Generating Formative Assessments for STEM With AI

A fast STEM formative check can test genuine cross-disciplinary reasoning, but only a narrower slice of it than a graded quiz item reaches. Of the two item types that actually earn the "STEM" label rather than borrowing it, a cross-disciplinary compute item and an engineering-design trade-off item, only one compresses into a 30-90-second prompt, and even that one only compresses partway (NGSS Lead States, 2013).

That's a different answer than biology gave. Generating Formative Assessments for Biology With AI found that natural-selection misconceptions fit a fast "why" prompt well, sometimes better than a scored distractor, because they live entirely in a sentence's causal language (Bishop & Anderson, 1990). STEM's genuinely integrated item types are quantitative and constraint-based by definition, so that same move doesn't have anywhere obvious to land.

Quick Answer: Cross-disciplinary compute items partially survive a fast check, but only when the underlying concept determines a direction a student couldn't otherwise guess from generic math reasoning (Stavy & Tirosh, 2000). Engineering-design trade-off items mostly don't survive at all. Their full evaluation needs a real constraint table, and the only piece that fits the clock is a narrower, prerequisite question about how trade-offs work, not the trade-off itself.

This guide covers:

• Why the STEM quiz guide's two genuinely integrated item types resist compressing the way biology's misconceptions did
• Which cross-disciplinary compute items survive a 30-90-second directional-prediction prompt, and which only look like they do
• Why engineering-design trade-off items mostly don't fit the format, and the one prerequisite concept that does
• A step-by-step AI workflow, a tools comparison, and the mistakes that undercut a fast STEM check

The same design question runs through this site's other subject-specific guides, from Spanish quiz design to STEM's own quiz and worksheet work.

What a STEM Formative Check Would Actually Have to Test

A STEM formative check only earns its label if it tests one of the two item types that make a "STEM quiz" genuinely different from a science quiz or a math quiz, not a topic that merely sounds STEM-flavored. The STEM quiz guide already ruled out single-subject and mixed-topic requests as the wrong ground; what's left is narrow on purpose.

The Two Item Types It Would Need to Compress

A cross-disciplinary compute item needs a calculation that depends on a science or engineering concept, not just a fact recalled next to some unrelated arithmetic. An engineering-design trade-off item needs explicit, sometimes competing constraints, cost, weight, durability, evaluated against real numbers (NGSS Lead States, 2013). Both, by the quiz guide's own definition, need enough setup, numbers, a scenario, a data table, that a scored item typically runs several sentences before a student even reaches the question.

A formative check runs the opposite direction: short-cycle formative work is answered minute-by-minute inside a live lesson, not worked out over a homework block (Wiliam & Thompson, 2007). That's a real tension worth naming honestly before assuming it resolves the way biology's did.

Applying the Removal Test to a Compressed Prompt

The quiz guide's removal test asks whether stripping out either the science/engineering half or the numeric half still leaves an answerable question. A compressed formative version needs the same test, run differently: strip out the exact numbers while keeping the concept, and ask whether a decidable question survives.

• If the direction of the answer still depends on the concept, the compressed version is doing real STEM work.
• If the direction is obvious from generic math reasoning alone, stripping the numbers didn't just simplify the item, it removed the only thing that made it STEM.
• That second case fits the clock easily and still fails the test, the fast-check equivalent of the quiz guide's stapled item.

Why Biology's Answer Doesn't Transfer Directly

Biology's fast-fitting misconceptions, teleological reasoning and acquired-trait inheritance, never involved a number in the first place; a one-sentence "why" prompt was already the item's natural size (Gregory, 2009). STEM's two genuinely integrated types are defined by numbers and constraints, so there's no equivalent pure-language starting point to just ask. The compression question has to be answered structurally, item type by item type, not assumed to resolve the way it did for biology.

Cross-Disciplinary Compute Items: When the Direction Survives Without the Numbers

A cross-disciplinary compute item's full version, exact numbers and a worked calculation, doesn't fit a 30-90-second prompt. What sometimes fits is a narrower slice: asking a student to predict the direction of the answer, bigger or smaller, sooner or later, before any calculation happens.

Trivial Direction vs. Concept-Gated Direction

Not every directional prediction is doing STEM work, even when it's dressed in scientific vocabulary. Ruth Stavy and Dina Tirosh's research on intuitive reasoning documents a rule students apply constantly across math and science: "more of A, more of B." Most of the time that rule happens to be correct, and a directional prompt built on it tests nothing but generic monotonic reasoning (Stavy & Tirosh, 2000).

1. A trivial-direction prompt: the "more A, more B" intuition already gives the right answer, so a student needs no science or engineering concept to solve it. It fails the removal test even though it fits the clock.
2. A concept-gated direction prompt: the correct direction actually contradicts "more A, more B," so a student needs the specific concept to answer correctly. This is where a compute item genuinely compresses.

A Worked Example That Passes the Test

Physics education research built around the Force Concept Inventory documents a specific, persistent version of this: students commonly assume a heavier object exerts more force in a collision than a lighter one pushes back (Hestenes, Wells, & Swackhamer, 1992). A fast check can surface that directly, no numbers required.

"In a head-on collision between a loaded delivery truck and a compact car, is the force the truck exerts on the car bigger, smaller, or the same as the force the car exerts on the truck? Explain your reasoning in one sentence."

A student who answers "bigger, because the truck is heavier" is applying the "more mass, more force" intuition Newton's third law contradicts. A student who answers "the same, because forces always come in equal pairs" holds the concept. Removing the numbers here changes nothing, since the item never needed any; removing the physics concept leaves the question undecidable by generic reasoning alone.

A Worked Example That Fails Anyway

A heating-rate prompt looks just as science-flavored, but the direction never needed the science at all.

"A pot of water is heating at a constant rate. If you turn the burner to a higher setting, will the water reach boiling sooner or later?"

The answer, sooner, follows from "more heat, faster result" without any specific science concept. Strip the stove framing and the same claim, a bigger rate produces a result sooner, stays fully true. It's fast and real, but it tests monotonic reasoning applied to a labeled scenario, not STEM integration.

The table's middle row is the real find: a directional prompt only earns the STEM label when the concept, not generic intuition, is doing the deciding.

A Grade-Calibrated Illustration

Say you teach a grade 4 class exploring simple machines and want a fast check on levers before the unit's vocabulary gets introduced.

"Two students push down on a seesaw from the same distance from the middle, but one weighs more. Will the seesaw tip toward the heavier student, the lighter student, or stay level? Explain in one sentence."

A grade 8 physical science class studying momentum can handle the truck/car version directly, since Newton's third law is typically named by that point.

Engineering-Design Trade-Off Items: Why Even the Direction Doesn't Fit

An engineering-design trade-off item's whole point is that no single option wins on every constraint, which is exactly what keeps it from compressing into a fast check, even in direction-only form. The quiz guide's trade-off items need a real data table, cost, weight, durability, evaluated against explicit limits (NGSS Lead States, 2013).

What the Full Item Needs That a Sentence Can't Hold

Genetics' ratio-conflation error found a partial escape hatch in biology's formative guide: a direction-only prompt still worked because the dominance concept alone, no arithmetic required, determines which way a ratio skews. A trade-off item doesn't have that escape hatch, since its whole design intentionally sets constraints against each other, cost against durability, weight against strength, so no option looks obviously better without checking actual numbers.

• Strip the numbers from a trade-off item and there's no principled "direction" left to predict. Unlike a rate problem, there's no default assumption a student could reasonably hold about which constraint wins.
• A single-constraint shortcut fails the removal test the same way the stove example did. Asking whether $8 exceeds a $6 budget cap is a bare number comparison; the engineering framing adds nothing.
• This mirrors biology's procedural grid error: the mistake, and the correct answer, only exist inside the fully worked-out comparison. There's no sentence-sized substitute.

The One Prerequisite Concept That Does Fit the Clock

What survives isn't the trade-off evaluation itself, it's a check of whether a student understands that a design decision has to satisfy every stated constraint, not just the most visible one. NGSS's engineering-design strand treats defining a problem by its full set of criteria and constraints as a distinct step, prior to comparing solutions (NGSS Lead States, 2013; National Research Council, 2012).

"True or false: the cheapest option in a design comparison is always the best one to choose. Explain your reasoning in one sentence."

A student who answers "true, because cost is what matters most" is applying a single-criterion shortcut. A student who answers "false, because it also has to meet the other requirements, like weight or durability" holds the concept a real trade-off item will later test with actual numbers. That's a genuinely useful, genuinely fast check; it just isn't the trade-off item itself.

A Grade-Calibrated Illustration

Say you teach a grade 3 class about to design a paper bridge and want a fast check before handing out materials.

"If a bridge design uses the least paper, is it automatically the strongest one? Explain in one sentence."

A grade 9 engineering elective preparing to compare drone battery options can use the same true/false shape aimed at their own constraint set, cost, weight, flight time, before the full comparison table comes out.

No row in that table does what the full item does in under two minutes. That's the honest finding: trade-off reasoning, unlike part of the compute-item territory, doesn't have a genuine fast-check version of itself, only a genuine fast-check version of the concept underneath it.

A Step-by-Step AI Workflow for Fast STEM Checks

Building a usable fast STEM check starts with sorting the target into one of three buckets, full computation, concept-gated direction, or trade-off prerequisite, since only two of the three actually fit a 30-90-second prompt.

1. Name the exact concept, not just the unit topic — "Newton's third law in a collision," not "forces."
2. Run the removal test on the compressed version before writing it, not just on the full item: does the direction still depend on the concept once the numbers are gone?
3. For a compute item, check whether "more A, more B" already gives the right answer. If it does, the item is testing generic reasoning, not STEM integration, no matter how it's worded (Stavy & Tirosh, 2000).
4. For a trade-off item, don't try to compress the evaluation itself. Write a true/false or "does X always win" prompt about the underlying constraint concept instead.
5. Set your class profile's grade level and vocabulary density, so the scenario matches what the class can reason through in under a minute without needing new terms defined mid-question.
6. Anchor the concept to the standard your unit is already teaching (NGSS Lead States, 2013), so a fast check lines up with what's actually been covered.
7. Decide your response branches before reading a single answer: concept correctly applied → move on; mixed reasoning → quick class discussion; consistent naive-intuition answers → flag for direct instruction before the full quiz item.

You could use EduGenius to hold a class profile's grade and subject while drafting a concept-gated directional prompt or a trade-off prerequisite check, then export the finished prompt as a quick slide or a half-sheet to project.

Tools for Fast STEM Formative Checks

A tool built for this job needs to support open-ended directional prompts and a way to flag when a "more A, more B" shortcut would make a question trivial, neither of which a generic quiz generator checks for by default.

How to Write Better Formative Assessments Faster With AI covers the general short-cycle timing problem this guide narrows for STEM specifically. How to Generate STEM Worksheets With AI covers the full engineering-design loop a trade-off item eventually needs, once a class is past the fast prerequisite check.