Generating Question Banks for Computer Science With AI
A computer science question bank needs one real addition beyond the standard five tags, and it only applies to one of the four non-code strands: an evidentiary-basis tag, marking whether an Impacts of Computing item is anchored to a real, documented case or a hedged hypothetical scenario. Computing Systems, Networks & the Internet, and Data & Analysis need no such wrinkle — the standard five-tag scheme already serves them cleanly.
Quick Answer: A CS question bank inherits the standard five tags — content standard, Bloom's level, format, difficulty, last-used — with zero wrinkles required, unlike ESL's content-standard split or coding's language/version qualifier. The one real addition is a new evidentiary-basis tag, scoped only to Impacts of Computing items, distinguishing a documented-case item from a hedged-hypothetical one. It exists because the exam guide's point-weighted blueprint requires pulling a majority-share set of documented-case items at exam stakes — a retrieval job content standard and format can't do alone. Unlike this series' other new tags, it's also a candidate for periodic upgrade, not just re-eligibility, since a documented incident can later surface that matches a scenario a bank already holds as a hedged hypothetical.
This guide covers:
- What a CS bank inherits from general bank architecture, and why content standard needs no wrinkle at all
- Why Data & Analysis's chart-based items are the closest CS gets to music's asset-dependency problem — and why they still fail the test
- The real gap: an evidentiary-basis tag for Impacts of Computing, tested against the retrieval job the exams guide actually created
- Whether an item's evidentiary basis can change over time, and what that means for how a bank gets maintained
- A full tag set, a step-by-step build workflow, and a tools comparison
This entry sits inside the same subject-by-subject series as How to Create Spanish Quizzes With AI, where each guide tests a shared format question against one subject's real structure. Within computer science specifically, the origin hub mapped the four non-code CSTA strands this whole guide works inside — Algorithms & Programming, the coding strand, is its own separate series and doesn't appear here.
What a Computer Science Question Bank Inherits From General Bank Architecture
A CS bank starts from the same five tags any Bloom's-aligned bank needs, and every one of them carries over with no CS-specific rework at all. The Bloom's-aligned bank pillar set that baseline against a living table-of-specifications blueprint: content standard, Bloom's level, item format, difficulty, and last-used or exposure (Notar, Zuelke, Wilson, & Yunker, 2004).
Five Tags That Transfer Without Modification
- Content standard records the CSTA strand and concept — "Networks & the Internet: DNS lookup sequence," "Impacts of Computing: algorithmic bias" — the same way any subject's concept tag works.
- Bloom's level doesn't discriminate a binary-conversion item from a bias-scenario item any more than it discriminates a biology recall item from an inference item.
- Item format (MCQ, true/false, short-answer) carries over cleanly, and — unlike coding's four item types — doesn't hide a second axis underneath it here.
- Difficulty tracks content difficulty or a pass rate, unchanged in principle from any other subject's bank.
- Last-used/exposure works identically across subjects: it tracks whether a student has already seen an item.
Why Content Standard Needs No Wrinkle at All
The ESL bank guide found content standard needed a language-item/content-item lane split, and the coding bank guide found it needed a language/version qualifier. CS is a genuine exception to that pattern.
A CS concept tag like "input vs. output devices" or "algorithmic bias" already nests cleanly under one CSTA strand, with no second value competing for the same field the way a Python-2-versus-3 item does. Nothing about a router-vs-switch item, for instance, depends on an unstated qualifier the way a coding item depends on an unstated language version.
Content standard for computer science does exactly one job — name the strand and the concept — and it does that job completely on its own. That's a genuinely simpler starting point than either ESL or coding got, and it means CS's real complication lives somewhere else entirely.
Testing the Obvious Candidate: Does CS Need an Asset-Dependency Tag Like Music's?
The obvious next move copies music's bank guide directly: music needed a stimulus-asset tag for notation images and an audio-source tag for recordings, because neither dependency was nameable from content standard alone. Both the ESL and coding bank guides ran this same test and ruled it out quickly, since their items are self-contained text. CS deserves a harder look, because one of its strands visibly involves images.
Data & Analysis's Chart Items Are the Closest CS Gets — and Still Fail the Test
The origin hub's Data & Analysis items include chart-reading questions — "which chart type best represents this dataset" — and the true/false sibling built claims directly against an attached bar chart's bar order. At first glance, that chart looks exactly like music's staff excerpt: an image the item depends on.
Music's actual test is narrower than "does this item use an image." It's whether the image is an unrecoverable external asset:
- A staff excerpt or recording is a real, sourced artifact — AI cannot generate playable audio, and a lost recording can be permanently gone.
- A chart accompanying a Data & Analysis item is built from an invented dataset for that item, not sourced from an external recording — it can be regenerated for free, from the same numbers, at any time.
- Nothing points outside the document. A chart's underlying data lives in the item itself, unlike a "see attached staff excerpt" instruction that references something the item can't reproduce on its own.
| Property | Music's Audio Dependency | CS's Chart-Based Data Item |
|---|---|---|
| Is the underlying asset externally sourced? | Yes — a real recording or transcribed score | No — a synthetic dataset invented for the item |
| Can it be regenerated for free? | No — AI cannot generate real playable audio | Yes — the same chart renders identically from the same numbers |
| Does a lost asset make the item permanently unusable? | Yes | No — the chart is trivially rebuilt from the stored data |
So the direct parallel fails, even on CS's best candidate for it. That's a more honest answer than assuming it fails because "it's all text" — a chart genuinely is an image, and it genuinely doesn't carry music's risk anyway.
The Real Question: Does Impacts of Computing Need an Evidentiary-Basis Tag?
Yes — and the case for it comes directly from a retrieval job the exams guide already created, one neither content standard nor item format can do on their own. The true/false guide first established that an ethics claim only survives selected-response formats when it's anchored to a real, documented case rather than a hedged hypothetical.
The exams guide then found exam stakes push further: a graded Impacts of Computing section needs a majority of its points riding on documented-case items, reserving hedged-hypothetical items for a minority share or lower-stakes formative use.
The Retrieval Test That Point Table Sets Up
The exams guide's worked example split a 40-point Impacts of Computing section into 25 documented-case points and 15 project-based points. Building that exam from a standing bank means one concrete retrieval job: pull enough documented-case items to hit the majority share, then fill the remainder from hedged-hypothetical or lower-stakes items.
- Content standard doesn't answer this. "Algorithmic bias" tags both a documented-case item about Amazon's scrapped recruiting tool and a hedged-hypothetical item about a described-but-fictional hiring algorithm identically.
- Item format only partially answers it. The true/false guide found true/false items are functionally constrained to documented cases — a bare declarative sentence can't carry the hedge a hypothetical needs. MCQ and short-answer items carry no such constraint; both formats can legitimately hold either evidentiary basis.
- Difficulty and Bloom's level don't touch this axis at all. A documented-case recall item and a hedged-hypothetical analysis item can sit at the same difficulty and Bloom's level while differing completely in what a challenge from a student, parent, or administrator would need to survive on.
Why This Doesn't Collapse Into an Existing Tag
Two items can share an identical content-standard tag, format, and difficulty rating, while one rests on Reuters' reporting on Amazon's tool (Dastin, 2018) and the other rests on a scenario a writer invented to illustrate the same mechanism. Nothing already in the bank records which is which.
Skipping the tag doesn't make either item wrong. It makes a teacher assembling a graded exam re-read every Impacts of Computing item by hand to sort documented-case from hedged-hypothetical — exactly the manual audit a bank's tagging system exists to eliminate.
That's the same shape of argument the ESL guide made for its home-language tag: the new field is the one thing standing between a bank and a retrieval task tagging was supposed to make trivial.
Does an Item's Evidentiary Basis Ever Change? The Periodic-Review Property
An item's evidentiary-basis tag doesn't change on its own — a hedged-hypothetical item stays hedged-hypothetical exactly as written, forever, unless someone actively rewrites it. But the underlying world it describes keeps producing new documented cases, which makes this tag behave differently from every other new tag this series has found.
From Hypothetical to Documented: What Actually Has to Happen
Say a bank holds a hedged-hypothetical item describing a hiring tool that screens out older applicants based on patterns in past hiring data, phrased with the honest hedge "would likely." That mechanism isn't invented — it's close to what the U.S. Equal Employment Opportunity Commission alleged against iTutorGroup, whose hiring software was programmed to automatically reject female applicants 55 and older and male applicants 60 and older, a case the agency settled in 2023 (EEOC, 2023).
A bank writer who learns of a real, matching case has a choice the original item never offered: rewrite it to cite the real case, upgrading its tag from hedged-hypothetical to documented-case. Nothing forces that rewrite — the old hedged version keeps working exactly as it did.
Why This Differs From Every Other New Tag in This Series
- ESL's home-language tag never changes; only a roster's relevance to an already-correct item shifts.
- Coding's item-type tag is fixed the moment an item is written — predict-output stays predict-output.
- Coding's language/version qualifier can go stale, but gets re-verified for free against a fixed standard: a live interpreter.
- Evidentiary basis is the first tag in this series that's a genuine candidate for editorial upgrade, not just re-eligibility or re-verification — because the real world keeps producing new documented incidents a previously-hypothetical item can be rewritten to cite.
A publicly maintained resource makes that upgrade path realistic rather than speculative: the AI Incident Database, built to catalog real, reported AI-harm events so the same failure patterns don't keep recurring uncatalogued (McGregor, 2021). Checking a hedged-hypothetical item against a growing, citable case log is a genuinely different maintenance task than anything ESL's or coding's new tags require.
That doesn't make evidentiary basis a tag needing constant babysitting. Most hedged-hypothetical items will sit exactly as written for years — the review is worth doing periodically, not per item, per grading cycle.
Do the Other Three Non-Code Strands Need Any Tagging Wrinkle?
No — Computing Systems, Networks & the Internet, and Data & Analysis are all straightforwardly served by the standard five-tag scheme, with no evidentiary-basis-style addition and no content-standard wrinkle. That's worth confirming directly rather than assuming, the same way this guide tested the asset-dependency parallel instead of dismissing it outright.
Computing Systems and Networks & the Internet: Clean
Both strands rest on fixed definitions and sequences — a router's job versus a switch's, the order a webpage request actually follows. Neither item type carries a real-versus-hypothetical axis the way an Impacts of Computing claim does; a router definition is simply correct or it isn't. A specific device spec could theoretically date itself the way a coding item's language version can, but the origin hub's own item types for these two strands — input/output categorization, request-sequencing — never reference specific products or capacities in the first place.
Data & Analysis: Clean, With One Write-Time Caveat
Chart-type conventions and correlation-versus-causation claims rest on fixed, checkable rules — a pie chart's job versus a line chart's, a shared trend versus a causal claim. The true/false guide flagged that an unscoped comparative word like "best" can turn a clean claim into a judgment call, but that's a check performed once, while writing the item — not a property a bank needs a persistent field to track.
| Strand | Tagging Wrinkle Needed? | Why |
|---|---|---|
| Computing Systems | No | Fixed definitions and roles; no real-vs-hypothetical axis |
| Networks & the Internet | No | Fixed, checkable sequences; same reasoning |
| Data & Analysis | No | Fixed conventions; scope-creep words are a write-time check, not a bank field |
| Impacts of Computing | Yes — evidentiary basis | The only strand where two items can share every other tag and still differ in what a challenge needs to survive on |
Three of the four non-code strands stay clean. That's not a coincidence specific to CS — it mirrors the coding bank guide's finding that most of a subject's structure transfers unmodified, with the real complexity concentrated in one identifiable spot rather than spread evenly across the whole discipline.
The Tag Set a Computer Science Question Bank Actually Needs
A workable CS bank keeps all five general tags unmodified and adds exactly one new tag, scoped conditionally to a single strand. That's a lighter footprint than ESL's wrinkle-plus-two-tags shape, and a differently-scoped addition than coding's single unconditional tag.
| Tag | Carries Over Unmodified? | Notes for Computer Science |
|---|---|---|
| Content standard | Yes, no wrinkle | Strand + concept, e.g. "Impacts of Computing: algorithmic bias" |
| Bloom's level | Yes | Doesn't discriminate evidentiary basis or strand |
| Item format | Yes | MCQ, true/false, short-answer; true/false functionally implies documented-case for Impacts items |
| Difficulty | Yes | Content difficulty or pass rate, unrelated to evidentiary basis |
| Last used/exposure | Yes | Same purpose as any bank, any subject |
| Evidentiary basis (new) | No | Conditional — applies only to Impacts of Computing items; values: documented-case or hedged-hypothetical |
A general-values-judgment claim — "is AI hiring fair?" — never enters the bank at all, per the origin hub's own gradability standard, so the evidentiary-basis tag only ever needs those two real values for anything that legitimately makes it into the bank.
Building a Computer Science Question Bank: A Step-by-Step Workflow
Building the bank starts with naming the CSTA strand before writing anything, since that single decision determines whether an evidentiary-basis tag applies at all.
- Pick the strand first: Computing Systems, Networks & the Internet, Data & Analysis, or Impacts of Computing.
- For the first three strands, tag the standard five fields and stop. No wrinkle, no sixth field, confirmed above.
- For an Impacts of Computing item, decide the evidentiary basis before drafting, not after — a documented-case item and a hedged-hypothetical item need different sourcing before a single sentence gets written.
- For a documented-case item, cite the real source by name and year in the item's own notes, so the tag is checkable against something concrete.
- For a hedged-hypothetical item, keep the hedge language intact ("likely," "tends to") and flag it for periodic review against real, newly reported cases.
- If building toward a graded exam, use the exams guide's majority-documented-case rule to decide how many of each evidentiary-basis value the bank needs before drafting more.
- Tag every item on all six fields at creation. Retagging a backlog once a bank reaches a few hundred items costs far more than tagging while the decision is still fresh.
You could use EduGenius to hold a class profile's grade level and CSTA strand focus while drafting a batch of non-code CS items in one pass — binary-representation, network-sequencing, or bias-scenario items, tagged by strand — then export the finished set with an answer key once every documented-case citation checks out.
Can AI Tag Evidentiary Basis Accurately?
No dedicated accuracy research exists yet for classifying Impacts of Computing items by evidentiary basis, the same honest gap the ESL and coding bank guides found for their own new tags. The tagging-speed guide reports real figures for automated Bloom's-level classification specifically — roughly 72-73% for general-purpose AI tools, rising to the low-to-mid 90s for purpose-built classifiers (Kumar, Gulwani, & Singh, 2025) — but that evidence doesn't extend to evidentiary-basis tagging.
This gap matters more than it did for coding's item-type tag, which is a byproduct of a decision a writer already made before drafting. Evidentiary basis needs an extra, independent step: confirming a cited case is real and correctly matches the item. Ongoing item-pool quality assurance research treats this kind of verification as a continuous task, not a one-time check at creation (von Davier et al., 2026).
Tools for Building a Computer Science Question Bank
No mainstream AI content tool ships with an evidentiary-basis field by default; the real difference is whether a tool holds a stated strand and case citation across a whole drafting session or needs them retyped into every prompt.
| Tool | Tags Strand at Creation | Distinguishes Documented-Case From Hypothetical | Holds a Point-Weighted Blueprint | Starting Price |
|---|---|---|---|---|
| EduGenius | Yes, if named in the prompt | Only if the case is supplied | Yes, via class profile and prompt context | 25 welcome credits free; Starter $7.99/month (500 credits); Professional $15.99/month (1,000 credits) |
| General-purpose AI chat tools | Only if requested | Rarely, unprompted | Only if retyped into each prompt | Varies by provider |
| CS-specific curriculum platforms | Often standards-tagged | Varies by publisher | Varies by publisher | Free tier common; paid school licenses vary |
| Spreadsheet or LMS native item banks | Fully manual; teacher enters every field | Depends entirely on the teacher | Fully manual | Often bundled with an existing LMS, or free |
No tool automatically verifies that a documented-case citation is accurate or that a hedged-hypothetical item's hedge language survived a rewrite — both checks stay with the teacher. The broader bank-tagging landscape across subjects covers the tagging-speed question generally; this guide narrows it to computer science's one real addition.
Pro Tips for a Computer Science Question Bank
- Tag the strand before writing anything. It's the single decision that determines whether an evidentiary-basis tag even applies.
- Keep a running, citable list of documented cases — Amazon's scrapped recruiting tool (Dastin, 2018), the Gender Shades study (Buolamwini & Gebru, 2018), the iTutorGroup settlement (EEOC, 2023) — rather than researching one from scratch every time an item needs it.
- Check the AI Incident Database periodically (McGregor, 2021) against your bank's hedged-hypothetical items, since a scenario invented two years ago may match something that actually happened since.
- Never let a comparative word like "best" or "always" slip into a Data & Analysis claim unscoped — that's a write-time check, not a reason to add a bank field.
- Reserve true/false for Impacts of Computing items only when the documented case is already confirmed, per the true/false guide's own finding that the format can't carry a hedge.
What to Avoid When Building a Computer Science Question Bank
Copying Music's Asset-Dependency Tag Onto Data & Analysis
A chart item's dataset is invented for that item and regenerates for free. It carries none of the unrecoverable-asset risk that justifies music's stimulus-asset tag.
Leaving Evidentiary Basis Untagged Because "the Concept Tag Already Covers It"
Two items can share an identical content-standard tag and differ completely in whether they're anchored to a real case or a hedged scenario. Content standard was never built to carry that distinction.
Treating a Hedged-Hypothetical Item as Permanently Fixed
An item's tag doesn't update itself, but the case log it could be checked against keeps growing. A hedged-hypothetical item worth writing once is worth periodically checking against a real incident that might now exist.
Letting a General Values Judgment Into the Bank at All
"Is AI hiring fair?" has no gradable answer at any evidentiary basis. The origin hub's gradability standard rules it out before a tag becomes relevant.
Key Takeaways
- A computer science question bank needs exactly one new tag: evidentiary basis, scoped only to Impacts of Computing items, distinguishing documented-case from hedged-hypothetical.
- Content standard needs zero wrinkle for CS, unlike ESL's language-item/content-item split or coding's language/version qualifier — a genuine, tested exception in this series.
- The music-style asset-dependency tag fails even against CS's best candidate, Data & Analysis's chart-based items, since a chart's dataset is invented per item and regenerates for free.
- The evidentiary-basis tag exists to serve a real retrieval job: assembling the majority-documented-case exam split the exams guide already established, which content standard and format can't do alone.
- Evidentiary basis is the first tag in this series that's a genuine candidate for editorial upgrade, not just re-eligibility, since a real documented case can later surface matching a scenario the bank already holds as hedged-hypothetical.
- Three of the four non-code strands stay fully clean on the standard five-tag scheme — only Impacts of Computing needs the new field.
- No published accuracy research yet covers automated evidentiary-basis classification, and the check it needs — verifying a citation is real — is a step beyond what coding's or ESL's new tags required.
Frequently Asked Questions
Does a computer science question bank need a tag for external files, the way a music question bank needs one for audio and notation?
No, even for its most image-heavy strand. A Data & Analysis chart item's dataset is invented for that item and regenerates for free at any time, unlike a real recording or transcribed score that can go permanently unrecoverable.
What's the one new tag a computer science question bank actually needs?
An evidentiary-basis tag, applied only to Impacts of Computing items, marking whether the item is anchored to a real, documented case or a hedged hypothetical scenario. It supports the majority-documented-case exam split the exams guide established — a retrieval job content standard and item format can't do alone.
Can a hedged-hypothetical item become a documented-case item later?
Not automatically — a tag doesn't change on its own. But a writer aware of a newly reported, matching real incident can rewrite the item to cite it, upgrading the tag deliberately. That makes evidentiary basis worth periodically re-checking, unlike this series' other new tags, which stay fixed once assigned.
Do Computing Systems, Networks & the Internet, or Data & Analysis need any special tagging treatment?
No. All three rest on fixed definitions, sequences, or conventions with no real-versus-hypothetical axis to track, so the standard five-tag scheme serves them unmodified. Only Impacts of Computing needs the evidentiary-basis addition.
Sources
- Computer Science Teachers Association. (2017). CSTA K-12 Computer Science Standards.
- Notar, C. E., Zuelke, D. C., Wilson, J. D., & Yunker, B. D. (2004). The table of specification: Insuring accountability in teacher made tests. Journal of Instructional Psychology, 31(2), 115-129.
- Dastin, J. (2018). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters.
- Buolamwini, J., & Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of Machine Learning Research (Conference on Fairness, Accountability, and Transparency), 81, 1–15.
- U.S. Equal Employment Opportunity Commission. (2023). iTutorGroup to pay $365,000 to settle EEOC discrimination suit. EEOC Press Release.
- McGregor, S. (2021). Preventing repeated real world AI failures by cataloging incidents: The AI Incident Database. Proceedings of the AAAI Conference on Artificial Intelligence / Innovative Applications of Artificial Intelligence, 35(17), 15458-15463.
- American Educational Research Association, American Psychological Association, & National Council on Measurement in Education. (2014). Standards for Educational and Psychological Testing.
- Kumar, R., Gulwani, D., & Singh, S. (2025). Automated analysis of learning outcomes and exam questions based on Bloom's taxonomy. arXiv preprint arXiv:2511.10903.
- von Davier, A. A., Zhang, X., Attali, Y., Park, Y., Church, J., Runge, A., LaFlair, G. T., & Tsigler, A. (2026). Analytics for quality assurance for item pools (AQuAP): Monitoring and maintaining item bank health in AI-driven assessment systems. arXiv preprint arXiv:2606.18536.