation as the ground AI cannot take from a student who’s actually thinking.
”The AI Can Do It” Is a Design Signal
Most instructors experience “I pasted my assignment into ChatGPT and it aced it” as a threat. Reframe it as a diagnostic instrument — the fastest course-design feedback you’ve ever had.
When a model completes your assignment well, it’s telling you the assignment lives on a rung where the student’s effort was serving as a proxy for thinking, and the proxy just broke. That’s not the machine cheating. That’s the machine revealing that the task was never really measuring what you hoped.
So run the test deliberately, on everything:
- If the AI aces it and you can’t tell what the student learned → below the line, redesign.
- If the AI produces a starting point that a student must then judge, extend, and take responsibility for → you can design above the line by grading the judgment, not the artifact.
- If the AI is confidently wrong in ways only a knowledgeable student would catch → you’ve found gold; that gap is the assessment.
The higher you climb, the more AI produces output that is fluent and wrong — a confident analysis of the wrong variable, a well-argued evaluation resting on a fabricated fact. Below the line, fluent usually means correct. Above the line, fluency and correctness come apart. This is exactly why higher-order tasks resist automation: catching the difference requires a human who actually understands the domain. Your assessments can be built directly on that gap — “here is the model’s answer; find where it’s wrong and prove it.”
The Diagonal, Applied to Your Target
Take the artifact you picked in Module 1 and place its deliverable on the ladder — not what you hope students think about, but what they actually hand in.
- A lab report that’s mostly a procedure write-up? Apply — low cost, high AI substitutability.
- A “compare and contrast these two theories” essay? Sounds like analyze, but if the two theories are famous, the contrast is in the model’s training data — it collapses back to understand.
- “Given our messy class dataset, decide which of these three methods to trust and defend it to a skeptic”? Evaluate — the constraints are local, the judgment is owned, the AI is a helper not an author.
The redesign move is almost always the same: add local constraints, require a defended judgment, and make the student accountable for it. The next lesson explains why you can’t do that by throwing out the foundations.
You paste your midterm question into an AI and the answer is excellent. A colleague says 'that proves the question is now worthless.' Why is that the wrong conclusion?
Reveal Answer & Explanation
It proves the question sits on a low rung — it’s a design signal about where the task lives, not a verdict that the underlying skill is worthless.
The skill the question was trying to reach may still matter enormously; the problem is that the deliverable was a proxy that AI can now produce. The right response isn’t to discard the question, it’s to move it up the ladder: keep the same content, but ask students to critique the AI’s excellent answer, apply the idea to an unfamiliar local case, or defend a judgment the model can’t own. “The AI can do it” tells you the rung, and the rung tells you how to redesign.
Up Next
If AI made the lower rungs cheap, the tempting conclusion is “so stop teaching them.” The next lesson explains why that’s a trap — and how foundational knowledge changes role without becoming optional.